{ "schemaVersion": "1.0", "name": "ScribeIndex Modellreferenzen", "checked": "2026-10-20", "license": { "name": "CC BY 4.0", "url": "https://creativecommons.org/licenses/by/4.0/deed.de" }, "records": [ { "id": "aa-whisper-large-v3-fal", "model": "Whisper Large v3", "aliases": [ "Whisper large-v3", "OpenAI Whisper Large v3" ], "kind": "speech", "benchmark": "Artificial Analysis Word Error Rate Index", "version": "2.2", "value": 4.1, "unit": "%", "direction": "lower", "configuration": "ohne Streaming, fal.ai API", "scope": "Englischsprachiger Modelltest von Artificial Analysis (AA-AgentTalk 50 %, VoxPopuli 25 %, Earnings22 25 %). Keine medizinischen oder deutschen Aufnahmen, keine Messung der Anbieteranpassung.", "checked": "2026-10-02", "source": { "title": "Artificial Analysis: Whisper Large v3, fal.ai", "url": "https://artificialanalysis.ai/speech-to-text/non-streaming" }, "authority": "independent", "rating": "Solide" }, { "id": "aa-whisper-large-v3-replicate", "model": "Whisper Large v3", "aliases": [ "Whisper large-v3", "OpenAI Whisper Large v3" ], "kind": "speech", "benchmark": "Artificial Analysis Word Error Rate Index", "version": "2.2", "value": 10.1, "unit": "%", "direction": "lower", "configuration": "ohne Streaming, Replicate API", "scope": "Englischsprachiger Modelltest von Artificial Analysis (AA-AgentTalk 50 %, VoxPopuli 25 %, Earnings22 25 %). Keine medizinischen oder deutschen Aufnahmen, keine Messung der Anbieteranpassung.", "checked": "2026-10-02", "source": { "title": "Artificial Analysis: Whisper Large v3, Replicate", "url": "https://artificialanalysis.ai/speech-to-text/non-streaming" }, "authority": "independent", "rating": "Schwach" }, { "id": "aa-whisper-large-v3-turbo-groq", "model": "Whisper Large v3 Turbo", "aliases": [ "Whisper large-v3-turbo", "OpenAI Whisper Large v3 Turbo", "Whisper large-v3 Turbo", "Whisper V3 Turbo" ], "kind": "speech", "benchmark": "Artificial Analysis Word Error Rate Index", "version": "2.2", "value": 4.6, "unit": "%", "direction": "lower", "configuration": "ohne Streaming, Groq API", "scope": "Englischsprachiger Modelltest von Artificial Analysis (AA-AgentTalk 50 %, VoxPopuli 25 %, Earnings22 25 %). Keine medizinischen oder deutschen Aufnahmen, keine Messung der Anbieteranpassung.", "checked": "2026-10-02", "source": { "title": "Artificial Analysis: Whisper Large v3 Turbo, Groq", "url": "https://artificialanalysis.ai/speech-to-text/non-streaming" }, "authority": "independent", "rating": "Solide" }, { "id": "aa-whisper-large-v2-openai", "model": "Whisper Large v2", "aliases": [ "Whisper large-v2", "OpenAI Whisper Large v2" ], "kind": "speech", "benchmark": "Artificial Analysis Word Error Rate Index", "version": "2.2", "value": 4.1, "unit": "%", "direction": "lower", "configuration": "ohne Streaming, OpenAI API", "scope": "Englischsprachiger Modelltest von Artificial Analysis (AA-AgentTalk 50 %, VoxPopuli 25 %, Earnings22 25 %). Keine medizinischen oder deutschen Aufnahmen, keine Messung der Anbieteranpassung.", "checked": "2026-10-02", "source": { "title": "Artificial Analysis: Whisper Large v2, OpenAI", "url": "https://artificialanalysis.ai/speech-to-text/non-streaming" }, "authority": "independent", "rating": "Solide" }, { "id": "aa-claude-sonnet-5-max", "model": "Claude Sonnet 5", "aliases": [ "Claude Sonnet 5", "Claude Sonnet 5 (Reasoning-Stufe maximal)" ], "kind": "language", "benchmark": "Artificial Analysis Intelligence Index", "version": "4.3.2", "value": 38, "unit": "points", "direction": "higher", "configuration": "Reasoning-Stufe maximal", "scope": "Allgemeine Modellbewertung. Produktkonfiguration und klinische Ausgabequalität sind dadurch nicht belegt.", "checked": "2026-10-02", "source": { "title": "Artificial Analysis: Claude Sonnet 5", "url": "https://artificialanalysis.ai/models/claude-sonnet-5" }, "authority": "independent", "rating": "Stark" }, { "id": "aa-claude-opus-5-max", "model": "Claude Opus 5", "aliases": [ "Claude Opus 5", "Claude Opus 5 (Reasoning-Stufe maximal)" ], "kind": "language", "benchmark": "Artificial Analysis Intelligence Index", "version": "4.3.2", "value": 51, "unit": "points", "direction": "higher", "configuration": "Reasoning-Stufe maximal", "scope": "Allgemeine Modellbewertung. Produktkonfiguration und klinische Ausgabequalität sind dadurch nicht belegt.", "checked": "2026-10-02", "source": { "title": "Artificial Analysis: Claude Opus 5", "url": "https://artificialanalysis.ai/models/claude-opus-5" }, "authority": "independent", "rating": "Spitzenklasse" }, { "id": "aa-gpt-5-6-luna-max", "model": "GPT-5.6 Luna", "aliases": [ "GPT-5.6 Luna", "GPT-5.6 Luna (Reasoning-Stufe maximal)" ], "kind": "language", "benchmark": "Artificial Analysis Intelligence Index", "version": "4.3.2", "value": 37, "unit": "points", "direction": "higher", "configuration": "Reasoning-Stufe maximal", "scope": "Allgemeine Modellbewertung. Produktkonfiguration und klinische Ausgabequalität sind dadurch nicht belegt.", "checked": "2026-10-02", "source": { "title": "Artificial Analysis: GPT-5.6 Luna", "url": "https://artificialanalysis.ai/models/gpt-5-6-luna" }, "authority": "independent", "rating": "Stark" }, { "id": "aa-gpt-5-6-sol-max", "model": "GPT-5.6 Sol", "aliases": [ "GPT-5.6 Sol", "GPT-5.6 Sol (Reasoning-Stufe maximal)" ], "kind": "language", "benchmark": "Artificial Analysis Intelligence Index", "version": "4.3.2", "value": 47, "unit": "points", "direction": "higher", "configuration": "Reasoning-Stufe maximal", "scope": "Allgemeine Modellbewertung. Produktkonfiguration und klinische Ausgabequalität sind dadurch nicht belegt.", "checked": "2026-10-02", "source": { "title": "Artificial Analysis: GPT-5.6 Sol", "url": "https://artificialanalysis.ai/models/gpt-5-6-sol" }, "authority": "independent", "rating": "Stark" }, { "id": "aa-gemini-3-5-flash-high", "model": "Gemini 3.5 Flash", "aliases": [ "Gemini 3.5 Flash", "Gemini 3.5 Flash (Reasoning-Stufe hoch)" ], "kind": "language", "benchmark": "Artificial Analysis Intelligence Index", "version": "4.3.2", "value": 33, "unit": "points", "direction": "higher", "configuration": "Reasoning-Stufe hoch", "scope": "Allgemeine Modellbewertung. Produktkonfiguration und klinische Ausgabequalität sind dadurch nicht belegt.", "checked": "2026-10-02", "source": { "title": "Artificial Analysis: Gemini 3.5 Flash", "url": "https://artificialanalysis.ai/models/gemini-3-5-flash" }, "authority": "independent", "rating": "Stark" }, { "id": "aa-claude-fable-5-max-fallback", "model": "Claude Fable 5", "aliases": [ "Claude Fable 5", "Claude Fable 5 (Reasoning-Stufe maximal, mit Opus 4.8 als Rückfallmodell)" ], "kind": "language", "benchmark": "Artificial Analysis Intelligence Index", "version": "4.3.2", "value": 50, "unit": "points", "direction": "higher", "configuration": "Reasoning-Stufe maximal, mit Opus 4.8 als Rückfallmodell", "scope": "Allgemeine Modellbewertung. Produktkonfiguration und klinische Ausgabequalität sind dadurch nicht belegt.", "checked": "2026-10-02", "source": { "title": "Artificial Analysis: Claude Fable 5", "url": "https://artificialanalysis.ai/models/claude-fable-5" }, "authority": "independent", "rating": "Spitzenklasse" }, { "id": "aa-assemblyai-universal-3-5-pro", "model": "AssemblyAI Universal-3.5 Pro", "aliases": [ "Universal-3.5 Pro", "AssemblyAI Universal 3.5 Pro" ], "kind": "speech", "benchmark": "Artificial Analysis Word Error Rate Index", "version": "2.2", "value": 3, "unit": "%", "direction": "lower", "configuration": "ohne Streaming; Wert aus dem Modelldiagramm „Universal-3.5 Pro, AssemblyAI“, ohne eigene Zeile in der API-Tabelle; im Index derselben Modellvariante zugeordnet wie der Streaming-Eintrag „AssemblyAI U3.5 Realtime Pro - Max Accuracy“", "scope": "Englischsprachiger Modelltest von Artificial Analysis (AA-AgentTalk 50 %, VoxPopuli 25 %, Earnings22 25 %). Keine medizinischen oder deutschen Aufnahmen, keine Messung der Anbieteranpassung.", "checked": "2026-10-02", "source": { "title": "Artificial Analysis: Universal-3.5 Pro, AssemblyAI", "url": "https://artificialanalysis.ai/speech-to-text/non-streaming" }, "authority": "independent", "rating": "Stark" }, { "id": "aa-assemblyai-universal-3-pro", "model": "AssemblyAI Universal-3 Pro", "aliases": [ "Universal-3 Pro", "AssemblyAI Universal 3 Pro" ], "kind": "speech", "benchmark": "Artificial Analysis Word Error Rate Index", "version": "2.2", "value": 3.1, "unit": "%", "direction": "lower", "configuration": "ohne Streaming, AssemblyAI API; API-Modell universal-3-pro", "scope": "Englischsprachiger Modelltest von Artificial Analysis (AA-AgentTalk 50 %, VoxPopuli 25 %, Earnings22 25 %). Keine medizinischen oder deutschen Aufnahmen, keine Messung der Anbieteranpassung.", "checked": "2026-10-02", "source": { "title": "Artificial Analysis: Universal-3 Pro, AssemblyAI", "url": "https://artificialanalysis.ai/speech-to-text/non-streaming" }, "authority": "independent", "rating": "Stark" }, { "id": "aa-assemblyai-universal", "model": "AssemblyAI Universal", "aliases": [ "AssemblyAI Universal" ], "kind": "speech", "benchmark": "Artificial Analysis Word Error Rate Index", "version": "2.2", "value": 3.8, "unit": "%", "direction": "lower", "configuration": "ohne Streaming, AssemblyAI API; API-Modell universal; veröffentlicht am 2024-10-30", "scope": "Englischsprachiger Modelltest von Artificial Analysis (AA-AgentTalk 50 %, VoxPopuli 25 %, Earnings22 25 %). Keine medizinischen oder deutschen Aufnahmen, keine Messung der Anbieteranpassung.", "checked": "2026-10-02", "source": { "title": "Artificial Analysis: Universal, AssemblyAI", "url": "https://artificialanalysis.ai/speech-to-text/non-streaming" }, "authority": "independent", "rating": "Stark" }, { "id": "aa-assemblyai-u3-5-realtime-pro-max-accuracy", "model": "AssemblyAI Universal-3.5 Pro Realtime", "aliases": [ "AssemblyAI U3.5 Realtime Pro", "Universal-3.5 Pro Realtime" ], "kind": "speech", "benchmark": "Artificial Analysis Word Error Rate Streaming Index (Final Transcription)", "version": "2.2", "value": 4, "unit": "%", "direction": "lower", "configuration": "Streaming, AssemblyAI API; API-Modell universal-3-5-pro-max-accuracy; Modus Max Accuracy", "scope": "Englischsprachiger Streamingtest von Artificial Analysis, WER der finalen Transkripte. Keine medizinischen oder deutschen Aufnahmen, keine Messung der Anbieteranpassung.", "checked": "2026-10-02", "source": { "title": "Artificial Analysis: AssemblyAI U3.5 Realtime Pro - Max Accuracy", "url": "https://artificialanalysis.ai/speech-to-text/streaming" }, "authority": "independent", "rating": "Stark" }, { "id": "aa-assemblyai-u3-realtime-pro", "model": "AssemblyAI Universal-3 Pro Realtime", "aliases": [ "AssemblyAI U3 Realtime Pro", "Universal-3 Pro Realtime" ], "kind": "speech", "benchmark": "Artificial Analysis Word Error Rate Streaming Index (Final Transcription)", "version": "2.2", "value": 4.1, "unit": "%", "direction": "lower", "configuration": "Streaming, AssemblyAI API; API-Modell u3-rt-pro", "scope": "Englischsprachiger Streamingtest von Artificial Analysis, WER der finalen Transkripte. Keine medizinischen oder deutschen Aufnahmen, keine Messung der Anbieteranpassung.", "checked": "2026-10-02", "source": { "title": "Artificial Analysis: AssemblyAI U3 Realtime Pro", "url": "https://artificialanalysis.ai/speech-to-text/streaming" }, "authority": "independent", "rating": "Stark" }, { "id": "aa-deepgram-nova-3", "model": "Deepgram Nova-3", "aliases": [ "Nova-3", "Deepgram Nova 3" ], "kind": "speech", "benchmark": "Artificial Analysis Word Error Rate Index", "version": "2.2", "value": 5.2, "unit": "%", "direction": "lower", "configuration": "ohne Streaming, Deepgram API; API-Modell nova-3", "scope": "Englischsprachiger Modelltest von Artificial Analysis (AA-AgentTalk 50 %, VoxPopuli 25 %, Earnings22 25 %). Keine medizinischen oder deutschen Aufnahmen, keine Messung der Anbieteranpassung.", "checked": "2026-10-02", "source": { "title": "Artificial Analysis: Nova-3, Deepgram", "url": "https://artificialanalysis.ai/speech-to-text/non-streaming" }, "authority": "independent", "rating": "Solide" }, { "id": "aa-deepgram-nova-3-telnyx", "model": "Deepgram Nova-3", "aliases": [ "Nova-3", "Deepgram Nova 3" ], "kind": "speech", "benchmark": "Artificial Analysis Word Error Rate Index", "version": "2.2", "value": 4.8, "unit": "%", "direction": "lower", "configuration": "ohne Streaming, Telnyx API; API-Modell deepgram/nova-3; Hosting durch Telnyx, nicht Deepgram", "scope": "Englischsprachiger Modelltest von Artificial Analysis (AA-AgentTalk 50 %, VoxPopuli 25 %, Earnings22 25 %). Keine medizinischen oder deutschen Aufnahmen, keine Messung der Anbieteranpassung.", "checked": "2026-10-02", "source": { "title": "Artificial Analysis: Nova-3, Telnyx, Telnyx", "url": "https://artificialanalysis.ai/speech-to-text/non-streaming" }, "authority": "independent", "rating": "Solide" }, { "id": "aa-deepgram-nova-2-telnyx", "model": "Deepgram Nova-2", "aliases": [ "Nova-2", "Deepgram Nova 2" ], "kind": "speech", "benchmark": "Artificial Analysis Word Error Rate Index", "version": "2.2", "value": 5.1, "unit": "%", "direction": "lower", "configuration": "ohne Streaming, Telnyx API; API-Modell deepgram/nova-2; Hosting durch Telnyx, nicht Deepgram", "scope": "Englischsprachiger Modelltest von Artificial Analysis (AA-AgentTalk 50 %, VoxPopuli 25 %, Earnings22 25 %). Keine medizinischen oder deutschen Aufnahmen, keine Messung der Anbieteranpassung.", "checked": "2026-10-02", "source": { "title": "Artificial Analysis: Nova-2, Telnyx, Telnyx", "url": "https://artificialanalysis.ai/speech-to-text/non-streaming" }, "authority": "independent", "rating": "Solide" }, { "id": "aa-deepgram-nova-3-realtime", "model": "Deepgram Nova-3 Realtime", "aliases": [ "Nova-3 Streaming", "Deepgram Nova-3" ], "kind": "speech", "benchmark": "Artificial Analysis Word Error Rate Streaming Index (Final Transcription)", "version": "2.2", "value": 6.6, "unit": "%", "direction": "lower", "configuration": "Streaming, Deepgram API; API-Modell nova-3", "scope": "Englischsprachiger Streamingtest von Artificial Analysis, WER der finalen Transkripte. Keine medizinischen oder deutschen Aufnahmen, keine Messung der Anbieteranpassung.", "checked": "2026-10-02", "source": { "title": "Artificial Analysis: Deepgram Nova-3 Realtime", "url": "https://artificialanalysis.ai/speech-to-text/streaming" }, "authority": "independent", "rating": "Solide" }, { "id": "aa-deepgram-flux", "model": "Deepgram Flux General EN", "aliases": [ "Deepgram Flux", "Flux" ], "kind": "speech", "benchmark": "Artificial Analysis Word Error Rate Streaming Index (Final Transcription)", "version": "2.2", "value": 7.4, "unit": "%", "direction": "lower", "configuration": "Streaming, Deepgram API; API-Modell flux-general-en", "scope": "Englischsprachiger Streamingtest von Artificial Analysis, WER der finalen Transkripte. Keine medizinischen oder deutschen Aufnahmen, keine Messung der Anbieteranpassung.", "checked": "2026-10-02", "source": { "title": "Artificial Analysis: Deepgram Flux", "url": "https://artificialanalysis.ai/speech-to-text/streaming" }, "authority": "independent", "rating": "Solide" }, { "id": "aa-elevenlabs-scribe-v2", "model": "ElevenLabs Scribe v2", "aliases": [ "Scribe v2" ], "kind": "speech", "benchmark": "Artificial Analysis Word Error Rate Index", "version": "2.2", "value": 2.2, "unit": "%", "direction": "lower", "configuration": "ohne Streaming, ElevenLabs API; API-Modell scribe_v2", "scope": "Englischsprachiger Modelltest von Artificial Analysis (AA-AgentTalk 50 %, VoxPopuli 25 %, Earnings22 25 %). Keine medizinischen oder deutschen Aufnahmen, keine Messung der Anbieteranpassung.", "checked": "2026-10-02", "source": { "title": "Artificial Analysis: Scribe v2, ElevenLabs", "url": "https://artificialanalysis.ai/speech-to-text/non-streaming" }, "authority": "independent", "rating": "Spitzenklasse" }, { "id": "aa-elevenlabs-scribe-v2-realtime", "model": "ElevenLabs Scribe v2 Realtime", "aliases": [ "Scribe v2 Realtime" ], "kind": "speech", "benchmark": "Artificial Analysis Word Error Rate Streaming Index (Final Transcription)", "version": "2.2", "value": 3.6, "unit": "%", "direction": "lower", "configuration": "Streaming, ElevenLabs API; API-Modell scribe_v2_realtime", "scope": "Englischsprachiger Streamingtest von Artificial Analysis, WER der finalen Transkripte. Keine medizinischen oder deutschen Aufnahmen, keine Messung der Anbieteranpassung.", "checked": "2026-10-02", "source": { "title": "Artificial Analysis: ElevenLabs Scribe v2 Realtime", "url": "https://artificialanalysis.ai/speech-to-text/streaming" }, "authority": "independent", "rating": "Stark" }, { "id": "aa-speechmatics-enhanced", "model": "Speechmatics Enhanced", "aliases": [ "Speechmatics Enhanced" ], "kind": "speech", "benchmark": "Artificial Analysis Word Error Rate Index", "version": "2.2", "value": 4, "unit": "%", "direction": "lower", "configuration": "ohne Streaming, Speechmatics API; API-Modell enhanced; Rohwert 4,05 %, im Index gerundet 4,0 %", "scope": "Englischsprachiger Modelltest von Artificial Analysis (AA-AgentTalk 50 %, VoxPopuli 25 %, Earnings22 25 %). Keine medizinischen oder deutschen Aufnahmen, keine Messung der Anbieteranpassung.", "checked": "2026-10-02", "source": { "title": "Artificial Analysis: Enhanced, Speechmatics", "url": "https://artificialanalysis.ai/speech-to-text/non-streaming" }, "authority": "independent", "rating": "Stark" }, { "id": "aa-speechmatics-melia", "model": "Speechmatics Melia", "aliases": [ "Melia" ], "kind": "speech", "benchmark": "Artificial Analysis Word Error Rate Index", "version": "2.2", "value": 4.9, "unit": "%", "direction": "lower", "configuration": "ohne Streaming, Speechmatics API; API-Modell melia-1", "scope": "Englischsprachiger Modelltest von Artificial Analysis (AA-AgentTalk 50 %, VoxPopuli 25 %, Earnings22 25 %). Keine medizinischen oder deutschen Aufnahmen, keine Messung der Anbieteranpassung.", "checked": "2026-10-02", "source": { "title": "Artificial Analysis: Melia, Speechmatics", "url": "https://artificialanalysis.ai/speech-to-text/non-streaming" }, "authority": "independent", "rating": "Solide" }, { "id": "aa-speechmatics-standard", "model": "Speechmatics Standard", "aliases": [ "Speechmatics Standard" ], "kind": "speech", "benchmark": "Artificial Analysis Word Error Rate Index", "version": "2.2", "value": 5.1, "unit": "%", "direction": "lower", "configuration": "ohne Streaming, Speechmatics API; API-Modell standard", "scope": "Englischsprachiger Modelltest von Artificial Analysis (AA-AgentTalk 50 %, VoxPopuli 25 %, Earnings22 25 %). Keine medizinischen oder deutschen Aufnahmen, keine Messung der Anbieteranpassung.", "checked": "2026-10-02", "source": { "title": "Artificial Analysis: Standard, Speechmatics", "url": "https://artificialanalysis.ai/speech-to-text/non-streaming" }, "authority": "independent", "rating": "Solide" }, { "id": "aa-speechmatics-realtime-enhanced", "model": "Speechmatics Realtime (Enhanced)", "aliases": [ "Speechmatics Realtime" ], "kind": "speech", "benchmark": "Artificial Analysis Word Error Rate Streaming Index (Final Transcription)", "version": "2.2", "value": 8, "unit": "%", "direction": "lower", "configuration": "Streaming, Speechmatics API; API-Modell enhanced", "scope": "Englischsprachiger Streamingtest von Artificial Analysis, WER der finalen Transkripte. Keine medizinischen oder deutschen Aufnahmen, keine Messung der Anbieteranpassung.", "checked": "2026-10-02", "source": { "title": "Artificial Analysis: Speechmatics Realtime", "url": "https://artificialanalysis.ai/speech-to-text/streaming" }, "authority": "independent", "rating": "Solide" }, { "id": "aa-speechmatics-agent-stt-linden-1", "model": "Speechmatics Agent STT Linden 1", "aliases": [ "Linden 1" ], "kind": "speech", "benchmark": "Artificial Analysis Word Error Rate Streaming Index (Final Transcription)", "version": "2.2", "value": 4.4, "unit": "%", "direction": "lower", "configuration": "Streaming, Speechmatics API; API-Modell linden-1", "scope": "Englischsprachiger Streamingtest von Artificial Analysis, WER der finalen Transkripte. Keine medizinischen oder deutschen Aufnahmen, keine Messung der Anbieteranpassung.", "checked": "2026-10-02", "source": { "title": "Artificial Analysis: Speechmatics Agent STT Linden 1", "url": "https://artificialanalysis.ai/speech-to-text/streaming" }, "authority": "independent", "rating": "Stark" }, { "id": "aa-soniox-v5-async", "model": "Soniox v5 Async", "aliases": [ "Soniox v5", "Soniox stt-async-v5" ], "kind": "speech", "benchmark": "Artificial Analysis Word Error Rate Index", "version": "2.2", "value": 3.8, "unit": "%", "direction": "lower", "configuration": "ohne Streaming, Soniox API; API-Modell stt-async-v5", "scope": "Englischsprachiger Modelltest von Artificial Analysis (AA-AgentTalk 50 %, VoxPopuli 25 %, Earnings22 25 %). Keine medizinischen oder deutschen Aufnahmen, keine Messung der Anbieteranpassung.", "checked": "2026-10-02", "source": { "title": "Artificial Analysis: Soniox v5 Async, Soniox", "url": "https://artificialanalysis.ai/speech-to-text/non-streaming" }, "authority": "independent", "rating": "Stark" }, { "id": "aa-soniox-v4", "model": "Soniox v4 Async", "aliases": [ "Soniox V4", "Soniox stt-async-v4" ], "kind": "speech", "benchmark": "Artificial Analysis Word Error Rate Index", "version": "2.2", "value": 3.9, "unit": "%", "direction": "lower", "configuration": "ohne Streaming, Soniox API; API-Modell stt-async-v4", "scope": "Englischsprachiger Modelltest von Artificial Analysis (AA-AgentTalk 50 %, VoxPopuli 25 %, Earnings22 25 %). Keine medizinischen oder deutschen Aufnahmen, keine Messung der Anbieteranpassung.", "checked": "2026-10-02", "source": { "title": "Artificial Analysis: Soniox V4, Soniox", "url": "https://artificialanalysis.ai/speech-to-text/non-streaming" }, "authority": "independent", "rating": "Stark" }, { "id": "aa-soniox-v5-realtime", "model": "Soniox v5 Real-Time", "aliases": [ "Soniox v5 Realtime", "Soniox stt-rt-v5" ], "kind": "speech", "benchmark": "Artificial Analysis Word Error Rate Streaming Index (Final Transcription)", "version": "2.2", "value": 4.5, "unit": "%", "direction": "lower", "configuration": "Streaming, Soniox API; API-Modell stt-rt-v5", "scope": "Englischsprachiger Streamingtest von Artificial Analysis, WER der finalen Transkripte. Keine medizinischen oder deutschen Aufnahmen, keine Messung der Anbieteranpassung.", "checked": "2026-10-02", "source": { "title": "Artificial Analysis: Soniox v5 Real-Time", "url": "https://artificialanalysis.ai/speech-to-text/streaming" }, "authority": "independent", "rating": "Stark" }, { "id": "aa-soniox-v4-realtime", "model": "Soniox v4 Real-Time", "aliases": [ "Soniox v4 Realtime", "Soniox stt-rt-v4" ], "kind": "speech", "benchmark": "Artificial Analysis Word Error Rate Streaming Index (Final Transcription)", "version": "2.2", "value": 4.5, "unit": "%", "direction": "lower", "configuration": "Streaming, Soniox API; API-Modell stt-rt-v4", "scope": "Englischsprachiger Streamingtest von Artificial Analysis, WER der finalen Transkripte. Keine medizinischen oder deutschen Aufnahmen, keine Messung der Anbieteranpassung.", "checked": "2026-10-02", "source": { "title": "Artificial Analysis: Soniox v4 Real-Time", "url": "https://artificialanalysis.ai/speech-to-text/streaming" }, "authority": "independent", "rating": "Stark" }, { "id": "aa-google-chirp-3", "model": "Google Chirp 3", "aliases": [ "Chirp 3", "Google Cloud Speech-to-Text Chirp 3" ], "kind": "speech", "benchmark": "Artificial Analysis Word Error Rate Index", "version": "2.2", "value": 4.3, "unit": "%", "direction": "lower", "configuration": "ohne Streaming, Google API; API-Modell chirp_3", "scope": "Englischsprachiger Modelltest von Artificial Analysis (AA-AgentTalk 50 %, VoxPopuli 25 %, Earnings22 25 %). Keine medizinischen oder deutschen Aufnahmen, keine Messung der Anbieteranpassung.", "checked": "2026-10-02", "source": { "title": "Artificial Analysis: Chirp 3, Google", "url": "https://artificialanalysis.ai/speech-to-text/non-streaming" }, "authority": "independent", "rating": "Solide" }, { "id": "aa-google-chirp-2", "model": "Google Chirp 2", "aliases": [ "Chirp 2", "Google Cloud Speech-to-Text Chirp 2" ], "kind": "speech", "benchmark": "Artificial Analysis Word Error Rate Index", "version": "2.2", "value": 5.8, "unit": "%", "direction": "lower", "configuration": "ohne Streaming; Wert aus dem Modelldiagramm „Chirp 2, Google“, ohne Zeile in der API-Tabelle; im Diagramm nicht standardmäßig ausgewählt", "scope": "Englischsprachiger Modelltest von Artificial Analysis (AA-AgentTalk 50 %, VoxPopuli 25 %, Earnings22 25 %). Keine medizinischen oder deutschen Aufnahmen, keine Messung der Anbieteranpassung.", "checked": "2026-10-02", "source": { "title": "Artificial Analysis: Chirp 2, Google", "url": "https://artificialanalysis.ai/speech-to-text/non-streaming" }, "authority": "independent", "rating": "Solide" }, { "id": "aa-google-chirp", "model": "Google Chirp (erste Generation)", "aliases": [ "Chirp 1", "Chirp v1" ], "kind": "speech", "benchmark": "Artificial Analysis Word Error Rate Index", "version": "2.2", "value": 31.2, "unit": "%", "direction": "lower", "configuration": "ohne Streaming, Google API; API-Modell chirp; im Index als „Cloud Speech-To-Text (Chirp)“ geführt, veröffentlicht am 2023-05-01", "scope": "Englischsprachiger Modelltest von Artificial Analysis (AA-AgentTalk 50 %, VoxPopuli 25 %, Earnings22 25 %). Keine medizinischen oder deutschen Aufnahmen, keine Messung der Anbieteranpassung.", "checked": "2026-10-02", "source": { "title": "Artificial Analysis: Chirp, Google", "url": "https://artificialanalysis.ai/speech-to-text/non-streaming" }, "authority": "independent", "rating": "Schwach" }, { "id": "aa-google-chirp-3-streaming", "model": "Google Chirp 3 Streaming", "aliases": [ "Chirp 3 Streaming" ], "kind": "speech", "benchmark": "Artificial Analysis Word Error Rate Streaming Index (Final Transcription)", "version": "2.2", "value": 4.8, "unit": "%", "direction": "lower", "configuration": "Streaming, Google API; API-Modell chirp_3", "scope": "Englischsprachiger Streamingtest von Artificial Analysis, WER der finalen Transkripte. Keine medizinischen oder deutschen Aufnahmen, keine Messung der Anbieteranpassung.", "checked": "2026-10-02", "source": { "title": "Artificial Analysis: Chirp 3 Streaming", "url": "https://artificialanalysis.ai/speech-to-text/streaming" }, "authority": "independent", "rating": "Solide" }, { "id": "aa-gemini-3-5-transcribe", "model": "Gemini 3.5 Transcribe", "aliases": [ "Gemini 3.5 Transcribe" ], "kind": "speech", "benchmark": "Artificial Analysis Word Error Rate Index", "version": "2.2", "value": 2.6, "unit": "%", "direction": "lower", "configuration": "ohne Streaming, Google API; API-Modell gemini-3.5-transcribe", "scope": "Englischsprachiger Modelltest von Artificial Analysis (AA-AgentTalk 50 %, VoxPopuli 25 %, Earnings22 25 %). Keine medizinischen oder deutschen Aufnahmen, keine Messung der Anbieteranpassung.", "checked": "2026-10-02", "source": { "title": "Artificial Analysis: Gemini 3.5 Transcribe, Google", "url": "https://artificialanalysis.ai/speech-to-text/non-streaming" }, "authority": "independent", "rating": "Stark" }, { "id": "aa-gemini-3-1-pro-preview-high-stt", "model": "Gemini 3.1 Pro Preview (Audio-Transkription)", "aliases": [ "Gemini 3.1 Pro Preview" ], "kind": "speech", "benchmark": "Artificial Analysis Word Error Rate Index", "version": "2.2", "value": 2.8, "unit": "%", "direction": "lower", "configuration": "ohne Streaming, Google API; API-Modell gemini-3.1-pro-preview; Reasoning-Stufe hoch", "scope": "Englischsprachiger Modelltest von Artificial Analysis (AA-AgentTalk 50 %, VoxPopuli 25 %, Earnings22 25 %). Keine medizinischen oder deutschen Aufnahmen, keine Messung der Anbieteranpassung.", "checked": "2026-10-02", "source": { "title": "Artificial Analysis: Gemini 3.1 Pro Preview (High), Google", "url": "https://artificialanalysis.ai/speech-to-text/non-streaming" }, "authority": "independent", "rating": "Stark" }, { "id": "aa-gemini-3-flash-high-stt", "model": "Gemini 3 Flash (Audio-Transkription)", "aliases": [ "Gemini 3 Flash" ], "kind": "speech", "benchmark": "Artificial Analysis Word Error Rate Index", "version": "2.2", "value": 2.9, "unit": "%", "direction": "lower", "configuration": "ohne Streaming, Google API; API-Modell gemini-3-flash-preview; Reasoning-Stufe hoch", "scope": "Englischsprachiger Modelltest von Artificial Analysis (AA-AgentTalk 50 %, VoxPopuli 25 %, Earnings22 25 %). Keine medizinischen oder deutschen Aufnahmen, keine Messung der Anbieteranpassung.", "checked": "2026-10-02", "source": { "title": "Artificial Analysis: Gemini 3 Flash (High), Google", "url": "https://artificialanalysis.ai/speech-to-text/non-streaming" }, "authority": "independent", "rating": "Stark" }, { "id": "aa-azure-llm-speech", "model": "Azure LLM Speech", "aliases": [ "LLM Speech", "Azure LLM Speech" ], "kind": "speech", "benchmark": "Artificial Analysis Word Error Rate Index", "version": "2.2", "value": 3.7, "unit": "%", "direction": "lower", "configuration": "ohne Streaming; Wert aus dem Modelldiagramm „LLM Speech, Azure“, ohne Zeile in der API-Tabelle; im Diagramm nicht standardmäßig ausgewählt", "scope": "Englischsprachiger Modelltest von Artificial Analysis (AA-AgentTalk 50 %, VoxPopuli 25 %, Earnings22 25 %). Keine medizinischen oder deutschen Aufnahmen, keine Messung der Anbieteranpassung.", "checked": "2026-10-02", "source": { "title": "Artificial Analysis: LLM Speech, Azure", "url": "https://artificialanalysis.ai/speech-to-text/non-streaming" }, "authority": "independent", "rating": "Stark" }, { "id": "aa-mai-transcribe-2", "model": "Microsoft MAI-Transcribe-2", "aliases": [ "MAI-Transcribe-2" ], "kind": "speech", "benchmark": "Artificial Analysis Word Error Rate Index", "version": "2.2", "value": 2, "unit": "%", "direction": "lower", "configuration": "ohne Streaming, Microsoft AI API; API-Modell mai-transcribe-15-e1; Artificial Analysis ordnet das Modell der Familie Azure Speech Service zu", "scope": "Englischsprachiger Modelltest von Artificial Analysis (AA-AgentTalk 50 %, VoxPopuli 25 %, Earnings22 25 %). Keine medizinischen oder deutschen Aufnahmen, keine Messung der Anbieteranpassung.", "checked": "2026-10-02", "source": { "title": "Artificial Analysis: MAI-Transcribe-2, Microsoft AI", "url": "https://artificialanalysis.ai/speech-to-text/non-streaming" }, "authority": "independent", "rating": "Spitzenklasse" }, { "id": "aa-mai-transcribe-1-5", "model": "Microsoft MAI-Transcribe-1.5", "aliases": [ "MAI-Transcribe-1.5" ], "kind": "speech", "benchmark": "Artificial Analysis Word Error Rate Index", "version": "2.2", "value": 2.4, "unit": "%", "direction": "lower", "configuration": "ohne Streaming, Microsoft AI API; API-Modell mai-transcribe-1.5; Artificial Analysis ordnet das Modell der Familie Azure Speech Service zu", "scope": "Englischsprachiger Modelltest von Artificial Analysis (AA-AgentTalk 50 %, VoxPopuli 25 %, Earnings22 25 %). Keine medizinischen oder deutschen Aufnahmen, keine Messung der Anbieteranpassung.", "checked": "2026-10-02", "source": { "title": "Artificial Analysis: MAI-Transcribe-1.5, Microsoft AI", "url": "https://artificialanalysis.ai/speech-to-text/non-streaming" }, "authority": "independent", "rating": "Stark" }, { "id": "aa-azure-stt-realtime", "model": "Azure AI Speech, Echtzeit-Transkription", "aliases": [], "kind": "speech", "benchmark": "Artificial Analysis Word Error Rate Streaming Index (Final Transcription)", "version": "2.2", "value": 5.2, "unit": "%", "direction": "lower", "configuration": "Streaming, Microsoft Azure API; API-Modell azure.cognitiveservices.speech (Real time, 1.31.0); Speech SDK 1.31.0", "scope": "Englischsprachiger Streamingtest von Artificial Analysis, WER der finalen Transkripte. Keine medizinischen oder deutschen Aufnahmen, keine Messung der Anbieteranpassung.", "checked": "2026-10-02", "source": { "title": "Artificial Analysis: Azure STT Real-time Transcription", "url": "https://artificialanalysis.ai/speech-to-text/streaming" }, "authority": "independent", "rating": "Solide" }, { "id": "aa-mai-transcribe-2-streaming", "model": "Microsoft MAI-Transcribe-2 Streaming", "aliases": [ "MAI-Transcribe-2-Streaming" ], "kind": "speech", "benchmark": "Artificial Analysis Word Error Rate Streaming Index (Final Transcription)", "version": "2.2", "value": 2.5, "unit": "%", "direction": "lower", "configuration": "Streaming, Microsoft AI API; API-Modell mai-transcribe-2-streaming-aa-g4", "scope": "Englischsprachiger Streamingtest von Artificial Analysis, WER der finalen Transkripte. Keine medizinischen oder deutschen Aufnahmen, keine Messung der Anbieteranpassung.", "checked": "2026-10-02", "source": { "title": "Artificial Analysis: MAI-Transcribe-2-Streaming", "url": "https://artificialanalysis.ai/speech-to-text/streaming" }, "authority": "independent", "rating": "Spitzenklasse" }, { "id": "aa-parakeet-tdt-0-6b-v2", "model": "NVIDIA Parakeet TDT 0.6B v2", "aliases": [ "Parakeet TDT 0.6B V2", "Parakeet V2 0.6B TDT", "nvidia/parakeet-tdt-0.6b-v2" ], "kind": "speech", "benchmark": "Artificial Analysis Word Error Rate Index", "version": "2.2", "value": 6.4, "unit": "%", "direction": "lower", "configuration": "ohne Streaming, NVIDIA API; API-Modell d3fe9151-442b-4204-a70d-5fcc597fd610; Basismodell ohne Feinabstimmung", "scope": "Englischsprachiger Modelltest von Artificial Analysis (AA-AgentTalk 50 %, VoxPopuli 25 %, Earnings22 25 %). Keine medizinischen oder deutschen Aufnahmen, keine Messung der Anbieteranpassung.", "checked": "2026-10-02", "source": { "title": "Artificial Analysis: Parakeet TDT 0.6B V2, NVIDIA", "url": "https://artificialanalysis.ai/speech-to-text/non-streaming" }, "authority": "independent", "rating": "Solide" }, { "id": "aa-parakeet-rnnt-1-1b", "model": "NVIDIA Parakeet RNNT 1.1B", "aliases": [ "Parakeet RNNT 1.1B" ], "kind": "speech", "benchmark": "Artificial Analysis Word Error Rate Index", "version": "2.2", "value": 5.4, "unit": "%", "direction": "lower", "configuration": "ohne Streaming, Replicate API; API-Modell nvidia/parakeet-rnnt-1.1b:73ddbebaef172a47c8dfdd79381f110bfdc7691bcc7a4edde82f0a39e380ce50; Hosting durch Replicate", "scope": "Englischsprachiger Modelltest von Artificial Analysis (AA-AgentTalk 50 %, VoxPopuli 25 %, Earnings22 25 %). Keine medizinischen oder deutschen Aufnahmen, keine Messung der Anbieteranpassung.", "checked": "2026-10-02", "source": { "title": "Artificial Analysis: Parakeet RNNT 1.1B, Replicate, Replicate", "url": "https://artificialanalysis.ai/speech-to-text/non-streaming" }, "authority": "independent", "rating": "Solide" }, { "id": "aa-canary-qwen-2-5b", "model": "NVIDIA Canary Qwen 2.5B", "aliases": [ "Canary Qwen 2.5B" ], "kind": "speech", "benchmark": "Artificial Analysis Word Error Rate Index", "version": "2.2", "value": 4.3, "unit": "%", "direction": "lower", "configuration": "ohne Streaming, Replicate API; API-Modell nvidia/canary-qwen-2.5b:afba731fc7a4082730943a246233b09c7fa3dfb2c24b07fe199c1408a7c8cb2f; Hosting durch Replicate", "scope": "Englischsprachiger Modelltest von Artificial Analysis (AA-AgentTalk 50 %, VoxPopuli 25 %, Earnings22 25 %). Keine medizinischen oder deutschen Aufnahmen, keine Messung der Anbieteranpassung.", "checked": "2026-10-02", "source": { "title": "Artificial Analysis: Canary Qwen 2.5B, Replicate, Replicate", "url": "https://artificialanalysis.ai/speech-to-text/non-streaming" }, "authority": "independent", "rating": "Solide" }, { "id": "aa-nemotron-3-asr-streaming-1120ms", "model": "NVIDIA Nemotron 3.5 ASR Streaming 0.6B (1120 ms)", "aliases": [ "Nemotron ASR Streaming", "NVIDIA-Nemotron-3.5-ASR-Streaming-Multilingual-0.6b" ], "kind": "speech", "benchmark": "Artificial Analysis Word Error Rate Streaming Index (Final Transcription)", "version": "2.2", "value": 5.4, "unit": "%", "direction": "lower", "configuration": "Streaming, NVIDIA API; API-Modell cache-aware-parakeet-rnnt-en-US-asr-streaming-1120ms; Latenzstufe 1120 ms", "scope": "Englischsprachiger Streamingtest von Artificial Analysis, WER der finalen Transkripte. Keine medizinischen oder deutschen Aufnahmen, keine Messung der Anbieteranpassung.", "checked": "2026-10-02", "source": { "title": "Artificial Analysis: Nemotron 3 ASR 1120ms", "url": "https://artificialanalysis.ai/speech-to-text/streaming" }, "authority": "independent", "rating": "Solide" }, { "id": "aa-nemotron-3-asr-streaming-160ms", "model": "NVIDIA Nemotron 3.5 ASR Streaming 0.6B (160 ms)", "aliases": [ "Nemotron ASR Streaming", "NVIDIA-Nemotron-3.5-ASR-Streaming-Multilingual-0.6b" ], "kind": "speech", "benchmark": "Artificial Analysis Word Error Rate Streaming Index (Final Transcription)", "version": "2.2", "value": 7.6, "unit": "%", "direction": "lower", "configuration": "Streaming, NVIDIA API; API-Modell cache-aware-parakeet-rnnt-en-US-asr-streaming-160ms; Latenzstufe 160 ms", "scope": "Englischsprachiger Streamingtest von Artificial Analysis, WER der finalen Transkripte. Keine medizinischen oder deutschen Aufnahmen, keine Messung der Anbieteranpassung.", "checked": "2026-10-02", "source": { "title": "Artificial Analysis: Nemotron 3 ASR 160ms", "url": "https://artificialanalysis.ai/speech-to-text/streaming" }, "authority": "independent", "rating": "Solide" }, { "id": "aa-gpt-transcribe", "model": "OpenAI GPT Transcribe", "aliases": [ "GPT Transcribe", "gpt-transcribe" ], "kind": "speech", "benchmark": "Artificial Analysis Word Error Rate Index", "version": "2.2", "value": 3.3, "unit": "%", "direction": "lower", "configuration": "ohne Streaming, OpenAI API; API-Modell gpt-transcribe", "scope": "Englischsprachiger Modelltest von Artificial Analysis (AA-AgentTalk 50 %, VoxPopuli 25 %, Earnings22 25 %). Keine medizinischen oder deutschen Aufnahmen, keine Messung der Anbieteranpassung.", "checked": "2026-10-02", "source": { "title": "Artificial Analysis: GPT Transcribe, OpenAI", "url": "https://artificialanalysis.ai/speech-to-text/non-streaming" }, "authority": "independent", "rating": "Stark" }, { "id": "aa-gpt-4o-transcribe", "model": "OpenAI GPT-4o Transcribe", "aliases": [ "GPT-4o Transcribe", "gpt-4o-transcribe" ], "kind": "speech", "benchmark": "Artificial Analysis Word Error Rate Index", "version": "2.2", "value": 4, "unit": "%", "direction": "lower", "configuration": "ohne Streaming, OpenAI API; API-Modell gpt-4o-transcribe", "scope": "Englischsprachiger Modelltest von Artificial Analysis (AA-AgentTalk 50 %, VoxPopuli 25 %, Earnings22 25 %). Keine medizinischen oder deutschen Aufnahmen, keine Messung der Anbieteranpassung.", "checked": "2026-10-02", "source": { "title": "Artificial Analysis: GPT-4o Transcribe, OpenAI", "url": "https://artificialanalysis.ai/speech-to-text/non-streaming" }, "authority": "independent", "rating": "Stark" }, { "id": "aa-gpt-4o-mini-transcribe", "model": "OpenAI GPT-4o Mini Transcribe", "aliases": [ "GPT-4o Mini Transcribe", "gpt-4o-mini-transcribe" ], "kind": "speech", "benchmark": "Artificial Analysis Word Error Rate Index", "version": "2.2", "value": 4.5, "unit": "%", "direction": "lower", "configuration": "ohne Streaming, OpenAI API; API-Modell gpt-4o-mini-transcribe", "scope": "Englischsprachiger Modelltest von Artificial Analysis (AA-AgentTalk 50 %, VoxPopuli 25 %, Earnings22 25 %). Keine medizinischen oder deutschen Aufnahmen, keine Messung der Anbieteranpassung.", "checked": "2026-10-02", "source": { "title": "Artificial Analysis: GPT-4o Mini Transcribe, OpenAI", "url": "https://artificialanalysis.ai/speech-to-text/non-streaming" }, "authority": "independent", "rating": "Solide" }, { "id": "aa-gpt-realtime-whisper", "model": "OpenAI GPT Realtime Whisper", "aliases": [ "gpt-realtime-whisper" ], "kind": "speech", "benchmark": "Artificial Analysis Word Error Rate Streaming Index (Final Transcription)", "version": "2.2", "value": 4.9, "unit": "%", "direction": "lower", "configuration": "Streaming, OpenAI API; API-Modell gpt-realtime-whisper", "scope": "Englischsprachiger Streamingtest von Artificial Analysis, WER der finalen Transkripte. Keine medizinischen oder deutschen Aufnahmen, keine Messung der Anbieteranpassung.", "checked": "2026-10-02", "source": { "title": "Artificial Analysis: OpenAI GPT Realtime Whisper", "url": "https://artificialanalysis.ai/speech-to-text/streaming" }, "authority": "independent", "rating": "Solide" }, { "id": "aa-gpt-live-transcribe", "model": "OpenAI GPT Live Transcribe", "aliases": [ "gpt-live-transcribe" ], "kind": "speech", "benchmark": "Artificial Analysis Word Error Rate Streaming Index (Final Transcription)", "version": "2.2", "value": 3.9, "unit": "%", "direction": "lower", "configuration": "Streaming, OpenAI API; API-Modell gpt-live-transcribe", "scope": "Englischsprachiger Streamingtest von Artificial Analysis, WER der finalen Transkripte. Keine medizinischen oder deutschen Aufnahmen, keine Messung der Anbieteranpassung.", "checked": "2026-10-02", "source": { "title": "Artificial Analysis: GPT Live Transcribe", "url": "https://artificialanalysis.ai/speech-to-text/streaming" }, "authority": "independent", "rating": "Stark" }, { "id": "aa-voxtral-small", "model": "Mistral Voxtral Small", "aliases": [ "Voxtral Small", "voxtral-small-2507" ], "kind": "speech", "benchmark": "Artificial Analysis Word Error Rate Index", "version": "2.2", "value": 2.8, "unit": "%", "direction": "lower", "configuration": "ohne Streaming, Mistral API; API-Modell voxtral-small-2507", "scope": "Englischsprachiger Modelltest von Artificial Analysis (AA-AgentTalk 50 %, VoxPopuli 25 %, Earnings22 25 %). Keine medizinischen oder deutschen Aufnahmen, keine Messung der Anbieteranpassung.", "checked": "2026-10-02", "source": { "title": "Artificial Analysis: Voxtral Small, Mistral", "url": "https://artificialanalysis.ai/speech-to-text/non-streaming" }, "authority": "independent", "rating": "Stark" }, { "id": "aa-voxtral-mini-transcribe-2", "model": "Mistral Voxtral Mini Transcribe 2", "aliases": [ "Voxtral Mini Transcribe 2", "voxtral-mini-2602" ], "kind": "speech", "benchmark": "Artificial Analysis Word Error Rate Index", "version": "2.2", "value": 3.6, "unit": "%", "direction": "lower", "configuration": "ohne Streaming, Mistral API; API-Modell voxtral-mini-2602", "scope": "Englischsprachiger Modelltest von Artificial Analysis (AA-AgentTalk 50 %, VoxPopuli 25 %, Earnings22 25 %). Keine medizinischen oder deutschen Aufnahmen, keine Messung der Anbieteranpassung.", "checked": "2026-10-02", "source": { "title": "Artificial Analysis: Voxtral Mini Transcribe 2, Mistral", "url": "https://artificialanalysis.ai/speech-to-text/non-streaming" }, "authority": "independent", "rating": "Stark" }, { "id": "aa-claude-opus-5-5-max", "model": "Claude Opus 5.5", "aliases": [ "Claude Opus 5.5", "Claude Opus 5.5 (Max, Default Fallback)" ], "kind": "language", "benchmark": "Artificial Analysis Intelligence Index", "version": "4.3.2", "value": 58, "unit": "points", "direction": "higher", "configuration": "Reasoning-Stufe maximal, mit Standard-Rückfallmodell", "scope": "Allgemeine Modellbewertung. Produktkonfiguration und klinische Ausgabequalität sind dadurch nicht belegt.", "checked": "2026-10-02", "source": { "title": "Artificial Analysis: Claude Opus 5.5 (Max, Default Fallback)", "url": "https://artificialanalysis.ai/models/claude-opus-5-5" }, "authority": "independent", "rating": "Spitzenklasse" }, { "id": "aa-claude-sonnet-5-5-max", "model": "Claude Sonnet 5.5", "aliases": [ "Claude Sonnet 5.5", "Claude Sonnet 5.5 (Max, Default Fallback)" ], "kind": "language", "benchmark": "Artificial Analysis Intelligence Index", "version": "4.3.2", "value": 56, "unit": "points", "direction": "higher", "configuration": "Reasoning-Stufe maximal, mit Standard-Rückfallmodell", "scope": "Allgemeine Modellbewertung. Produktkonfiguration und klinische Ausgabequalität sind dadurch nicht belegt.", "checked": "2026-10-02", "source": { "title": "Artificial Analysis: Claude Sonnet 5.5 (Max, Default Fallback)", "url": "https://artificialanalysis.ai/models/claude-sonnet-5-5" }, "authority": "independent", "rating": "Spitzenklasse" }, { "id": "aa-claude-fable-5-1-max", "model": "Claude Fable 5.1", "aliases": [ "Claude Fable 5.1", "Claude Fable 5.1 (Max, Default Fallback)" ], "kind": "language", "benchmark": "Artificial Analysis Intelligence Index", "version": "4.3.2", "value": 53, "unit": "points", "direction": "higher", "configuration": "Reasoning-Stufe maximal, mit Standard-Rückfallmodell", "scope": "Allgemeine Modellbewertung. Produktkonfiguration und klinische Ausgabequalität sind dadurch nicht belegt.", "checked": "2026-10-02", "source": { "title": "Artificial Analysis: Claude Fable 5.1 (Max, Default Fallback)", "url": "https://artificialanalysis.ai/models/claude-fable-5-1" }, "authority": "independent", "rating": "Spitzenklasse" }, { "id": "aa-claude-haiku-4-5-reasoning", "model": "Claude Haiku 4.5", "aliases": [ "Claude 4.5 Haiku", "Claude Haiku 4.5", "Claude 4.5 Haiku (Reasoning)" ], "kind": "language", "benchmark": "Artificial Analysis Intelligence Index", "version": "4.3.2", "value": 17, "unit": "points", "direction": "higher", "configuration": "mit Reasoning", "scope": "Allgemeine Modellbewertung. Produktkonfiguration und klinische Ausgabequalität sind dadurch nicht belegt.", "checked": "2026-10-02", "source": { "title": "Artificial Analysis: Claude 4.5 Haiku (Reasoning)", "url": "https://artificialanalysis.ai/models/claude-4-5-haiku-reasoning" }, "authority": "independent", "rating": "Solide" }, { "id": "aa-gpt-6-1-sol-max", "model": "GPT-6.1 Sol", "aliases": [ "GPT-6.1 Sol", "GPT-6.1 Sol (Max)" ], "kind": "language", "benchmark": "Artificial Analysis Intelligence Index", "version": "4.3.2", "value": 52, "unit": "points", "direction": "higher", "configuration": "Reasoning-Stufe maximal", "scope": "Allgemeine Modellbewertung. Produktkonfiguration und klinische Ausgabequalität sind dadurch nicht belegt.", "checked": "2026-10-02", "source": { "title": "Artificial Analysis: GPT-6.1 Sol (Max)", "url": "https://artificialanalysis.ai/models/gpt-6-1-sol" }, "authority": "independent", "rating": "Spitzenklasse" }, { "id": "aa-gpt-5-6-terra-max", "model": "GPT-5.6 Terra", "aliases": [ "GPT-5.6 Terra", "GPT-5.6 Terra (Max)" ], "kind": "language", "benchmark": "Artificial Analysis Intelligence Index", "version": "4.3.2", "value": 42, "unit": "points", "direction": "higher", "configuration": "Reasoning-Stufe maximal", "scope": "Allgemeine Modellbewertung. Produktkonfiguration und klinische Ausgabequalität sind dadurch nicht belegt.", "checked": "2026-10-02", "source": { "title": "Artificial Analysis: GPT-5.6 Terra (Max)", "url": "https://artificialanalysis.ai/models/gpt-5-6-terra" }, "authority": "independent", "rating": "Stark" }, { "id": "aa-gpt-5-5-xhigh", "model": "GPT-5.5", "aliases": [ "GPT-5.5", "OpenAI GPT-5.5", "GPT-5.5 (Xhigh)" ], "kind": "language", "benchmark": "Artificial Analysis Intelligence Index", "version": "4.3.2", "value": 38, "unit": "points", "direction": "higher", "configuration": "Reasoning-Stufe sehr hoch; laut Modellseite eingestellt", "scope": "Allgemeine Modellbewertung. Produktkonfiguration und klinische Ausgabequalität sind dadurch nicht belegt.", "checked": "2026-10-02", "source": { "title": "Artificial Analysis: GPT-5.5 (Xhigh)", "url": "https://artificialanalysis.ai/models/gpt-5-5" }, "authority": "independent", "rating": "Stark" }, { "id": "aa-gpt-5-4-xhigh", "model": "GPT-5.4", "aliases": [ "GPT-5.4", "OpenAI GPT-5.4", "GPT-5.4 (Xhigh)" ], "kind": "language", "benchmark": "Artificial Analysis Intelligence Index", "version": "4.3.2", "value": 39, "unit": "points", "direction": "higher", "configuration": "Reasoning-Stufe sehr hoch; von Artificial Analysis als Schätzung markiert, unabhängige Messung ausstehend; laut Modellseite eingestellt", "scope": "Schätzwert von Artificial Analysis, keine eigene Messung. Allgemeine Modellbewertung, Produktkonfiguration und klinische Ausgabequalität sind dadurch nicht belegt.", "checked": "2026-10-02", "source": { "title": "Artificial Analysis: GPT-5.4 (Xhigh)", "url": "https://artificialanalysis.ai/models/gpt-5-4" }, "authority": "independent", "rating": "Stark" }, { "id": "aa-gpt-5-2-xhigh", "model": "GPT-5.2", "aliases": [ "GPT-5.2", "OpenAI GPT-5.2", "GPT-5.2 (Xhigh)" ], "kind": "language", "benchmark": "Artificial Analysis Intelligence Index", "version": "4.3.2", "value": 30, "unit": "points", "direction": "higher", "configuration": "Reasoning-Stufe sehr hoch; von Artificial Analysis als Schätzung markiert, unabhängige Messung ausstehend; laut Modellseite eingestellt", "scope": "Schätzwert von Artificial Analysis, keine eigene Messung. Allgemeine Modellbewertung, Produktkonfiguration und klinische Ausgabequalität sind dadurch nicht belegt.", "checked": "2026-10-02", "source": { "title": "Artificial Analysis: GPT-5.2 (Xhigh)", "url": "https://artificialanalysis.ai/models/gpt-5-2" }, "authority": "independent", "rating": "Stark" }, { "id": "aa-gpt-4o-nov-24", "model": "GPT-4o", "aliases": [ "GPT-4o", "GPT-4o (Nov '24)" ], "kind": "language", "benchmark": "Artificial Analysis Intelligence Index", "version": "4.3.2", "value": 8, "unit": "points", "direction": "higher", "configuration": "Version November 2024, ohne Reasoning; von Artificial Analysis als Schätzung markiert, unabhängige Messung ausstehend; laut Modellseite eingestellt", "scope": "Schätzwert von Artificial Analysis, keine eigene Messung. Allgemeine Modellbewertung, Produktkonfiguration und klinische Ausgabequalität sind dadurch nicht belegt.", "checked": "2026-10-02", "source": { "title": "Artificial Analysis: GPT-4o (Nov '24)", "url": "https://artificialanalysis.ai/models/gpt-4o" }, "authority": "independent", "rating": "Schwach" }, { "id": "aa-gpt-4", "model": "GPT-4", "aliases": [ "GPT-4", "OpenAI GPT-4" ], "kind": "language", "benchmark": "Artificial Analysis Intelligence Index", "version": "4.3.2", "value": 7, "unit": "points", "direction": "higher", "configuration": "ohne Reasoning; von Artificial Analysis als Schätzung markiert, unabhängige Messung ausstehend; laut Modellseite eingestellt", "scope": "Schätzwert von Artificial Analysis, keine eigene Messung. Allgemeine Modellbewertung, Produktkonfiguration und klinische Ausgabequalität sind dadurch nicht belegt.", "checked": "2026-10-02", "source": { "title": "Artificial Analysis: GPT-4", "url": "https://artificialanalysis.ai/models/gpt-4" }, "authority": "independent", "rating": "Schwach" }, { "id": "aa-gemini-4-argon-high", "model": "Gemini 4 Argon", "aliases": [ "Gemini 4 Argon", "Gemini 4 Argon (High)" ], "kind": "language", "benchmark": "Artificial Analysis Intelligence Index", "version": "4.3.2", "value": 53, "unit": "points", "direction": "higher", "configuration": "Reasoning-Stufe hoch", "scope": "Allgemeine Modellbewertung. Produktkonfiguration und klinische Ausgabequalität sind dadurch nicht belegt.", "checked": "2026-10-02", "source": { "title": "Artificial Analysis: Gemini 4 Argon (High)", "url": "https://artificialanalysis.ai/models/gemini-4-argon" }, "authority": "independent", "rating": "Spitzenklasse" }, { "id": "aa-gemini-3-8-flash-high", "model": "Gemini 3.8 Flash", "aliases": [ "Gemini 3.8 Flash", "Gemini 3.8 Flash (High)" ], "kind": "language", "benchmark": "Artificial Analysis Intelligence Index", "version": "4.3.2", "value": 41, "unit": "points", "direction": "higher", "configuration": "Reasoning-Stufe hoch", "scope": "Allgemeine Modellbewertung. Produktkonfiguration und klinische Ausgabequalität sind dadurch nicht belegt.", "checked": "2026-10-02", "source": { "title": "Artificial Analysis: Gemini 3.8 Flash (High)", "url": "https://artificialanalysis.ai/models/gemini-3-8-flash" }, "authority": "independent", "rating": "Stark" }, { "id": "aa-gemini-3-1-pro-preview", "model": "Gemini 3.1 Pro Preview", "aliases": [ "Gemini 3.1 Pro", "Gemini 3.1 Pro Preview" ], "kind": "language", "benchmark": "Artificial Analysis Intelligence Index", "version": "4.3.2", "value": 30, "unit": "points", "direction": "higher", "configuration": "Standardeinstellung", "scope": "Allgemeine Modellbewertung. Produktkonfiguration und klinische Ausgabequalität sind dadurch nicht belegt.", "checked": "2026-10-02", "source": { "title": "Artificial Analysis: Gemini 3.1 Pro Preview", "url": "https://artificialanalysis.ai/models/gemini-3-1-pro-preview" }, "authority": "independent", "rating": "Stark" }, { "id": "aa-gemini-3-5-flash-lite", "model": "Gemini 3.5 Flash-Lite", "aliases": [ "Gemini 3.5 Flash-Lite", "Gemini 3.5 Flash Lite" ], "kind": "language", "benchmark": "Artificial Analysis Intelligence Index", "version": "4.3.2", "value": 22, "unit": "points", "direction": "higher", "configuration": "Standardeinstellung", "scope": "Allgemeine Modellbewertung. Produktkonfiguration und klinische Ausgabequalität sind dadurch nicht belegt.", "checked": "2026-10-02", "source": { "title": "Artificial Analysis: Gemini 3.5 Flash-Lite", "url": "https://artificialanalysis.ai/models/gemini-3-5-flash-lite" }, "authority": "independent", "rating": "Solide" }, { "id": "aa-llama-4-maverick", "model": "Llama 4 Maverick", "aliases": [ "Llama 4 Maverick", "Meta Llama 4 Maverick" ], "kind": "language", "benchmark": "Artificial Analysis Intelligence Index", "version": "4.3.2", "value": 10, "unit": "points", "direction": "higher", "configuration": "ohne Reasoning; von Artificial Analysis als Schätzung markiert, unabhängige Messung ausstehend", "scope": "Schätzwert von Artificial Analysis, keine eigene Messung. Allgemeine Modellbewertung, Produktkonfiguration und klinische Ausgabequalität sind dadurch nicht belegt.", "checked": "2026-10-02", "source": { "title": "Artificial Analysis: Llama 4 Maverick", "url": "https://artificialanalysis.ai/models/llama-4-maverick" }, "authority": "independent", "rating": "Schwach" }, { "id": "aa-llama-4-scout", "model": "Llama 4 Scout", "aliases": [ "Llama 4 Scout", "Meta Llama 4 Scout" ], "kind": "language", "benchmark": "Artificial Analysis Intelligence Index", "version": "4.3.2", "value": 8, "unit": "points", "direction": "higher", "configuration": "ohne Reasoning; von Artificial Analysis als Schätzung markiert, unabhängige Messung ausstehend", "scope": "Schätzwert von Artificial Analysis, keine eigene Messung. Allgemeine Modellbewertung, Produktkonfiguration und klinische Ausgabequalität sind dadurch nicht belegt.", "checked": "2026-10-02", "source": { "title": "Artificial Analysis: Llama 4 Scout", "url": "https://artificialanalysis.ai/models/llama-4-scout" }, "authority": "independent", "rating": "Schwach" }, { "id": "aa-mistral-small-3", "model": "Mistral Small 3", "aliases": [ "Mistral Small 3", "Mistral-Small-24B-Instruct-2501" ], "kind": "language", "benchmark": "Artificial Analysis Intelligence Index", "version": "4.3.2", "value": 7, "unit": "points", "direction": "higher", "configuration": "ohne Reasoning, Basismodell ohne Feinabstimmung; von Artificial Analysis als Schätzung markiert, unabhängige Messung ausstehend; laut Modellseite eingestellt", "scope": "Schätzwert von Artificial Analysis, keine eigene Messung. Allgemeine Modellbewertung, Produktkonfiguration und klinische Ausgabequalität sind dadurch nicht belegt.", "checked": "2026-10-02", "source": { "title": "Artificial Analysis: Mistral Small 3", "url": "https://artificialanalysis.ai/models/mistral-small-3" }, "authority": "independent", "rating": "Schwach" }, { "id": "aa-mistral-small-3-1", "model": "Mistral Small 3.1", "aliases": [ "Mistral Small 3.1" ], "kind": "language", "benchmark": "Artificial Analysis Intelligence Index", "version": "4.3.2", "value": 7, "unit": "points", "direction": "higher", "configuration": "ohne Reasoning; laut Modellseite eingestellt", "scope": "Allgemeine Modellbewertung. Produktkonfiguration und klinische Ausgabequalität sind dadurch nicht belegt.", "checked": "2026-10-02", "source": { "title": "Artificial Analysis: Mistral Small 3.1", "url": "https://artificialanalysis.ai/models/mistral-small-3-1" }, "authority": "independent", "rating": "Schwach" }, { "id": "aa-mistral-small-3-2", "model": "Mistral Small 3.2", "aliases": [ "Mistral Small 3.2" ], "kind": "language", "benchmark": "Artificial Analysis Intelligence Index", "version": "4.3.2", "value": 8, "unit": "points", "direction": "higher", "configuration": "ohne Reasoning; von Artificial Analysis als Schätzung markiert, unabhängige Messung ausstehend; laut Modellseite eingestellt", "scope": "Schätzwert von Artificial Analysis, keine eigene Messung. Allgemeine Modellbewertung, Produktkonfiguration und klinische Ausgabequalität sind dadurch nicht belegt.", "checked": "2026-10-02", "source": { "title": "Artificial Analysis: Mistral Small 3.2", "url": "https://artificialanalysis.ai/models/mistral-small-3-2" }, "authority": "independent", "rating": "Schwach" }, { "id": "aa-mistral-small-4-reasoning", "model": "Mistral Small 4", "aliases": [ "Mistral Small 4", "Mistral Small 4 (Reasoning)" ], "kind": "language", "benchmark": "Artificial Analysis Intelligence Index", "version": "4.3.2", "value": 11, "unit": "points", "direction": "higher", "configuration": "mit Reasoning", "scope": "Allgemeine Modellbewertung. Produktkonfiguration und klinische Ausgabequalität sind dadurch nicht belegt.", "checked": "2026-10-02", "source": { "title": "Artificial Analysis: Mistral Small 4 (Reasoning)", "url": "https://artificialanalysis.ai/models/mistral-small-4" }, "authority": "independent", "rating": "Schwach" }, { "id": "aa-mistral-medium-3", "model": "Mistral Medium 3", "aliases": [ "Mistral Medium 3" ], "kind": "language", "benchmark": "Artificial Analysis Intelligence Index", "version": "4.3.2", "value": 9, "unit": "points", "direction": "higher", "configuration": "ohne Reasoning; von Artificial Analysis als Schätzung markiert, unabhängige Messung ausstehend; laut Modellseite eingestellt", "scope": "Schätzwert von Artificial Analysis, keine eigene Messung. Allgemeine Modellbewertung, Produktkonfiguration und klinische Ausgabequalität sind dadurch nicht belegt.", "checked": "2026-10-02", "source": { "title": "Artificial Analysis: Mistral Medium 3", "url": "https://artificialanalysis.ai/models/mistral-medium-3" }, "authority": "independent", "rating": "Schwach" }, { "id": "aa-mistral-medium-3-5", "model": "Mistral Medium 3.5", "aliases": [ "Mistral Medium 3.5" ], "kind": "language", "benchmark": "Artificial Analysis Intelligence Index", "version": "4.3.2", "value": 14, "unit": "points", "direction": "higher", "configuration": "Standardeinstellung", "scope": "Allgemeine Modellbewertung. Produktkonfiguration und klinische Ausgabequalität sind dadurch nicht belegt.", "checked": "2026-10-02", "source": { "title": "Artificial Analysis: Mistral Medium 3.5", "url": "https://artificialanalysis.ai/models/mistral-medium-3-5" }, "authority": "independent", "rating": "Schwach" }, { "id": "aa-mistral-large-2-nov-24", "model": "Mistral Large 2", "aliases": [ "Mistral Large", "Mistral Large 2 (Nov '24)" ], "kind": "language", "benchmark": "Artificial Analysis Intelligence Index", "version": "4.3.2", "value": 8, "unit": "points", "direction": "higher", "configuration": "Version November 2024, ohne Reasoning; von Artificial Analysis als Schätzung markiert, unabhängige Messung ausstehend; laut Modellseite eingestellt", "scope": "Schätzwert von Artificial Analysis, keine eigene Messung. Allgemeine Modellbewertung, Produktkonfiguration und klinische Ausgabequalität sind dadurch nicht belegt.", "checked": "2026-10-09", "source": { "title": "Artificial Analysis: Mistral Large 2 (Nov '24)", "url": "https://artificialanalysis.ai/models/mistral-large-2" }, "authority": "independent", "rating": "Schwach" }, { "id": "aa-nemotron-3-ultra", "model": "NVIDIA Nemotron 3 Ultra", "aliases": [ "Nemotron 3 Ultra", "Nemotron 3 Ultra 550B A55B", "Nemotron 3 Ultra 550B A55B (Reasoning)" ], "kind": "language", "benchmark": "Artificial Analysis Intelligence Index", "version": "4.3.2", "value": 23, "unit": "points", "direction": "higher", "configuration": "mit Reasoning", "scope": "Allgemeine Modellbewertung. Produktkonfiguration und klinische Ausgabequalität sind dadurch nicht belegt.", "checked": "2026-10-02", "source": { "title": "Artificial Analysis: Nemotron 3 Ultra 550B A55B (Reasoning)", "url": "https://artificialanalysis.ai/models/nvidia-nemotron-3-ultra-550b-a55b" }, "authority": "independent", "rating": "Solide" }, { "id": "aa-nemotron-3-super", "model": "NVIDIA Nemotron 3 Super", "aliases": [ "Nemotron 3 Super", "Nemotron 3 Super 120B A12B", "Nemotron 3 Super 120B A12B (Reasoning)" ], "kind": "language", "benchmark": "Artificial Analysis Intelligence Index", "version": "4.3.2", "value": 13, "unit": "points", "direction": "higher", "configuration": "mit Reasoning", "scope": "Allgemeine Modellbewertung. Produktkonfiguration und klinische Ausgabequalität sind dadurch nicht belegt.", "checked": "2026-10-02", "source": { "title": "Artificial Analysis: Nemotron 3 Super 120B A12B (Reasoning)", "url": "https://artificialanalysis.ai/models/nvidia-nemotron-3-super-120b-a12b" }, "authority": "independent", "rating": "Schwach" }, { "id": "aa-nemotron-3-nano-30b-reasoning", "model": "NVIDIA Nemotron 3 Nano", "aliases": [ "Nemotron 3 Nano", "Nemotron 3 Nano 30B A3B", "NVIDIA Nemotron 3 Nano 30B A3B (Reasoning)" ], "kind": "language", "benchmark": "Artificial Analysis Intelligence Index", "version": "4.3.2", "value": 9, "unit": "points", "direction": "higher", "configuration": "mit Reasoning", "scope": "Allgemeine Modellbewertung. Produktkonfiguration und klinische Ausgabequalität sind dadurch nicht belegt.", "checked": "2026-10-02", "source": { "title": "Artificial Analysis: NVIDIA Nemotron 3 Nano 30B A3B (Reasoning)", "url": "https://artificialanalysis.ai/models/nvidia-nemotron-3-nano-30b-a3b-reasoning" }, "authority": "independent", "rating": "Schwach" }, { "id": "aa-nemotron-3-5-lightning", "model": "NVIDIA Nemotron 3.5 Lightning", "aliases": [ "Nemotron 3.5 Lightning" ], "kind": "language", "benchmark": "Artificial Analysis Intelligence Index", "version": "4.3.2", "value": 13, "unit": "points", "direction": "higher", "configuration": "Standardeinstellung", "scope": "Allgemeine Modellbewertung. Produktkonfiguration und klinische Ausgabequalität sind dadurch nicht belegt.", "checked": "2026-10-02", "source": { "title": "Artificial Analysis: Nemotron 3.5 Lightning", "url": "https://artificialanalysis.ai/models/nemotron-3-5-lightning" }, "authority": "independent", "rating": "Schwach" }, { "id": "aa-grok-4-7-xhigh", "model": "Grok 4.7", "aliases": [ "Grok 4.7", "xAI Grok 4.7", "Grok 4.7 (Xhigh)" ], "kind": "language", "benchmark": "Artificial Analysis Intelligence Index", "version": "4.3.2", "value": 46, "unit": "points", "direction": "higher", "configuration": "Reasoning-Stufe sehr hoch", "scope": "Allgemeine Modellbewertung. Produktkonfiguration und klinische Ausgabequalität sind dadurch nicht belegt.", "checked": "2026-10-02", "source": { "title": "Artificial Analysis: Grok 4.7 (Xhigh)", "url": "https://artificialanalysis.ai/models/grok-4-7" }, "authority": "independent", "rating": "Stark" }, { "id": "aa-qwen3-8-27b-non-reasoning", "model": "Qwen3.8 27B", "aliases": [ "Qwen3.8 27B", "Qwen3.8 27B (Non-reasoning)", "qwen3.8:27b" ], "kind": "language", "benchmark": "Artificial Analysis Intelligence Index", "version": "4.3.2", "value": 20, "unit": "points", "direction": "higher", "configuration": "ohne Reasoning", "scope": "Allgemeine Modellbewertung. Produktkonfiguration und klinische Ausgabequalität sind dadurch nicht belegt.", "checked": "2026-10-20", "source": { "title": "Artificial Analysis: Qwen3.8 27B (Non-reasoning)", "url": "https://artificialanalysis.ai/models/qwen3-8-27b-non-reasoning" }, "authority": "independent", "rating": "Solide" }, { "id": "de-anamnese-assemblyai-universal", "model": "AssemblyAI Universal", "aliases": [ "AssemblyAI Universal", "assemblyai/universal" ], "kind": "speech", "benchmark": "Deutsche Arzt-Patienten-Gespräche (arXiv 2601.19945)", "version": "arXiv:2601.19945v1 (23.01.2026)", "value": 2.99, "unit": "%", "direction": "lower", "configuration": "Tabelle 2: WER 0.0299 (Fließtext: 3,0 %), API über AssemblyAI SDK, Version nicht genauer benannt (Tabelle 1 listet Universal-1, Universal-2, Universal-nano); Szenarien 0,6 / 5,2 / 4,7 / 1,4 %, identisch mit der Zeile „assemblyai/universal-nano“ in Abbildung 2; 16 kHz, Peak-Normalisierung, Greedy Decoding, Kleinschreibung und Entfernung der Interpunktion; ungewichteter Mittelwert über vier Szenarien", "scope": "Deutsche simulierte Anamnesegespräche: vier Szenarien aus öffentlichen YouTube-Videos, manuell transkribiert, teils mit Akzent und Dialekt. Preprint ohne Peer Review, sehr kleine Stichprobe, keine Messung der Anbieteranpassung.", "checked": "2026-10-02", "source": { "title": "Schuster et al.: Benchmarking von ASR-Modellen im deutschen medizinischen Kontext, Tabelle 2", "url": "https://arxiv.org/html/2601.19945v1" }, "authority": "independent", "rating": null }, { "id": "de-anamnese-voxtral-small", "model": "Mistral Voxtral Small", "aliases": [ "Voxtral Small", "voxtral-small-latest" ], "kind": "speech", "benchmark": "Deutsche Arzt-Patienten-Gespräche (arXiv 2601.19945)", "version": "arXiv:2601.19945v1 (23.01.2026)", "value": 7.11, "unit": "%", "direction": "lower", "configuration": "Tabelle 2: „Voxtral Small Latest“ WER 0.0711, Mistral API (voxtral-small-latest); Szenarien 8,9 / 6,2 / 6,4 / 6,9 %; 16 kHz, Peak-Normalisierung, Greedy Decoding, Kleinschreibung und Entfernung der Interpunktion; ungewichteter Mittelwert über vier Szenarien", "scope": "Deutsche simulierte Anamnesegespräche: vier Szenarien aus öffentlichen YouTube-Videos, manuell transkribiert, teils mit Akzent und Dialekt. Preprint ohne Peer Review, sehr kleine Stichprobe, keine Messung der Anbieteranpassung.", "checked": "2026-10-02", "source": { "title": "Schuster et al.: Benchmarking von ASR-Modellen im deutschen medizinischen Kontext, Tabelle 2", "url": "https://arxiv.org/html/2601.19945v1" }, "authority": "independent", "rating": null }, { "id": "de-anamnese-voxtral-mini", "model": "Mistral Voxtral Mini", "aliases": [ "Voxtral Mini", "voxtral-mini-latest" ], "kind": "speech", "benchmark": "Deutsche Arzt-Patienten-Gespräche (arXiv 2601.19945)", "version": "arXiv:2601.19945v1 (23.01.2026)", "value": 8.46, "unit": "%", "direction": "lower", "configuration": "Tabelle 2: „Voxtral Mini Latest“ WER 0.0846, Mistral API (voxtral-mini-latest); Szenarien 11,2 / 6,5 / 7,7 / 8,5 %; 16 kHz, Peak-Normalisierung, Greedy Decoding, Kleinschreibung und Entfernung der Interpunktion; ungewichteter Mittelwert über vier Szenarien", "scope": "Deutsche simulierte Anamnesegespräche: vier Szenarien aus öffentlichen YouTube-Videos, manuell transkribiert, teils mit Akzent und Dialekt. Preprint ohne Peer Review, sehr kleine Stichprobe, keine Messung der Anbieteranpassung.", "checked": "2026-10-02", "source": { "title": "Schuster et al.: Benchmarking von ASR-Modellen im deutschen medizinischen Kontext, Tabelle 2", "url": "https://arxiv.org/html/2601.19945v1" }, "authority": "independent", "rating": null }, { "id": "de-anamnese-whisper-large-v3-faster-whisper", "model": "Whisper Large v3", "aliases": [ "Whisper large-v3", "faster-whisper large-v3" ], "kind": "speech", "benchmark": "Deutsche Arzt-Patienten-Gespräche (arXiv 2601.19945)", "version": "arXiv:2601.19945v1 (23.01.2026)", "value": 12.58, "unit": "%", "direction": "lower", "configuration": "Tabelle 2: „Whisper Large-v3 (Distilled)“ WER 0.1258; Fließtext nennt „OpenAI Whisper Large-v3 … 12,6 %“; die Heatmapzeile „large-v3“ entspricht dem Checkpoint der lokalen Faster-Whisper-Implementierung (CTranslate2) aus Tabelle 1; Szenarien 13,5 / 11,4 / 8,5 / 17,0 %; 16 kHz, Peak-Normalisierung, Greedy Decoding, Kleinschreibung und Entfernung der Interpunktion; ungewichteter Mittelwert über vier Szenarien", "scope": "Deutsche simulierte Anamnesegespräche: vier Szenarien aus öffentlichen YouTube-Videos, manuell transkribiert, teils mit Akzent und Dialekt. Preprint ohne Peer Review, sehr kleine Stichprobe, keine Messung der Anbieteranpassung.", "checked": "2026-10-02", "source": { "title": "Schuster et al.: Benchmarking von ASR-Modellen im deutschen medizinischen Kontext, Tabelle 2", "url": "https://arxiv.org/html/2601.19945v1" }, "authority": "independent", "rating": null }, { "id": "de-anamnese-deepgram-nova-2", "model": "Deepgram Nova-2", "aliases": [ "Nova-2", "Deepgram Nova 2" ], "kind": "speech", "benchmark": "Deutsche Arzt-Patienten-Gespräche (arXiv 2601.19945)", "version": "arXiv:2601.19945v1 (23.01.2026)", "value": 14.33, "unit": "%", "direction": "lower", "configuration": "Tabelle 2: WER 0.1433, API über Deepgram SDK; Szenarien 17,0 / 13,5 / 11,3 / 15,6 %; 16 kHz, Peak-Normalisierung, Greedy Decoding, Kleinschreibung und Entfernung der Interpunktion; ungewichteter Mittelwert über vier Szenarien", "scope": "Deutsche simulierte Anamnesegespräche: vier Szenarien aus öffentlichen YouTube-Videos, manuell transkribiert, teils mit Akzent und Dialekt. Preprint ohne Peer Review, sehr kleine Stichprobe, keine Messung der Anbieteranpassung.", "checked": "2026-10-02", "source": { "title": "Schuster et al.: Benchmarking von ASR-Modellen im deutschen medizinischen Kontext, Tabelle 2", "url": "https://arxiv.org/html/2601.19945v1" }, "authority": "independent", "rating": null }, { "id": "oasr-de-assemblyai-universal-3-5-pro", "model": "AssemblyAI Universal-3.5 Pro", "aliases": [ "Universal-3.5 Pro", "assemblyai/universal-3-5-pro" ], "kind": "speech", "benchmark": "Hugging Face Open ASR Leaderboard, Deutsch (FLEURS und Common Voice)", "version": "Stand 2. Oktober 2026 (multilingual_de.csv, Revision 54d26266)", "value": 2.43, "unit": "%", "direction": "lower", "configuration": "Eintrag assemblyai/universal-3-5-pro; FLEURS 2.17 %, Common Voice 2.69 %; German Avg = Mittelwert beider Datensätze", "scope": "Deutsche vorgelesene Einzelsätze (FLEURS, Mozilla Common Voice), Mittelwert beider Datensätze. Keine Gespräche, keine medizinischen Aufnahmen, keine Messung der Anbieteranpassung.", "checked": "2026-10-09", "source": { "title": "Hugging Face Open ASR Leaderboard, Multilingual, Deutsch: assemblyai/universal-3-5-pro", "url": "https://huggingface.co/spaces/hf-audio/open_asr_leaderboard" }, "authority": "independent", "rating": "Spitzenklasse" }, { "id": "oasr-de-assemblyai-universal-3-pro", "model": "AssemblyAI Universal-3 Pro", "aliases": [ "Universal-3 Pro", "assemblyai/universal-3-pro" ], "kind": "speech", "benchmark": "Hugging Face Open ASR Leaderboard, Deutsch (FLEURS und Common Voice)", "version": "Stand 2. Oktober 2026 (multilingual_de.csv, Revision 54d26266)", "value": 2.59, "unit": "%", "direction": "lower", "configuration": "Eintrag assemblyai/universal-3-pro; FLEURS 2.42 %, Common Voice 2.76 %; German Avg = Mittelwert beider Datensätze", "scope": "Deutsche vorgelesene Einzelsätze (FLEURS, Mozilla Common Voice), Mittelwert beider Datensätze. Keine Gespräche, keine medizinischen Aufnahmen, keine Messung der Anbieteranpassung.", "checked": "2026-10-09", "source": { "title": "Hugging Face Open ASR Leaderboard, Multilingual, Deutsch: assemblyai/universal-3-pro", "url": "https://huggingface.co/spaces/hf-audio/open_asr_leaderboard" }, "authority": "independent", "rating": "Stark" }, { "id": "oasr-de-elevenlabs-scribe-v2", "model": "ElevenLabs Scribe v2", "aliases": [ "Scribe v2", "elevenlabs/scribe_v2" ], "kind": "speech", "benchmark": "Hugging Face Open ASR Leaderboard, Deutsch (FLEURS und Common Voice)", "version": "Stand 2. Oktober 2026 (multilingual_de.csv, Revision 54d26266)", "value": 2.25, "unit": "%", "direction": "lower", "configuration": "Eintrag elevenlabs/scribe_v2; FLEURS 2.3 %, Common Voice 2.19 %; German Avg = Mittelwert beider Datensätze", "scope": "Deutsche vorgelesene Einzelsätze (FLEURS, Mozilla Common Voice), Mittelwert beider Datensätze. Keine Gespräche, keine medizinischen Aufnahmen, keine Messung der Anbieteranpassung.", "checked": "2026-10-09", "source": { "title": "Hugging Face Open ASR Leaderboard, Multilingual, Deutsch: elevenlabs/scribe_v2", "url": "https://huggingface.co/spaces/hf-audio/open_asr_leaderboard" }, "authority": "independent", "rating": "Spitzenklasse" }, { "id": "oasr-de-azure-speech-07-2026", "model": "Azure AI Speech (Juli 2026)", "aliases": [ "microsoft/azure-speech-07-2026" ], "kind": "speech", "benchmark": "Hugging Face Open ASR Leaderboard, Deutsch (FLEURS und Common Voice)", "version": "Stand 2. Oktober 2026 (multilingual_de.csv, Revision 54d26266)", "value": 1.92, "unit": "%", "direction": "lower", "configuration": "Eintrag microsoft/azure-speech-07-2026; FLEURS 1.83 %, Common Voice 2.01 %; German Avg = Mittelwert beider Datensätze; Modellversion nur als Datum benannt", "scope": "Deutsche vorgelesene Einzelsätze (FLEURS, Mozilla Common Voice), Mittelwert beider Datensätze. Keine Gespräche, keine medizinischen Aufnahmen, keine Messung der Anbieteranpassung.", "checked": "2026-10-09", "source": { "title": "Hugging Face Open ASR Leaderboard, Multilingual, Deutsch: microsoft/azure-speech-07-2026", "url": "https://huggingface.co/spaces/hf-audio/open_asr_leaderboard" }, "authority": "independent", "rating": "Spitzenklasse" }, { "id": "oasr-de-speechmatics-enhanced", "model": "Speechmatics Enhanced", "aliases": [ "Speechmatics Enhanced", "speechmatics/enhanced" ], "kind": "speech", "benchmark": "Hugging Face Open ASR Leaderboard, Deutsch (FLEURS und Common Voice)", "version": "Stand 2. Oktober 2026 (multilingual_de.csv, Revision 54d26266)", "value": 2.77, "unit": "%", "direction": "lower", "configuration": "Eintrag speechmatics/enhanced; FLEURS 3.56 %, Common Voice 1.99 %; German Avg = Mittelwert beider Datensätze", "scope": "Deutsche vorgelesene Einzelsätze (FLEURS, Mozilla Common Voice), Mittelwert beider Datensätze. Keine Gespräche, keine medizinischen Aufnahmen, keine Messung der Anbieteranpassung.", "checked": "2026-10-09", "source": { "title": "Hugging Face Open ASR Leaderboard, Multilingual, Deutsch: speechmatics/enhanced", "url": "https://huggingface.co/spaces/hf-audio/open_asr_leaderboard" }, "authority": "independent", "rating": "Stark" }, { "id": "oasr-de-soniox-stt-async-v5", "model": "Soniox v5 Async", "aliases": [ "Soniox v5", "Soniox stt-async-v5", "soniox/stt-async-v5" ], "kind": "speech", "benchmark": "Hugging Face Open ASR Leaderboard, Deutsch (FLEURS und Common Voice)", "version": "Stand 2. Oktober 2026 (multilingual_de.csv, Revision 54d26266)", "value": 3.64, "unit": "%", "direction": "lower", "configuration": "Eintrag soniox/stt-async-v5; FLEURS 2.66 %, Common Voice 4.62 %; German Avg = Mittelwert beider Datensätze", "scope": "Deutsche vorgelesene Einzelsätze (FLEURS, Mozilla Common Voice), Mittelwert beider Datensätze. Keine Gespräche, keine medizinischen Aufnahmen, keine Messung der Anbieteranpassung.", "checked": "2026-10-09", "source": { "title": "Hugging Face Open ASR Leaderboard, Multilingual, Deutsch: soniox/stt-async-v5", "url": "https://huggingface.co/spaces/hf-audio/open_asr_leaderboard" }, "authority": "independent", "rating": "Stark" }, { "id": "oasr-de-whisper-large-v3", "model": "Whisper Large v3", "aliases": [ "Whisper large-v3", "OpenAI Whisper Large v3", "openai/whisper-large-v3" ], "kind": "speech", "benchmark": "Hugging Face Open ASR Leaderboard, Deutsch (FLEURS und Common Voice)", "version": "Stand 2. Oktober 2026 (multilingual_de.csv, Revision 54d26266)", "value": 4, "unit": "%", "direction": "lower", "configuration": "Eintrag openai/whisper-large-v3; FLEURS 3.2 %, Common Voice 4.79 %; German Avg = Mittelwert beider Datensätze", "scope": "Deutsche vorgelesene Einzelsätze (FLEURS, Mozilla Common Voice), Mittelwert beider Datensätze. Keine Gespräche, keine medizinischen Aufnahmen, keine Messung der Anbieteranpassung.", "checked": "2026-10-09", "source": { "title": "Hugging Face Open ASR Leaderboard, Multilingual, Deutsch: openai/whisper-large-v3", "url": "https://huggingface.co/spaces/hf-audio/open_asr_leaderboard" }, "authority": "independent", "rating": "Stark" }, { "id": "oasr-de-whisper-large-v3-turbo", "model": "Whisper Large v3 Turbo", "aliases": [ "Whisper large-v3-turbo", "OpenAI Whisper Large v3 Turbo", "openai/whisper-large-v3-turbo", "Whisper large-v3 Turbo", "Whisper V3 Turbo" ], "kind": "speech", "benchmark": "Hugging Face Open ASR Leaderboard, Deutsch (FLEURS und Common Voice)", "version": "Stand 2. Oktober 2026 (multilingual_de.csv, Revision 54d26266)", "value": 6.12, "unit": "%", "direction": "lower", "configuration": "Eintrag openai/whisper-large-v3-turbo; FLEURS 3.67 %, Common Voice 8.58 %; German Avg = Mittelwert beider Datensätze", "scope": "Deutsche vorgelesene Einzelsätze (FLEURS, Mozilla Common Voice), Mittelwert beider Datensätze. Keine Gespräche, keine medizinischen Aufnahmen, keine Messung der Anbieteranpassung.", "checked": "2026-10-09", "source": { "title": "Hugging Face Open ASR Leaderboard, Multilingual, Deutsch: openai/whisper-large-v3-turbo", "url": "https://huggingface.co/spaces/hf-audio/open_asr_leaderboard" }, "authority": "independent", "rating": "Solide" }, { "id": "oasr-de-parakeet-tdt-0-6b-v3", "model": "NVIDIA Parakeet TDT 0.6B v3", "aliases": [ "Parakeet TDT 0.6B v3", "nvidia/parakeet-tdt-0.6b-v3" ], "kind": "speech", "benchmark": "Hugging Face Open ASR Leaderboard, Deutsch (FLEURS und Common Voice)", "version": "Stand 2. Oktober 2026 (multilingual_de.csv, Revision 54d26266)", "value": 4.12, "unit": "%", "direction": "lower", "configuration": "Eintrag nvidia/parakeet-tdt-0.6b-v3; FLEURS 4.16 %, Common Voice 4.07 %; German Avg = Mittelwert beider Datensätze; Parakeet TDT 0.6B v2 ist rein englisch und hat keinen deutschen Eintrag", "scope": "Deutsche vorgelesene Einzelsätze (FLEURS, Mozilla Common Voice), Mittelwert beider Datensätze. Keine Gespräche, keine medizinischen Aufnahmen, keine Messung der Anbieteranpassung.", "checked": "2026-10-09", "source": { "title": "Hugging Face Open ASR Leaderboard, Multilingual, Deutsch: nvidia/parakeet-tdt-0.6b-v3", "url": "https://huggingface.co/spaces/hf-audio/open_asr_leaderboard" }, "authority": "independent", "rating": "Stark" }, { "id": "oasr-de-canary-1b-v2", "model": "NVIDIA Canary 1B v2", "aliases": [ "Canary 1B v2", "nvidia/canary-1b-v2" ], "kind": "speech", "benchmark": "Hugging Face Open ASR Leaderboard, Deutsch (FLEURS und Common Voice)", "version": "Stand 2. Oktober 2026 (multilingual_de.csv, Revision 54d26266)", "value": 4.06, "unit": "%", "direction": "lower", "configuration": "Eintrag nvidia/canary-1b-v2; FLEURS 3.43 %, Common Voice 4.69 %; German Avg = Mittelwert beider Datensätze", "scope": "Deutsche vorgelesene Einzelsätze (FLEURS, Mozilla Common Voice), Mittelwert beider Datensätze. Keine Gespräche, keine medizinischen Aufnahmen, keine Messung der Anbieteranpassung.", "checked": "2026-10-09", "source": { "title": "Hugging Face Open ASR Leaderboard, Multilingual, Deutsch: nvidia/canary-1b-v2", "url": "https://huggingface.co/spaces/hf-audio/open_asr_leaderboard" }, "authority": "independent", "rating": "Stark" }, { "id": "oasr-de-nemotron-3-5-asr-streaming", "model": "NVIDIA Nemotron 3.5 ASR Streaming 0.6B", "aliases": [ "Nemotron ASR Streaming", "nvidia/nemotron-3.5-asr-streaming-0.6b" ], "kind": "speech", "benchmark": "Hugging Face Open ASR Leaderboard, Deutsch (FLEURS und Common Voice)", "version": "Stand 2. Oktober 2026 (multilingual_de.csv, Revision 54d26266)", "value": 8.96, "unit": "%", "direction": "lower", "configuration": "Eintrag nvidia/nemotron-3.5-asr-streaming-0.6b; FLEURS 8.46 %, Common Voice 9.46 %; German Avg = Mittelwert beider Datensätze", "scope": "Deutsche vorgelesene Einzelsätze (FLEURS, Mozilla Common Voice), Mittelwert beider Datensätze. Keine Gespräche, keine medizinischen Aufnahmen, keine Messung der Anbieteranpassung.", "checked": "2026-10-09", "source": { "title": "Hugging Face Open ASR Leaderboard, Multilingual, Deutsch: nvidia/nemotron-3.5-asr-streaming-0.6b", "url": "https://huggingface.co/spaces/hf-audio/open_asr_leaderboard" }, "authority": "independent", "rating": "Schwach" }, { "id": "oasr-de-voxtral-small-24b", "model": "Mistral Voxtral Small", "aliases": [ "Voxtral Small", "Voxtral-Small-24B-2507", "mistralai/Voxtral-Small-24B-2507" ], "kind": "speech", "benchmark": "Hugging Face Open ASR Leaderboard, Deutsch (FLEURS und Common Voice)", "version": "Stand 2. Oktober 2026 (multilingual_de.csv, Revision 54d26266)", "value": 2.9, "unit": "%", "direction": "lower", "configuration": "Eintrag mistralai/Voxtral-Small-24B-2507; FLEURS 2.61 %, Common Voice 3.19 %; German Avg = Mittelwert beider Datensätze", "scope": "Deutsche vorgelesene Einzelsätze (FLEURS, Mozilla Common Voice), Mittelwert beider Datensätze. Keine Gespräche, keine medizinischen Aufnahmen, keine Messung der Anbieteranpassung.", "checked": "2026-10-09", "source": { "title": "Hugging Face Open ASR Leaderboard, Multilingual, Deutsch: mistralai/Voxtral-Small-24B-2507", "url": "https://huggingface.co/spaces/hf-audio/open_asr_leaderboard" }, "authority": "independent", "rating": "Stark" }, { "id": "oasr-de-voxtral-mini-3b", "model": "Mistral Voxtral Mini", "aliases": [ "Voxtral Mini", "Voxtral-Mini-3B-2507", "mistralai/Voxtral-Mini-3B-2507" ], "kind": "speech", "benchmark": "Hugging Face Open ASR Leaderboard, Deutsch (FLEURS und Common Voice)", "version": "Stand 2. Oktober 2026 (multilingual_de.csv, Revision 54d26266)", "value": 4.46, "unit": "%", "direction": "lower", "configuration": "Eintrag mistralai/Voxtral-Mini-3B-2507; FLEURS 3.64 %, Common Voice 5.29 %; German Avg = Mittelwert beider Datensätze", "scope": "Deutsche vorgelesene Einzelsätze (FLEURS, Mozilla Common Voice), Mittelwert beider Datensätze. Keine Gespräche, keine medizinischen Aufnahmen, keine Messung der Anbieteranpassung.", "checked": "2026-10-09", "source": { "title": "Hugging Face Open ASR Leaderboard, Multilingual, Deutsch: mistralai/Voxtral-Mini-3B-2507", "url": "https://huggingface.co/spaces/hf-audio/open_asr_leaderboard" }, "authority": "independent", "rating": "Solide" }, { "id": "oasr-de-cohere-transcribe-03-2026", "model": "Cohere Transcribe", "aliases": [ "cohere-transcribe-03-2026", "CohereLabs/cohere-transcribe-03-2026" ], "kind": "speech", "benchmark": "Hugging Face Open ASR Leaderboard, Deutsch (FLEURS und Common Voice)", "version": "Stand 2. Oktober 2026 (multilingual_de.csv, Revision 54d26266)", "value": 3.1, "unit": "%", "direction": "lower", "configuration": "Eintrag CohereLabs/cohere-transcribe-03-2026; FLEURS 3.33 %, Common Voice 2.87 %; German Avg = Mittelwert beider Datensätze", "scope": "Deutsche vorgelesene Einzelsätze (FLEURS, Mozilla Common Voice), Mittelwert beider Datensätze. Keine Gespräche, keine medizinischen Aufnahmen, keine Messung der Anbieteranpassung.", "checked": "2026-10-09", "source": { "title": "Hugging Face Open ASR Leaderboard, Multilingual, Deutsch: CohereLabs/cohere-transcribe-03-2026", "url": "https://huggingface.co/spaces/hf-audio/open_asr_leaderboard" }, "authority": "independent", "rating": "Stark" } ], "familyReferences": { "Whisper": [ "aa-whisper-large-v3-fal", "aa-whisper-large-v3-replicate", "aa-whisper-large-v3-turbo-groq", "aa-whisper-large-v2-openai", "de-anamnese-whisper-large-v3-faster-whisper", "oasr-de-whisper-large-v3", "oasr-de-whisper-large-v3-turbo" ], "OpenAI Whisper": [ "aa-whisper-large-v3-fal", "aa-whisper-large-v3-replicate", "aa-whisper-large-v3-turbo-groq", "aa-whisper-large-v2-openai", "de-anamnese-whisper-large-v3-faster-whisper", "oasr-de-whisper-large-v3", "oasr-de-whisper-large-v3-turbo" ], "GPT": [ "aa-gpt-5-6-luna-max", "aa-gpt-5-6-sol-max", "aa-gpt-6-1-sol-max", "aa-gpt-5-6-terra-max", "aa-gpt-5-5-xhigh", "aa-gpt-5-4-xhigh", "aa-gpt-5-2-xhigh", "aa-gpt-4o-nov-24", "aa-gpt-4" ], "Claude": [ "aa-claude-sonnet-5-max", "aa-claude-opus-5-max", "aa-claude-fable-5-max-fallback", "aa-claude-opus-5-5-max", "aa-claude-sonnet-5-5-max", "aa-claude-fable-5-1-max", "aa-claude-haiku-4-5-reasoning" ], "Gemini": [ "aa-gemini-3-5-flash-high", "aa-gemini-4-argon-high", "aa-gemini-3-8-flash-high", "aa-gemini-3-1-pro-preview", "aa-gemini-3-5-flash-lite" ], "Gemini Audio": [ "aa-gemini-3-5-transcribe", "aa-gemini-3-1-pro-preview-high-stt", "aa-gemini-3-flash-high-stt" ], "Llama": [ "aa-llama-4-maverick", "aa-llama-4-scout" ], "Mistral": [ "aa-mistral-small-3", "aa-mistral-small-3-1", "aa-mistral-small-3-2", "aa-mistral-small-4-reasoning", "aa-mistral-medium-3", "aa-mistral-medium-3-5" ], "Voxtral": [ "aa-voxtral-small", "aa-voxtral-mini-transcribe-2", "de-anamnese-voxtral-small", "de-anamnese-voxtral-mini", "oasr-de-voxtral-small-24b", "oasr-de-voxtral-mini-3b" ], "AssemblyAI": [ "aa-assemblyai-universal-3-5-pro", "aa-assemblyai-universal-3-pro", "aa-assemblyai-universal", "aa-assemblyai-u3-5-realtime-pro-max-accuracy", "aa-assemblyai-u3-realtime-pro", "de-anamnese-assemblyai-universal", "oasr-de-assemblyai-universal-3-5-pro", "oasr-de-assemblyai-universal-3-pro" ], "Deepgram": [ "aa-deepgram-nova-3", "aa-deepgram-nova-3-telnyx", "aa-deepgram-nova-2-telnyx", "aa-deepgram-nova-3-realtime", "aa-deepgram-flux", "de-anamnese-deepgram-nova-2" ], "ElevenLabs": [ "aa-elevenlabs-scribe-v2", "aa-elevenlabs-scribe-v2-realtime", "oasr-de-elevenlabs-scribe-v2" ], "Speechmatics": [ "aa-speechmatics-enhanced", "aa-speechmatics-melia", "aa-speechmatics-standard", "aa-speechmatics-realtime-enhanced", "aa-speechmatics-agent-stt-linden-1", "oasr-de-speechmatics-enhanced" ], "Soniox": [ "aa-soniox-v5-async", "aa-soniox-v4", "aa-soniox-v5-realtime", "aa-soniox-v4-realtime", "oasr-de-soniox-stt-async-v5" ], "Google Speech": [ "aa-google-chirp-3", "aa-google-chirp-2", "aa-google-chirp", "aa-google-chirp-3-streaming" ], "Azure Speech": [ "aa-azure-llm-speech", "aa-mai-transcribe-2", "aa-mai-transcribe-1-5", "aa-azure-stt-realtime", "aa-mai-transcribe-2-streaming", "oasr-de-azure-speech-07-2026" ], "OpenAI Speech": [ "aa-gpt-transcribe", "aa-gpt-4o-transcribe", "aa-gpt-4o-mini-transcribe", "aa-gpt-realtime-whisper", "aa-gpt-live-transcribe" ], "Parakeet": [ "aa-parakeet-tdt-0-6b-v2", "aa-parakeet-rnnt-1-1b", "oasr-de-parakeet-tdt-0-6b-v3" ], "Canary": [ "aa-canary-qwen-2-5b", "oasr-de-canary-1b-v2" ], "Nemotron": [ "aa-nemotron-3-ultra", "aa-nemotron-3-super", "aa-nemotron-3-nano-30b-reasoning", "aa-nemotron-3-5-lightning" ], "Nemotron Speech": [ "aa-nemotron-3-asr-streaming-1120ms", "aa-nemotron-3-asr-streaming-160ms", "oasr-de-nemotron-3-5-asr-streaming" ], "Grok": [ "aa-grok-4-7-xhigh" ], "Qwen": [ "aa-qwen3-8-27b-non-reasoning" ] }, "indexStats": { "speech": { "benchmark": "Artificial Analysis Word Error Rate Index", "version": "2.2", "mode": "ohne Streaming, Modellebene (repräsentativer Endpunkt je Modell)", "best": 1.7, "median": 4, "count": 61, "bestModel": "StepAudio 3 ASR (StepFun)", "direction": "lower", "unit": "%", "source": { "title": "Artificial Analysis: Speech to Text, ohne Streaming", "url": "https://artificialanalysis.ai/speech-to-text/non-streaming" }, "checked": "2026-10-02" }, "speechStreaming": { "benchmark": "Artificial Analysis Word Error Rate Streaming Index (Final Transcription)", "version": "2.2", "best": 2.5, "median": 4.5, "count": 38, "bestModel": "MAI-Transcribe-2-Streaming (Microsoft AI)", "direction": "lower", "unit": "%", "source": { "title": "Artificial Analysis: Speech to Text, Streaming", "url": "https://artificialanalysis.ai/speech-to-text/streaming" }, "checked": "2026-10-02" }, "speechGerman": { "benchmark": "Hugging Face Open ASR Leaderboard, Deutsch (FLEURS und Common Voice)", "version": "Stand 2. Oktober 2026", "best": 1.92, "median": 4.12, "count": 39, "bestModel": "microsoft/azure-speech-07-2026", "direction": "lower", "unit": "%", "source": { "title": "Hugging Face Open ASR Leaderboard, Multilingual, Deutsch", "url": "https://huggingface.co/spaces/hf-audio/open_asr_leaderboard" }, "checked": "2026-10-09" }, "language": { "benchmark": "Artificial Analysis Intelligence Index", "version": "4.3.2", "best": 58, "median": 30, "count": 118, "bestModel": "Claude Opus 5.5 (Max, Default Fallback)", "direction": "higher", "unit": "points", "source": { "title": "Artificial Analysis: LLM Leaderboard", "url": "https://artificialanalysis.ai/leaderboards/models" }, "checked": "2026-10-02" } }, "unavailable": [ { "model": "Whisper", "aliases": [ "OpenAI Whisper" ], "reason": "Die Anbieter nennen die Modellfamilie, nicht eine konkrete Version oder Hostingkonfiguration. Es gibt deshalb keinen einzelnen zugeordneten Modellwert. Die Referenzwerte zeigen explizit benannte Varianten.", "checked": "2026-10-02", "sources": [ { "title": "Artificial Analysis: Whisper Familie", "url": "https://artificialanalysis.ai/speech-to-text/models/whisper" } ] }, { "model": "Claude", "aliases": [], "reason": "Die dokumentierte Familie hat keine eindeutig zugeordnete Version oder Reasoningkonfiguration. Konkrete Versionen werden ausschließlich als getrennte Familienreferenz aufgeführt.", "checked": "2026-10-02", "sources": [ { "title": "Artificial Analysis: Modellvergleich", "url": "https://artificialanalysis.ai/models" } ] }, { "model": "GPT", "aliases": [], "reason": "Die dokumentierte Familie hat keine eindeutig zugeordnete Version oder Reasoningkonfiguration. Konkrete Versionen werden ausschließlich als getrennte Familienreferenz aufgeführt.", "checked": "2026-10-02", "sources": [ { "title": "Artificial Analysis: Modellvergleich", "url": "https://artificialanalysis.ai/models" } ] }, { "model": "Gemini", "aliases": [], "reason": "Die dokumentierte Familie hat keine eindeutig zugeordnete Version oder Reasoningkonfiguration. Konkrete Versionen werden ausschließlich als getrennte Familienreferenz aufgeführt.", "checked": "2026-10-02", "sources": [ { "title": "Artificial Analysis: Modellvergleich", "url": "https://artificialanalysis.ai/models" } ] } ], "vendorClaims": [ { "provider": "heidi", "modelName": "NVIDIA Parakeet TDT 0.6B v2 (von Heidi feinabgestimmt)", "metric": "Nicht kuratierte WER („non-curated“) auf Heidis internem Testset (Englisch, Testset nicht näher beschrieben)", "value": 9.4, "unit": "%", "comparison": "Basismodell Parakeet v2 11,5 %, bisherige, nicht benannte Anbieterlösung 13,0 %", "published": "2026-07-31", "source": { "title": "Heidi: How Heidi Cut ASR Costs 64% and Latency 75% with NVIDIA Nemotron Open ASR (31.07.2026)", "url": "https://www.heidihealth.com/blog/how-heidi-improved-asr-nvidia-nemotron" }, "checked": "2026-10-02", "authority": "vendor" }, { "provider": "heidi", "modelName": "NVIDIA Parakeet TDT 0.6B v2 (von Heidi feinabgestimmt)", "metric": "F1 für medizinische Fachbegriffe", "value": 93.5, "unit": "%", "comparison": "Basismodell Parakeet v2 89,7 %, bisherige Anbieterlösung 89,5 %", "published": "2026-07-31", "source": { "title": "Heidi: How Heidi Cut ASR Costs 64% and Latency 75% with NVIDIA Nemotron Open ASR (31.07.2026)", "url": "https://www.heidihealth.com/blog/how-heidi-improved-asr-nvidia-nemotron" }, "checked": "2026-10-02", "authority": "vendor" }, { "provider": "heidi", "modelName": "NVIDIA Parakeet TDT 0.6B v2 (von Heidi feinabgestimmt)", "metric": "Gewinnrate in verblindeten Paarvergleichen", "value": 54.5, "unit": "%", "comparison": "Gegenüber dem bisher eingesetzten System, Bewerter und Stichprobengröße nennt Heidi nicht", "published": "2026-07-31", "source": { "title": "Heidi: How Heidi Cut ASR Costs 64% and Latency 75% with NVIDIA Nemotron Open ASR (31.07.2026)", "url": "https://www.heidihealth.com/blog/how-heidi-improved-asr-nvidia-nemotron" }, "checked": "2026-10-02", "authority": "vendor" }, { "provider": "heidi", "modelName": "NVIDIA Parakeet TDT 0.6B v2 (von Heidi feinabgestimmt)", "metric": "Latenz der Spracherkennung", "value": 0.7, "unit": "s", "comparison": "Bisherige Latenz etwa 3,0 s, laut Heidi mehr als 75 % Verbesserung", "published": "2026-07-31", "source": { "title": "Heidi: How Heidi Cut ASR Costs 64% and Latency 75% with NVIDIA Nemotron Open ASR (31.07.2026)", "url": "https://www.heidihealth.com/blog/how-heidi-improved-asr-nvidia-nemotron" }, "checked": "2026-10-02", "authority": "vendor" }, { "provider": "heidi", "modelName": "Drittanbieter-Spracherkennung (für Englisch abgelöst)", "metric": "Nicht kuratierte WER („non-curated“) auf Heidis internem Testset (Englisch)", "value": 13, "unit": "%", "comparison": "Heidis feinabgestimmtes Parakeet 9,4 %, Basismodell Parakeet v2 11,5 %, Anbieter nicht benannt", "published": "2026-07-31", "source": { "title": "Heidi: How Heidi Cut ASR Costs 64% and Latency 75% with NVIDIA Nemotron Open ASR (31.07.2026)", "url": "https://www.heidihealth.com/blog/how-heidi-improved-asr-nvidia-nemotron" }, "checked": "2026-10-02", "authority": "vendor" }, { "provider": "heidi", "modelName": "NVIDIA Parakeet TDT 0.6B v2 (von Heidi feinabgestimmt)", "metric": "Durchschnittliche WER des Basismodells im Hugging Face Open ASR Leaderboard, im AWS-Beitrag zitiert", "value": 6.05, "unit": "%", "comparison": "Gilt für das Basismodell ohne Heidis Feinabstimmung, gemessen auf englischen Benchmarkdatensätzen, einen Messwert für das feinabgestimmte Modell nennt der Beitrag nicht", "published": "2026-03-12", "source": { "title": "AWS Machine Learning Blog: Fine-tuning NVIDIA Nemotron Speech ASR on Amazon EC2 for domain adaptation (12.03.2026)", "url": "https://aws.amazon.com/blogs/machine-learning/fine-tuning-nvidia-nemotron-speech-asr-on-amazon-ec2-for-domain-adaptation" }, "checked": "2026-10-02", "authority": "vendor" }, { "provider": "tldv", "modelName": "tl;dv-Standardmodell", "metric": "Genauigkeit laut Anbieter (Messmethode, Sprache und Testset nicht genannt)", "value": 96, "unit": "%", "comparison": "Keine Vergleichsangabe", "published": "2025-06-19", "source": { "title": "tl;dv Hilfe: Why choose Whisper as a transcription model (19.06.2025)", "url": "https://intercom.help/tldv/en/articles/11427586-why-choose-whisper-as-a-transcription-model" }, "checked": "2026-10-02", "authority": "vendor" }, { "provider": "tldv", "modelName": "ElevenLabs Scribe", "metric": "Punktzahl für die Transkription im eigenen französischen Produktvergleich, bewertet nach sieben Kriterien, darunter 5 von 5 Punkten bei der Wortfehlerrate", "value": 62, "unit": "von 65 Punkten", "comparison": "Fireflies 52, Noota 49 und Google Gemini 43 Punkte, verglichen vom Anbieter an 10 bis 12 Minuten langen Clips, tl;dv laut Beitrag mit der lizenzierten ElevenLabs-Engine im Business-Konto", "published": "2026-07-09", "source": { "title": "tl;dv: French Meeting Transcription Accuracy: We Tested 4 AI Tools (09.07.2026)", "url": "https://tldv.io/blog/french-meeting-transcription-accuracy/" }, "checked": "2026-10-02", "authority": "vendor" }, { "provider": "fireflies", "modelName": "nicht modellspezifisch (Fireflies-Transkription mit AssemblyAI bzw. Soniox)", "metric": "Genauigkeit laut Startseite (Messmethode, Sprache und Testset nicht genannt)", "value": 95, "unit": "%", "comparison": "Keine Vergleichsangabe. Werbeaussage auf der Startseite: „95 % genau“", "published": null, "source": { "title": "Fireflies: Startseite", "url": "https://fireflies.ai/" }, "checked": "2026-10-02", "authority": "vendor" } ], "methodologySources": [ { "title": "Artificial Analysis: Speech to Text Methodology, aktuelle Unterversion 2.2", "url": "https://artificialanalysis.ai/methodology/speech-to-text" }, { "title": "Artificial Analysis: Intelligence Benchmarking, aktuelle Version 4.3.2", "url": "https://artificialanalysis.ai/methodology/intelligence-benchmarking" }, { "title": "Hugging Face Open ASR Leaderboard, deutsche Teilwertung", "url": "https://huggingface.co/spaces/hf-audio/open_asr_leaderboard" }, { "title": "arXiv 2601.19945: Spracherkennung in deutschen Arzt-Patienten-Gesprächen (Preprint)", "url": "https://arxiv.org/abs/2601.19945" } ], "notes": [ "Artificial Analysis führt den Word Error Rate Index in der Ergebnistabelle als AA-WER v2. Das Feld version nennt die genaue Unterversion 2.2 aus der Methodik von Artificial Analysis, gültig seit Mai 2026.", "Die Werte des Artificial Analysis Intelligence Index 4.3.2 stammen von den Modellseiten auf artificialanalysis.ai, die je Modell einen Indexwert nennen.", "Ein externer Modellwert ist kein ScribeIndex-Score. Keine Umrechnung von Indexpunkten in medizinische Genauigkeit.", "Der Wert für Claude Fable 5 stammt aus einer Messung mit Opus 4.8 als Rückfallmodell und gilt nur für diese kombinierte Konfiguration.", "Die Bewertungsstufen beziehen sich je Benchmark auf den Bestwert und den Median aller dort gemessenen Modelle. Werte verschiedener Benchmarks werden nicht miteinander verrechnet.", "Die arXiv-Werte zu deutschen Anamnesegesprächen stammen aus einem Preprint mit vier Szenarien. Sie werden angezeigt, gehen aber nicht in die Bewertungsstufe ein.", "Mit „Schätzwert von Artificial Analysis“ gekennzeichnete Indexwerte sind keine eigene Messung von Artificial Analysis.", "Herstellerangaben (vendorClaims) stammen aus eigenen Tests der Anbieter mit nicht offengelegten Testsets und sind nicht mit unabhängigen Benchmarks vergleichbar." ] }