Audio benchmarks
Every benchmark the Audio index tracks, with each core benchmark's share of the composite and how many models it covers.
31 benchmarks tracked.
- Agentic Performance (AA, tau-Voice based)
- AA Speech to Speech Index
- AA-WER Streaming
- AA-WER v2 (non-streaming)
- MMAR
- MMAU (v05.15.25) accuracy
- MMAU-Pro
- MMSU accuracy
- Open ASR AMI meeting speech WER
- Open ASR Earnings-22 earnings calls WER
- Open ASR English short-form average WER
- Open ASR GigaSpeech WER
- Open ASR long-form average WER
- Open ASR LibriSpeech test-clean WER
- Open ASR LibriSpeech test-other WER
- Open ASR multilingual DE WER
- Open ASR multilingual ES WER
- Open ASR multilingual FR WER
- Open ASR multilingual IT WER
- Open ASR multilingual PT WER
- Open ASR multilingual average WER
- Open ASR RTFx (throughput)
- Open ASR SPGISpeech WER
- Open ASR VoxPopuli WER
- Diarization error rate (pyannoteAI benchmark)
- Speech Agent Arena preference Elo
- Conversational Dynamics (Full-Duplex-Bench subset)
- Speech Reasoning (Big Bench Audio)
- Task Success Rate (AA)
- tau-Voice pass^1 (Sierra)
- VoiceBench overall