AssemblyAI Author Benchmark - prerecorded - Verbatim transcription - Noisy Speech - WER: leaderboard
AssemblyAI Author Benchmark - prerecorded - Verbatim transcription - Noisy Speech - WER — AssemblyAI primary author evaluation, captured 2026-11-16; no fusion with another campaign sharing a corpus. Exact system versions and run ranges retained in source_systems_rosters. Prerecorded: public API defaults, auto language except explicitly pinned Cohere/Azure; models shown only when >=90% of files returned relative to AssemblyAI. Streaming: real-time SDK/WebSocket delivery, fixed language, final transcripts rather than partials, shared normalization. All streaming 0.00 means no data. Language WER/CER >=60 means unsupported. Language-specific boards join all-language and partial-language rosters but exclude incomparable means over different supported languages. Native percentage error; no upper cap of100. Metric: WER (%). Source: https://www.assemblyai.com/benchmarks
Metric: WER (%). Source: www.assemblyai.com. Saturation forecast: Rough model projection: around 2026. 14 models tracked. Updated 16 Nov 2026.
Top models
| # | Model | Setting | Score | Overall rank |
|---|---|---|---|---|
| 1 | AssemblyAI Universal-3.5 Pro | 6.3 | ||
| 2 | Soniox v5 Async | 7.5 | ||
| 3 | Speechmatics Enhanced | 8.2 | ||
| 4 | Scribe v2 | 9 | ||
| 5 | Voxtral Mini Transcribe 2 | assemblyai_author_benchmarks_2026 run | 9.1 | |
| 6 | Cohere Transcribe (language pinned) | assemblyai_author_benchmarks_2026 run | 9.1 | |
| 7 | Azure Batch Transcription | assemblyai_author_benchmarks_2026 run | 9.4 | |
| 8 | Deepgram Nova 3 | 10 | ||
| 9 | parakeet-tdt-0.6B-v3 | 10 | ||
| 10 | Gemini 3.5 Transcribe | 10.7 | ||
| 11 | GPT-4o Transcribe | 11.4 | ||
| 12 | Grok STT | assemblyai_author_benchmarks_2026 run | 11.6 | |
| 13 | AWS Transcribe | assemblyai_author_benchmarks_2026 run | 19.9 | |
| 14 | Gladia Solaria-1 | assemblyai_author_benchmarks_2026 run | 38.6 |
Interactive version: theaggregate.ai/benchmark?slug=assemblyai-author-benchmark-prerecorded-verbatim-transcription-noisy-speech-wer · How It Works · Data refreshed daily, snapshot 2026-11-16.