Benchmarks
Industry-leading accuracy for pre-recorded speech-to-text.
Benchmarks are an important first step before running your own evaluation. Below are the current benchmarks for our pre-recorded models so you can assess performance across accuracy, latency, and error rates.
Public benchmarks can be misleading due to overfitting and benchmark gaming. We strongly recommend running your own evaluation on your audio data to identify the best model for your use case.
For the full interactive benchmark experience with competitive comparisons, visit assemblyai.com/benchmarks.
Word error rate (WER)
Word Error Rate (WER) is the classical metric for speech-to-text accuracy. It counts substitutions, deletions, and insertions against a reference transcript, divided by the total word count in the ground truth. AssemblyAI Universal-3 Pro achieves a mean WER of 5.6% (median 4.9%) on English benchmarks.
WER weights every word equally, so a misrecognized filler word counts the same as a misrecognized email address or medication name. For production voice workflows, we recommend pairing WER with Missed entity rate, which measures accuracy on the high-stakes entities — names, emails, phone numbers, and medical terms — that actually drive end-user outcomes.
English benchmarks
Most recent update: January 2026.
Multilingual benchmarks
Most recent update: January 2026. Dataset: FLEURS.
Missed entity rate
For production voice workflows, the actual words that matter most are entities — names, organizations, emails, phone numbers, and medical terms. The Missed Entity Rate (MER) measures how often a model fails to correctly transcribe these high-stakes terms. See Missed Entity Rate for the full definition.
Universal-3 Pro reduces missed entities by 27.1% overall compared to Universal-2, with the largest gains on emails, organization names, and medical terms.
Hallucinations and consecutive errors
Hallucinations are a critical concern in production STT systems. AssemblyAI reduces hallucinations by 30% compared to Whisper, across three error categories:
- Fabrications — words inserted that were never spoken
- Omissions — spoken words that are missing from the transcript
- Hallucinations — extended sequences of fabricated content
Benchmark challenges
Models are often trained on publicly available datasets — sometimes the very same datasets used for evaluation. When this happens, the model becomes overfit to the evaluation set and will show artificially strong performance on standard WER tests. This makes WER potentially misleading, as real-world performance on unseen audio will be significantly worse.
External benchmarks
For third-party benchmarks, we recommend the Hugging Face ASR Leaderboard. Note that many models listed require self-hosting and lack production features like speaker diarization and automatic language detection.
Methodology
Our benchmarks are evaluated across 250+ hours of audio data, 80,000+ audio files, and 26 datasets. We apply standard text normalization before calculating metrics. For full details on our methodology, visit assemblyai.com/benchmarks.
Run your own benchmark
We’d be happy to help. AssemblyAI has a benchmarking tool to help you run a custom evaluation against your real audio files. Contact us for more information.
You can also run your own benchmarks following the Hugging Face framework which provides a GitHub repo with full instructions.
