Benchmarks
Industry-leading accuracy for streaming speech-to-text.
Industry-leading accuracy for streaming speech-to-text.
Benchmarks are an important first step before running your own evaluation. Below are the current benchmarks for our streaming 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) 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.
WER weights every word equally, so a misrecognized filler word counts the same as a misrecognized email or phone number. For voice agents, we recommend pairing WER with Missed entity rate, which measures accuracy on the high-stakes entities — names, emails, phone numbers, and medical terms — that break agent conversion in production.
Most recent update: March 2026.
Most recent update: March 2026.
For voice agents, the words that break conversion are entities — emails, phone numbers, organizations, names, 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 Streaming delivers double-digit relative gains over Universal Streaming across most categories, with the biggest improvements on emails, locations, and medical terms.
Source: AssemblyAI Streaming product page.
Most recent update: May 2026.
These figures are measured using in-domain voice agent conversation data from our internal evaluation set. In production, actual latency will vary depending on audio characteristics, silence duration, and turn detection configuration. Latency claims referenced elsewhere in our documentation may reflect different evaluation conditions or measurement approaches.
If you have questions about benchmarking methodology or need help interpreting these numbers for your use case, reach out to our team at support@assemblyai.com.
Time from when a word is spoken to when that word is returned by the API. Best metric for use cases that consume partial transcripts in real time.
Universal-3 Pro Streaming does not emit partial transcripts, so emission latency is not applicable.
Time from end of speaker turn to receipt of the finalized transcript. Best metric for voice agent responsiveness.
Time to First Token (TTFT) is a useful metric for LLMs but is not well suited to measuring streaming STT performance. For streaming STT, use emission latency (for partial transcript speed) and TTCT (for voice agent responsiveness).
In streaming, speed is critical. To achieve lower TTFT (time to first token) metrics, some providers emit tokens before any audio is actually spoken. These early tokens are hallucinations designed to game the benchmark, making TTFT a misleading measure of actual latency.
For third-party streaming benchmarks, we recommend the Coval Speech-to-Text Playground.
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.
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.