Start with 45 free minutes of callsTry for free

Request a new model

Tell us which model you want measured. We review every request.

Speech-to-Text Benchmark

Benchmarks

X
Y

WER

End of turn

lower is better

Time from the last spoken word to the final transcript, decided by the model itself.

Time to first token

lower is better

Time from the caller starting to speak to the first word of the live transcript.

Live words

lower is better

Time from a word being spoken to it appearing in the live transcript for good.

Rewrites

lower is better

Words the live transcript takes back, per word in the final transcript.

Cost

lower is better

Published list price per thousand minutes; batch-only models show their batch rate.

FAQ

Which STT API Has the Lowest Latency for Phone Calls?
Meta Muse Voice Transcribe had the lowest end of turn on this board, 301ms p50 from the last spoken word to the final transcript, timed vendor-direct from US East. Azure MAI-Transcribe-2 Streaming had the lowest live-word time, 114ms p50 from a word being spoken to it staying in the live transcript.
What Is the Most Accurate Speech-to-Text API for Live Conversation?
Azure MAI-Transcribe-2 Streaming measured the lowest streaming WER on this board, 6.2% on 448 English conversation clips (95% interval 5.2% to 7.1%). The 95% intervals of 10 other models overlap it.
Why Is Streaming WER Higher Than Batch WER?
One reason is that a streaming model commits words before it hears the rest of the sentence, while a batch model reads the whole recording. On the same 448 conversation clips, 10 of the 12 models measured both ways had a higher WER streaming; Deepgram Nova-3 measured 11.0% streaming and 10.1% batch.
What Is the Cheapest Streaming STT API?
Modulate Velma 2 (8.4% WER) and Hopper nemotron-asr (14.4% WER) cost $0.83 per 1,000 minutes, the lowest streaming price on this board.
What Does End-of-Turn Latency Mean for a Voice Agent?
End of turn is the time from the caller's last spoken word to the final transcript, with the model deciding on its own that the turn is over. On this board the p50 runs from 301ms (Meta Muse Voice Transcribe) to 1.90s (Inworld Realtime STT-1).
How Is This Speech-to-Text Benchmark Measured?
Streaming WER pools the word errors over 448 English conversation clips (EdAcc 98, PriMock57 150, AMI 50 and Earnings-22 150), sent at real-time pace to each vendor's own API; batch WER sends the same clips as whole files. Latency is timed from US East, end of turn uses each model's own turn detection, and cost is each model's price per 1,000 billed minutes, with no Speko markup.