Speech-to-text

How it works

Speech-to-text (STT) models, also called automatic speech recognition, convert spoken audio into text. They are typically priced per minute or hour of input audio, and some token-based models bill audio input as tokens.

Models differ in accuracy across accents and languages, handling of noise and crosstalk, timestamps, speaker labels (diarization), custom vocabulary and real-time streaming. Accuracy is commonly reported as word error rate.

Test on recordings that match your real conditions, such as phone audio, meeting rooms or technical vocabulary. Accuracy on clean benchmark audio can be much better than on yours.

Worked example

Illustrative: at $0.006 per minute, transcribing 2,000 hours of call recordings (120,000 minutes) costs $720. Switching models and re-running costs the same again, so compare on a sample first.

Common mistakes

• 1Choosing a model from accuracy on clean audio instead of testing your own recordings.
• 2Not supplying custom vocabulary or prompts for names and domain terms.
• 3Forgetting that silence and hold music usually count as billed audio minutes unless trimmed.