What is speech-to-text (STT)?

Accuracy is usually quoted as word error rate on clean read speech, which bears little resemblance to a phone call. Telephone audio is narrowband, compressed, and carries background noise, and callers use names, postcodes and numbers that general models handle poorly.

For agents, streaming matters as much as accuracy. A model that returns a perfect transcript only after the caller stops speaking forces the whole latency budget to start from scratch; a streaming model lets the pipeline begin working mid-sentence.

Domain vocabulary is the usual fix for the accuracy gap. Feeding the recogniser the actual list of treatments, product names or place names it should expect measurably reduces errors on exactly the words that decide whether the call succeeds.