Speech-to-text (STT)
Speech-to-text, also called automatic speech recognition (ASR), turns spoken audio into written words.
In a voice agent, STT is usually the first model in the chain: the caller’s audio is transcribed, the transcript goes to a language model, and the reply is spoken back with text-to-speech. For live calls it runs in streaming mode, producing partial words as the caller speaks rather than waiting for the whole sentence.
Accuracy on phone calls depends on the audio (8 kHz telephone lines carry less detail than a microphone), accents, background noise and domain words such as drug names or product codes. Most providers let you add custom vocabulary for those.
Some newer realtime models skip the separate transcription step and listen to the audio directly.
