ASR (speech recognition)

ASR (automatic speech recognition) converts speech into text in real time. Its quality is measured by word error rate (WER): the lower it is, the less the AI misunderstands. Current models handle regional accents, construction-site background noise, and FR/EN code-switching — a Quebecker switching to "not bad, that works" mid-sentence must still be understood.

In a voice agent, ASR is the front-end stage: anything the model mishears gets mishandled, no matter how good the downstream LLM is. It's also one of the latency bottlenecks — streaming models transcribe while the caller is still talking, rather than after a pause. Without good speech recognition, there is no good voice AI.