Speech-to-text. STT, sometimes called ASR for automatic speech recognition, is the layer that turns the caller's audio into the text that the LLM works with. It runs in two modes: streaming (partial transcripts arrive in real time) and final (the locked-in transcript after the caller stops speaking).
Voice agents in 2026 rely on streaming STT because the latency budget cannot afford to wait for final. The model starts thinking on the partial transcript and updates if the final is materially different. The quality bar is whether the partial-final gap introduces errors the customer notices.
STT vendors compete on accuracy in noisy environments (drive-thrus, busy clinics), on accented English, and on languages beyond English. The differences are real, but they are not where voice-agent quality actually lives, see voice-agent-latency for where the seconds actually go.
Vorel uses Deepgram + custom-tuned endpointing as the STT layer; the partial-final gap is benchmarked against vertical-specific noise profiles.