Long pauses make voice interactions harder to follow. A voice feature needs an explicit latency budget and tests on its intended network, devices and providers. The clinic sample on this website is scripted; it does not demonstrate a live speech pipeline or measured latency.
Where time is spent
Speech recognition, endpoint detection, model inference, speech synthesis and network transport each contribute to response time. Measure each stage before choosing an optimisation.
- Streaming recognition. Partial transcripts can overlap with the caller’s speech. Test endpoint detection against pauses and background noise.
- Incremental generation. Streaming model output can let synthesis start before the complete reply is available. Test incomplete clauses and cancellation.
- Streaming synthesis. Audio chunks can reduce the wait for playback, with trade-offs in phrasing and prosody.
- Transport. Account for telephony codecs, buffering and network location in the end-to-end measurement.
Handle interruptions explicitly
Define how playback stops when a caller interrupts, how queued audio is cancelled and how the conversation resumes. Test overlapping speech and interrupted tool actions, rather than assuming a streaming pipeline handles them automatically.
Agree measurable targets
Useful measures include time to first audible response, interruption recovery and task completion. Set acceptance criteria for representative conditions, including slow networks and provider failures. These are evaluation recommendations, not performance results for the browser sample.
Read about scoping voice AI integrations or discuss an AI project.