Evaluate voice latency before promising natural conversation
First text is only part of the wait
A model can show incremental text quickly while the user still waits for audio. The current speech path uploads a recorded turn and buffers a completed spoken reply. Measure the time from the end of speech to first audible output and to finished playback separately. First-text timing alone does not describe conversational responsiveness.
Identify the actual speech path
Configured server endpoints handle transcription and synthesis. When a server function is unconfigured, supported browser recognition or synthesis can be used. Browser recognition may use an external service and is not guaranteed offline. A configured server failure is reported instead of silently rerouting data. Confirm the path and destination before recording sensitive speech.
Measure real answer lengths
Test short, medium and long replies, including cold and warm inference. One earlier local CPU run took 18.7 seconds to generate 7.4 seconds of audio. That observation belongs to that hardware and model; it is not a release guarantee. Faster inference needs measurement on the target deployment, not a better first-text number.
Test failure and cancellation
Interrupt stops local playback and cancels in-flight model work while generation revisions reject stale output. It is not acoustic full-duplex barge-in. Deny microphone permission, stop the speech service and interrupt during generation and playback. Physical device acceptance remains open in v0.10. Define a pilot target using real users, recognition errors and latency percentiles before expanding access.
Product and guides
Explore creekX Agents · Quick start · Documentation. Written for v0.10 early access, 23 October 2026.
Frequently asked questions
Where should I start?
Read the separate Agents quick start and evaluate v0.10 in a private local workspace before a customer pilot.