Spaces:

---

smolagents

/

hf-realtime-voice

like 381

Running

The playback worklet previously started on the first audio chunk with no prebuffer, so mid-turn network jitter produced audible ticks (out→silence→in).

This adds a 120 ms jitter/prebuffer (with a 300 ms max-wait flush), longer de-click ramps (4 ms), underrun re-arm, and a high (~30 s) runaway queue ceiling that never trims a normal TTS reply.

Ported from a production fork of this Space UI.