[davencyw]

Adds an optional vad_start_threshold to the Gradium STT. When set, the server VAD in step messages starts speech as soon as the inactivity probability drops below it, instead of waiting for the first transcribed word. This allows fast barge-in without a local VAD, like Gradbot and Unmute (which use 0.4). Off by default.

• Steps sent before any audio was processed (total_duration_s == 0) are ignored: their estimate reads as activity whatever the input, so they would open a false turn on every new connection.
• A turn opened by VAD with no text still ends on silence as before.

Checked against the real server, speech start detected after speech onset:

clip	vad_start_threshold=0.4	default (first word)
long sentence	+0.38s	+1.09s
"Wait, stop."	+0.40s	+1.27s
"mm-hmm"	+0.33s	+1.26s
sentence from the very first audio	+0.27s	+1.13s
sentence 0.3s after connecting	+0.14s	+1.12s

No false starts on 10s of near silence or 6s of white noise.

With vad=None and turn_detection="stt", this start of speech sets the user state to speaking right away, so an agent can barge in on it:

@session.on("user_state_changed")
def _on_user_state(ev):
    if ev.new_state == "speaking" and session.agent_state == "speaking":
        try:
            session.interrupt()
        except RuntimeError:
            pass  # current speech is not interruptible

Note: this overlaps with #7870 and #7871 (same file), will rebase on whichever lands first.

[CLAassistant]

All committers have signed the CLA.

[davencyw]

@longcw @tinalenguyen could one of you take a look when you have a moment? A customer asked for this: fast barge-in on Gradium STT without adding a local VAD. It's opt-in (vad_start_threshold, default None), so default behavior doesn't change. Live results against our server are in the description. Happy to adjust anything.

[longcw]

could you resolve the conflicts before we can merge?

[davencyw]

Done, merged main and resolved the conflicts (both sides only added tests). Thanks for the review!