fix(speech): language ID minSpeechMs counts voiced speech, not trailing silence

[akirilyuk]

Summary

Symptom: a ~1 s utterance followed by silence passed a 1500 ms languageId.minSpeechMs threshold, so language ID ran on audio that was mostly silence.

Cause: both threshold checks (the early/continuous check and the identify gate) compared minSpeechMs with the length of the LID PCM buffer. That buffer runs from utterance start to close and includes the endpoint tail and STT gate hold after VAD speech end, so trailing silence counted as speech.

Change:

• The agent tracks lid_voiced_ms: speech from the first to the last voiced VAD frame of the utterance. Gaps between words count (they are folded in once speech resumes). Silence after the last voiced frame never counts.
• Both threshold checks and the too_short skip use lid_voiced_ms. The classifier still receives the whole buffered audio. Agents without a VAD treat every frame as voiced.
• force semantics, the continuous re-identify cadence and the defaults are unchanged.
• The roundtrip-language-id example threshold is retuned to speech units (2000 ms, a gate only under the default end-of-utterance timing).
• Docs (SDK LanguageIdConfig.minSpeechMs, VOICE-API.md) and CHANGELOG updated.

Tests (crates/speech/tests/language_id_nonblocking_test.rs):

• early_timing_short_speech_with_long_trailing_silence_is_too_short
• end_of_utterance_timing_short_speech_with_trailing_silence_is_too_short
• early_timing_identifies_once_voiced_reaches_min
• pauses_between_words_count_trailing_silence_does_not
• early_timing_does_not_identify_during_trailing_silence

Verified locally and via CI; no extra manual steps.

🤖 Generated with Claude Code