voice_core 0.2.0

Cancellable provider-neutral voice conversation orchestration.

Changelog

0.2.0

• Add opt-in full-duplex conversation mode. VoiceDuplexMode, VoiceEchoCancellationFallback, and VoiceDuplexConfig select between the default half-duplex flow and a full-duplex flow in which the user can speak over synthesized playback. Half duplex remains the default and its behavior is unchanged: pass nothing and nothing moves.
• VoiceConversationController takes a duplex policy. In full duplex it never calls VoiceInput.setRecognitionEnabled — the microphone is expected to be echo-cancelled upstream, so recognition is never torn down and the onset of an interrupting utterance is no longer lost — and voice activity alone no longer interrupts playback. A committed final transcript still starts a new turn in both modes, and interruptTurn() still works in both.
• Add VoiceDuplexConfig.interruptAfterSustainedSpeech, an opt-in dwell after which sustained user speech interrupts full-duplex playback early. It stacks on the detector's own hysteresis rather than replacing it. Default null.
• Add VoiceConversationSnapshot.duplexMode, the mode actually in effect. Resolution happens at construction, so the first snapshot a listener sees already reports an honest mode.
• Full duplex requested without echo cancellation resolves to half duplex by default (VoiceDuplexConfig.isDegraded), on the principle that the degraded mode must never be worse than the no-AEC baseline: an open, uncancelled microphone makes the recognizer transcribe the assistant. Running full duplex anyway requires explicitly passing VoiceEchoCancellationFallback.acceptEchoRisk.
• voice_core still has no dependency on audio_aec and no knowledge of echo cancellation. It is told whether the microphone is clean; the composition lives in voice_flutter.

0.1.0

• Initial public release.