Voice input is lazy and local. Typed-only sessions do not load the VAD stack until voice is actually used.

​
Input path

VoiceInput records 16 kHz mono audio through sounddevice.InputStream. Silero VAD receives 512-sample frames, about 32 ms each. A lookback deque preserves speech onset. When VAD emits speech_end, Eyra stops recording if enough frames were captured. Very short clips are treated as false triggers.

​
Transcription

Transcription prefers Local Whisper’s Unix socket: 
~/.whisper/cmd.sock
If the socket is unavailable, Eyra falls back to the resolved wh binary from preflight. Runtime code uses state.wh_bin, not bare wh from PATH.

​
Speech output

SpeechController calls Local Whisper TTS through the resolved wh_bin. interrupt() kills active TTS so the user can cut in. listen() waits for speech to finish before opening the microphone.

​
Diagnostics

VoiceDiagnostics checks device selection, non-zero input, VAD behavior, Local Whisper ASR, generated WAV transcription, and barge-in flows. /voice-diagnose is the normal first check.

​
Recovery

/voice on re-runs the Local Whisper check and enables whichever side is available. ASR and TTS are tracked separately.