Dodaios is an open-source agent OS for your desktop: it runs your coding agents, browser and apps by voice or chat, on your own machine. Formerly Personal Jarvis.
Python agent-orchestrationai-agentai-coding-assistantclaude-codecoding-agentcomputer-usedesktop-applocal-firstmcppersonal-jarvisvoice-assistantvoice-coding
fix(voice-engine): cut a reply after 160 ms of user speech, not 250 ms
rubenluetke10-beep673e72a9deb9
diff --git a/jarvis/voice_engine/bench/worker_suite.py b/jarvis/voice_engine/bench/worker_suite.pyindex 70c9f23e1d..7b2bc9f518 100644--- a/jarvis/voice_engine/bench/worker_suite.py+++ b/jarvis/voice_engine/bench/worker_suite.py@@ -58,6 +58,10 @@ class _Conversation:self.audio_frames = 0self.done = asyncio.Event()self.interrupted_at: float | None = None+ # Audio after this point answers the user's next utterance, so it is+ # no leak of the interrupted reply.+ self.next_turn_at: float | None = None+ self.leaked_until: float | None = Noneself.metrics: dict[str, Any] = {}def reset(self) -> None:@@ -67,6 +71,8 @@ class _Conversation:self.audio_frames = 0self.done = asyncio.Event()self.interrupted_at = None+ self.next_turn_at = None+ self.leaked_until = Noneself.metrics = {}async def pump(self) -> None:@@ -80,6 +86,8 @@ class _Conversation:self.events.append((now, message))kind = message.get("type")if kind == p.TRANSCRIPT_INPUT and message.get("final"):+ if self.interrupted_at is not None and self.next_turn_at is None:+ self.next_turn_at = nowawait self.client.send({"type": p.RESPONSE_REQUEST, "session": self.session,"language": self.language})elif kind == p.TOOL_CALL:@@ -101,6 +109,8 @@ class _Conversation:if self.first_audio_at is None:self.first_audio_at = nowself.last_audio_at = now+ if self.interrupted_at is not None and self.next_turn_at is None:+ self.leaked_until = nowawait asyncio.gather(messages(), audio())@@ -123,6 +133,11 @@ class _Conversation:return start+def _speech_start(samples: np.ndarray, threshold: float = 0.01) -> float:+ voiced = np.flatnonzero(np.abs(samples) > threshold)+ return voiced[0] / STT_RATE if voiced.size else 0.0++def _speech_end(samples: np.ndarray, threshold: float = 0.01) -> float:voiced = np.flatnonzero(np.abs(samples) > threshold)return (voiced[-1] + 1) / STT_RATE if voiced.size else 0.0@@ -194,12 +209,12 @@ async def _run(model: str, tts: str, languages: list[str], voice_kind: str,continueawait convo.stream(silence(0.4, STT_RATE))barge = voice.render(_BARGE[language], language, pad_s=0.0)- barge_start = time.monotonic()- await convo.stream(np.concatenate([barge, silence(0.6, STT_RATE)]))+ sent_at = await convo.stream(np.concatenate([barge, silence(0.6, STT_RATE)]))+ barge_start = sent_at + _speech_start(barge)interrupted = convo.interrupted_atleak = 0.0- if interrupted is not None and convo.last_audio_at is not None:- leak = max(0.0, convo.last_audio_at - interrupted)+ if interrupted is not None and convo.leaked_until is not None:+ leak = max(0.0, convo.leaked_until - interrupted)barges.append({"language": language,"interrupt_ms": round((interrupted - barge_start) * 1000.0, 1)diff --git a/jarvis/voice_engine/engine.py b/jarvis/voice_engine/engine.pyindex 4a2b798834..fd8b89c68c 100644--- a/jarvis/voice_engine/engine.py+++ b/jarvis/voice_engine/engine.py@@ -92,7 +92,10 @@ class EngineModels:class EngineConfig:silence_ms: int = 200min_speech_ms: int = 160- barge_in_ms: int = 250+ # Voiced time before the user's speech cancels a reply. 250 ms put the+ # measured cut at ~300 ms after speech onset, over the 250 ms release SLO;+ # the browser's echo cancellation keeps Jarvis' own voice out of the VAD.+ barge_in_ms: int = 160preroll_ms: int = 300turn_threshold: float = 0.5incomplete_wait_ms: int = 1000