jarvis-voice-pipeline

Use one persistent sounddevice OutputStream opened at startup instead of sd.play() per call, which stalls about 1.4 seconds opening and closing a PortAudio stream. Serialize writes with a threading.Lock and play the wake acknowledgment in a background thread so listening starts immediately.

Install piper-tts, download the voice .onnx and .onnx.json files from HuggingFace, then load with PiperVoice.load(model, config). The synthesize() method yields AudioChunk objects with audio_float_array and sample_rate, and speed is controlled via SynthesisConfig(length_scale=...).

Piper is faster and more characterful for this use case, outputting at 22050 Hz versus Kokoro's 24000 Hz. Piper's API differs: synthesize returns an iterator of AudioChunk objects rather than a single audio tuple, and the code falls back to Kokoro bm_george if Piper is unavailable.

subprocess.PIPE stalls because sd.InputStream mic initialization blocks in a headless subprocess context without a real console. Launch via start "" /min cmd /c with output redirected to a log file, then tail that log file for state updates instead of reading pipes.

Use the tiny model with cpu_threads set to os.cpu_count() for maximum decode speed. Enable vad_filter=True with aggressive vad_parameters and pre-trim silence from the audio before calling transcribe to cut the amount of audio processed.

The assistant's own TTS output feeds back into the microphone and re-triggers the wake detector, causing repeated acknowledgment loops. Drain the mic queue after the assistant finishes speaking so its own voice cannot re-feed the wake detection.