Adds mlx-audio-based Qwen3-TTS as an alternative to ElevenLabs,
enabling fully offline voice synthesis with voice cloning from a
short reference audio clip. Benchmarked at 0.52x RTF (sub-realtime)
on Apple Silicon with the 1.7B-Base-4bit model.
Switch from webrtcvad's binary is_speech to Silero VAD's calibrated
float probability via direct ONNX session calls with numpy. The LSTM
provides temporal smoothing natively, eliminating the sliding window
hack. Frame size changes from 480 (30ms) to 512 (32ms) end-to-end
to match Silero's requirements.
Consolidate pipeline/requirements.txt into root requirements.txt,
swap webrtcvad+setuptools for silero-vad+onnxruntime.
Move venv to repo root with combined requirements.txt, fix libopus/portaudio
discovery on macOS, replace deprecated audioop with numpy u-law encoder,
add colored pipeline logging with suppressed third-party noise, fix mic
deadlock on non-speech rejection, fix localhost IP mismatch for test client,
add VAD visualization bar, tune VAD for conversational speech, and move
runtime data to gitignored data/ directory.
Pipeline: async voice pipeline replacing monolithic threaded server.
ASR, LLM, and TTS are independent pluggable services. ASR calls
external parakeet-asr-server, LLM uses any OpenAI-compatible
endpoint, TTS uses ElevenLabs with pluggable backend interface.
Firmware: add mDNS hostname resolution as fallback when multicast
discovery doesn't work. Resolves configured server_hostname via
MDNS.queryHost() on boot, falls back to multicast if resolution fails.
Also adds test_client.py that emulates an ESP32 device for testing
without hardware (TCP server, Opus decode, mic streaming).
