mirror of
https://github.com/justLV/onju-v2
synced 2026-04-21 15:47:55 +00:00
13f9d59245
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.
52 lines
1.6 KiB
Text
52 lines
1.6 KiB
Text
asr:
|
|
url: "http://localhost:8100"
|
|
|
|
llm:
|
|
base_url: "http://localhost:8080/v1" # mlx_lm.server (or Ollama, OpenRouter, OpenAI, etc.)
|
|
api_key: "none" # set if using a hosted API
|
|
model: "mlx-community/gemma-3-4b-it-qat-4bit"
|
|
max_messages: 20
|
|
max_tokens: 300
|
|
system_prompt: "You are a helpful voice assistant. Keep responses concise (under 2 sentences)."
|
|
|
|
tts:
|
|
backend: "qwen3" # "qwen3" (local) or "elevenlabs" (cloud)
|
|
qwen3:
|
|
url: "http://localhost:8880" # mlx-audio server (start with: python -m mlx_audio.server --port 8880)
|
|
model: "mlx-community/Qwen3-TTS-12Hz-1.7B-Base-4bit"
|
|
ref_audio: "" # path to 3-10s reference clip for voice cloning
|
|
ref_text: "" # optional transcript of reference audio
|
|
elevenlabs:
|
|
api_key: "" # your ElevenLabs API key
|
|
default_voice: "Rachel"
|
|
voices:
|
|
Rachel: "21m00Tcm4TlvDq8ikWAM" # add your voice IDs here
|
|
|
|
vad:
|
|
threshold: 0.5 # speech onset probability
|
|
neg_threshold: 0.35 # speech offset probability (hysteresis)
|
|
silence_time: 1.5
|
|
pre_buffer_s: 1.0
|
|
|
|
network:
|
|
udp_port: 3000
|
|
tcp_port: 3001
|
|
multicast_group: "239.0.0.1"
|
|
multicast_port: 12345
|
|
|
|
audio:
|
|
sample_rate: 16000
|
|
chunk_size: 512 # 32ms at 16kHz (Silero VAD requirement)
|
|
opus_frame_size: 320 # 20ms at 16kHz
|
|
|
|
device:
|
|
default_volume: 14
|
|
default_mic_timeout: 60
|
|
led_fade: 6
|
|
led_power: 35
|
|
led_update_period: 0.2
|
|
persist_file: "data/devices.json"
|
|
greeting_wav: "data/hello_imhere.wav"
|
|
|
|
logging:
|
|
level: "INFO"
|