config.example.yaml

# vTTS Configuration
# Copy this file to config.yaml and adjust to your setup.

model:
  # HuggingFace model ID (downloaded automatically on first run)
  # Options:
  #   - Qwen/Qwen3-TTS-12Hz-0.6B-Base          (voice cloning, 1.2GB VRAM)
  #   - Qwen/Qwen3-TTS-12Hz-1.7B-Base          (voice cloning, 3.4GB VRAM)
  #   - Qwen/Qwen3-TTS-12Hz-0.6B-CustomVoice   (built-in speakers, 1.2GB VRAM)
  #   - Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice   (built-in speakers, 3.4GB VRAM)
  name: "Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice"

  # GPU device
  device: "cuda:0"

  # Model precision (bfloat16 recommended for Ampere+ GPUs)
  dtype: "bfloat16"

  # Optional: local path if you already have the model downloaded.
  # If set, vTTS uses this path instead of downloading from HuggingFace.
  # local_path: "/path/to/your/local/model"

worker:
  # Maximum concurrent audio generations
  max_slots: 16

  # Frames between audio chunk emissions.
  # Lower = lower latency (faster first byte), more vocoder overhead.
  #   3 = ~250ms chunks, ~200ms TTFB (best for voice agents)
  #   6 = ~500ms chunks, ~400ms TTFB (good balance)
  #  20 = ~1.7s chunks, ~1.4s TTFB (best for batch dubbing)
  emit_every_frames: 6

  # Vocoder context window (more = better quality, more compute)
  decode_window_frames: 80

  # Crossfade overlap between chunks (samples at 24kHz)
  overlap_samples: 1024

  # Maximum generation length in frames (12 frames = 1 second of audio)
  # Safety limit to prevent infinite generation.
  max_frames: 3600  # 5 minutes

server:
  host: "0.0.0.0"
  port: 8080

# Voice cloning configuration (only for Base models).
# Each voice needs:
#   - ref_audio: path to a WAV file (~5-10 seconds of clean speech)
#   - ref_text: exact transcript of what is said in the audio
#
# voices:
#   my_voice:
#     ref_audio: "./voices/my_voice.wav"
#     ref_text: "The exact words spoken in the reference audio."
#   another_voice:
#     ref_audio: "./voices/another.wav"
#     ref_text: "Another reference transcript."