backend/configs/config.yaml

whisper:
  # Default implementation is faster-whisper. This indicates model name within `models\Whisper\faster-whisper`
  model_size: large-v2
  # Compute type. 'float16' for CUDA, 'float32' for CPU.
  compute_type: float16

bgm_separation:
  # UVR model sizes between ["UVR-MDX-NET-Inst_HQ_4", "UVR-MDX-NET-Inst_3"]
  model_size: UVR-MDX-NET-Inst_HQ_4
  # Whether to offload the model after the inference. Should be true if your setup has a VRAM less than <16GB
  enable_offload: true
  # Device to load BGM separation model
  device: cuda

# Settings that apply to the `cache' directory. The output files for `/bgm-separation` are stored in the `cache' directory,
# (You can check out the actual generated files by testing `/bgm-separation`.)
# You can adjust the TTL/cleanup frequency of the files in the `cache' directory here.
cache:
  # TTL (Time-To-Live) in seconds, defaults to 10 minutes
  ttl: 600
  # Clean up frequency in seconds, defaults to 1 minutes
  frequency: 60