HT-Demucs ONNX Models (ipsilondev/htdemucs-onnx)

Clean ONNX model distribution of HT-Demucs (Demucs v4) for 4-stem real-time audio separation (Drums, Bass, Other, Vocals). Optimized for mobile, edge, and desktop runtimes via ONNX Runtime.

Model Files

File	Size	Precision	Target Deployment
htdemucs_fp16weights.onnx	166 MB	FP16 weights / FP32 I/O	Mobile (Android / iOS), low RAM footprint
htdemucs.onnx	316 MB	Full FP32	Desktop, server, high-precision inference

Tensor Specifications

• Input (mix):
  • Shape: [1, 2, 343980] (Batch: 1, Channels: 2 Stereo, Samples: 343,980 at 44.1 kHz, ~7.8 seconds)
  • Type: loat32
• Output (stems):
  • Shape: [1, 4, 2, 343980]
  • Stem channel order:
    1. Index 0: drums
    2. Index 1: ass
    3. Index 2: other
    4. Index 3: ocals
  • Type: loat32

Recommended Inference Configuration

• Chunking: Chunked overlap-add with linear/cosine trapezoidal window
  • Segment: 343,980 samples (~7.8s)
  • Overlap: 85,995 samples (25%)
  • Stride: 257,985 samples
• Sample Rate: 44,100 Hz stereo