HTDemucs — GGUF

GGUF conversions of Meta's HTDemucs (Hybrid Transformer Demucs) for use with CrispASR.

HTDemucs is a state-of-the-art music source separation model that splits audio into 4 stems: drums, bass, other, and vocals.

Files

40 of 533 tensors are quantized (the CrossTransformer attention/FFN projections). Encoder/decoder conv weights remain at F16 (required by the ggml Conv2d kernel). Norms, biases, LayerScale, and frequency embeddings stay at F32.

Usage with CrispASR

# Separate audio into stems
crispasr --separate -m htdemucs-f16.gguf -f input.wav

# Auto-download
crispasr --separate --backend htdemucs --auto-download -f input.wav

# Select specific stems
crispasr --separate -m htdemucs-f16.gguf --stems vocals,drums -f input.wav

Python

import crispasr
s = crispasr.Session("htdemucs-f16.gguf", backend="htdemucs")
stems = s.separate(stereo_pcm_44100hz)
# stems = {"drums": array, "bass": array, "other": array, "vocals": array}

Model Details

• Architecture: Hybrid Transformer Demucs (HTDemucs)
• Parameters: 42M
• Input: Stereo audio at 44100 Hz
• Output: 4 stereo stems (drums, bass, other, vocals)
• Segment length: 7.8 seconds (with overlap-add for longer audio)
• License: MIT (Meta)

Architecture

• Dual-branch U-Net: frequency (Conv2d on STFT spectrograms) + time (Conv1d on waveforms)
• 4-layer encoder/decoder with skip connections
• 5-layer CrossTransformer at the bottleneck (alternating self-attention and cross-attention between freq and time branches)
• DConv residual blocks (dilated convolutions) in the encoder
• Complex-as-Channels (CaC) for spectrogram processing
• Frequency embeddings (ScaledEmbedding with smooth initialization)

Conversion

Converted from the official htdemucs pretrained model using:

python models/convert-htdemucs-to-gguf.py --model htdemucs --output htdemucs-f16.gguf --dtype f16
crispasr-quantize htdemucs-f16.gguf htdemucs-q8_0.gguf q8_0
crispasr-quantize htdemucs-f16.gguf htdemucs-q4_k.gguf q4_k

Provenance and EU AI Act Art. 53 note

• Upstream model: facebook/htdemucs.
• Upstream licence: mit. This repository redistributes under the same terms; it grants no rights the upstream licence does not.
• What was done here: format conversion and/or quantisation only (GGUF). No training, no fine-tuning, no merging, no distillation, no change to architecture, vocabulary or capability. Only the numeric representation of the upstream weights differs.
• Training data: documented — where it is documented at all — by the upstream provider; see the upstream model card. No training data was used, added or selected by this repository.
• Provider status: under Regulation (EU) 2024/1689 the upstream authors remain the provider of this model. Converting the serialisation format does not make this repository the provider of a new general-purpose AI model, and no such claim is made. Questions about training content, copyright policy or model capability belong upstream.