Moises-Light

This is an unofficial PyTorch implementation of the Moises-Light architecture from "Moises-Light: Resource-efficient Band-split U-Net for Music Source Separation" (Hung et al., WASPAA 2025). The paper does not release code; this is an independent implementation based on the paper's description and the open-source implementations of DTTNet, BS-RoFormer, and SCNet.

Installation

Install from PyPI

pip install moises-light

Install from GitHub

pip install git+https://github.com/crlandsc/moises-light.git

Or, you can clone the repository and install it in editable mode for development:

git clone https://github.com/crlandsc/moises-light.git
cd moises-light
pip install -e .

Dependencies

• PyTorch (>=2.0)
• einops
• rotary-embedding-torch

Quick Start

import torch
from moises_light import MoisesLight, configs

# Use a preset
model = MoisesLight(**configs['paper_large'])

# Forward pass
x = torch.randn(1, 2, 264600)  # [batch, channels, samples] 6s @ 44.1kHz
y = model(x)                   # [1, 4, 2, 264600] = [batch, stems, channels, samples]

# With auxiliary outputs interface (for training framework compatibility)
y, aux = model(x, return_auxiliary_outputs=True)

Preset Configurations

All presets use n_fft=6144, hop_size=1024, stereo input, and 4-stem output (vocals, drums, bass, other).

The paper truncates the STFT at 2048 bins (~14.7 kHz), zeroing everything above. While the original DTTNet paper noted that this truncation has little to no effect on SI-SDR scores, in practice these high frequencies are critical for perceptual audio quality — vocal air, cymbal shimmer, synth brightness, etc. all live above 15 kHz. This package includes fullband presets that extend processing to the full 0-22 kHz spectrum.

Extending to fullband requires increasing n_bands from 4 to 6 (to maintain 512 bins per band), and G must be divisible by n_bands for group convolutions. Since 56 is not divisible by 6, G must change. Two strategies are provided:

1. Fullband matched-param — Pick the nearest valid G that keeps total params similar to the paper (G=60 or 36). This trades per-band capacity for full spectrum coverage within the same parameter budget. SI-SDR may decrease slightly since the same capacity is spread across 2 additional high-frequency bands.
2. Fullband wide — Pick G so that G/n_bands matches the paper's per-group channel count (84/6=14, matching 56/4=14). Each band retains the same representation power as the paper model, but total params increase ~1.8x. This may preserve metric performance while gaining full spectrum coverage.

Paper-Faithful (truncated spectrum, 0-14.7 kHz)

Faithful to the paper's architecture. Frequencies above ~14.7 kHz are zeroed.

Fullband Matched-Param (full spectrum, 0-22 kHz, similar param budget)

Full spectrum via 6 bands of 512 bins (freq_dim=3072). G adjusted to keep param count close to paper variants.

Fullband Wide (full spectrum, 0-22 kHz, matched per-group capacity)

Full spectrum with the same per-group channel capacity as the paper models.

Architecture

Moises-Light builds on the DTTNet foundation (a symmetric U-Net with TFC-TDF encoder/decoder blocks and dual-path RNN bottleneck) and integrates improvements from BS-RoFormer and SCNet:

• Band splitting via group convolutions (inspired by BSRNN/BS-RoFormer): Instead of DTTNet's full-spectrum convolutions, the STFT is divided into n_bands equal-width subbands and processed with group convolutions (Split Module). This replaces DTTNet's first/last 1x1 convolutions and dramatically reduces parameters compared to the original band-split MLPs in BSRNN.
• Split and Merge Module (replaces DTTNet's TFC-TDF V3 blocks): Group conv blocks with n_bands groups replace the original TFC layers, so each band is processed independently. The TDF (Time-Distributed Frequency FC) bottleneck is retained but now operates on per-band frequency dimensions (freq_dim / n_bands), which is n_bands times cheaper.
• RoPE transformer bottleneck (from BS-RoFormer): DTTNet's dual-path RNN is replaced with dual-path RoPE transformers for sequence modeling along both frequency and time axes. This improves performance without significantly increasing parameters.
• Asymmetric encoder/decoder (from SCNet): The encoder has n_enc heavy stages (each with a full Split and Merge block), while the decoder uses only n_dec heavy stages plus n_enc - n_dec light stages (upsample + skip connection only, no Split and Merge). This saves significant compute in the decoder.
• Frequency truncation (from DTTNet): Only freq_dim of the n_fft/2+1 STFT bins are processed; the rest are zero-padded for iSTFT reconstruction. Paper presets truncate at ~14.7 kHz; fullband presets extend to ~22 kHz.
• Multiplicative skip connections (from DTTNet): Decoder stages combine upsampled features with encoder skip connections via element-wise multiplication rather than concatenation or addition.

Implementation Notes

This is an independent implementation — the paper does not release code. The following decisions were made where the paper was ambiguous or where I diverged:

• Asymmetric decoder interpretation: The paper specifies N_enc=3, N_dec=1 (Table 1) but doesn't explicitly state what happens with the remaining 2 decoder stages. I interpret N_dec=1 as 1 heavy stage (with Split and Merge processing) and 2 light stages (upsample + skip connection only), matching the SCNet asymmetric pattern.

• Time-only downsampling: DTTNet downsamples both time and frequency dimensions (T/2^N and F/2^N). Our implementation only downsamples time. The paper states that band-splitting "allows us to remove frequency pooling or upsampling across all DTTNet layers" (Sec 3.1), but doesn't explicitly confirm this removal in the final architecture.

• Transformer hyperparameters: The paper does not specify the RoPE transformer's internal dimensions. I use heads=4, dim_head=32, ff_mult=2 — chosen to keep the bottleneck lightweight and consistent with the model's parameter budget.

• Multiplicative masking: The paper states the model "directly generating the separated target spectrogram." By default (use_mask=True), our implementation applies multiplicative masking on the original STFT (i.e., the network predicts a mask rather than the spectrogram directly). This is a common and effective approach in other SOTA models like BS-RoFormer and often leads to better perceptual quality, particularly for silent segments. Setting use_mask=False switches to the paper's direct spectrogram generation mode, where the network output produces spectrograms directly.

• Z-score normalization: The paper does not mention input normalization. I apply Z-score normalization (zero mean, unit variance) to the STFT features before the U-Net, inspired by HTDemucs-style preprocessing. This is standard practice in similar architectures and stabilizes training.

• TDF bottleneck factor (bn_factor): The paper does not specify this parameter. DTTNet uses bn_factor=8 for vocals, drums, and other, and bn_factor=2 for bass (bass has narrower frequency range and more tonal structure, benefiting from higher TDF capacity). This implementation defaults to bn_factor=8 to match DTTNet's majority-stem setting. For single-stem bass models, consider bn_factor=2.

• Multi-stem output: The paper trains separate per-stem models (4x ~5M params for VDBO). This implementation outputs all stems simultaneously via a shared encoder and source head, as this paradigm has proven effective in other U-Net models like HTDemucs and SCNet. To reproduce the paper's approach, train 4 separate single-stem models (e.g., MoisesLight(sources=['vocals'])).

Key Parameters

Integration

Custom Training Loop

model = MoisesLight(**configs['paper_large'])
optimizer = torch.optim.AdamW(model.parameters(), lr=5e-4)

for batch in dataloader:
    mix = batch['mix']          # [B, 2, L]
    targets = batch['targets']  # [B, 4, 2, L]
    pred = model(mix)           # [B, 4, 2, L]
    loss = criterion(pred, targets)
    loss.backward()
    optimizer.step()
    optimizer.zero_grad()

Known Limitations

• MPS (Apple Silicon): There is a bug in the MPS implementation of torch.istft. The model automatically falls back to CPU for iSTFT when on MPS, which adds overhead. This is a PyTorch limitation, not a model issue.
• Frequency truncation: Paper presets zero frequencies above ~14.7 kHz. Use fullband presets if high-frequency content matters.

Citation

@inproceedings{hung2025moises,
  title={Moises-Light: Resource-efficient Band-split U-Net for Music Source Separation},
  author={Hung, Yun-Ning and Pereira, Igor and Korzeniowski, Filip},
  booktitle={2025 IEEE Workshop on Applications of Signal Processing to Audio and Acoustics (WASPAA)},
  pages={1--5},
  year={2025},
  doi={10.1109/WASPAA66052.2025.11230925}
}

Contributing

Contributions are welcome! Please open an issue or submit a pull request if you have any bug fixes, improvements, or new features to suggest.

License

This project is licensed under the MIT License - see the LICENSE file for details.