mossformer2-se-48k

MossFormer2 Speech Enhancement 48 kHz

A 48 kHz MossFormer2-based speech-enhancement model. The checkpoint contains 55,262,170 parameters and is stored in Safetensors format.

Model details

• Sampling rate: 48,000 Hz
• FFT/window length: 1,920 samples
• Hop length: 384 samples
• Mel bins: 60
• Input feature channels: 180
• Output frequency bins: 961
• MossFormer2 blocks: 24

The Transformers model predicts a non-negative time-frequency mask. Waveform loading, feature extraction, mask application, and waveform reconstruction are preprocessing/postprocessing responsibilities and are not bundled as a Transformers pipeline.

Loading the model

The architecture uses custom Transformers code from this repository, so review the code and enable trust_remote_code when loading it.

from transformers import AutoConfig, AutoModel

repo_id = "MigoXV/mossformer2-se-48k"
config = AutoConfig.from_pretrained(repo_id, trust_remote_code=True)
model = AutoModel.from_pretrained(repo_id, trust_remote_code=True)
model.eval()

print(config.sampling_rate)  # 48000

Runtime dependencies for the custom architecture include PyTorch, Transformers, Einops, and rotary-embedding-torch.

Input and output

input_features must be a float tensor shaped [batch, 180, frames]. The returned mask is shaped [batch, frames, 961]. Use the spectral parameters in config.json to build a compatible 48 kHz frontend and inverse transform.

Intended use

This checkpoint is intended for speech enhancement and denoising research or inference at 48 kHz. Validate output quality and behavior on your target audio domain before production use.

Limitations

Training data, evaluation metrics, demographic analysis, and the original training procedure are not documented in the available checkpoint artifacts. No claim is made about performance on unseen languages, speakers, recording devices, or noise conditions.