RE-SepFormer trained on Libri2Mix dataset
This model custom-trained Resource-Efficient Separation Transformer (RE-SepFormer) was designed to separate overlapping speech from audio mixtures into two separate sources.
Model Details
- Developed by: Jahmori Richardson
- Model type: Audio Source Separation (Deep Learning)
- Language(s) (NLP): English
- License: Apache 2.0
Model Sources [optional]
- Repository: [More Information Needed]
- Paper: https://arxiv.org/abs/2206.09507
Uses
Isolating individual speakers from mixed audio files containing overlapping speech.
Downstream Use [optional]
[More Information Needed]
Out-of-Scope Use
It is not intended for non-speech audio separation or audio with heavy music/complex noise.
Bias, Risks, and Limitations
Model performance may vary with audio inputs that contain high amounts of background, reverberation, or more than required speakers that the model was designed for.
How to Get Started with the Model
Use the code below to get started with the model.
[More Information Needed]
Training Details
Training Data
| Mean SI-SNRi (dB) | Mean SDRi (dB) |
|---|---|
| 15.8 | 16.1 |
[More Information Needed]
Training Procedure
Training Hyperparameters
- Training regime: fp32 mixed precision