shahules786 / mayavoz

Pytorch based speech enhancement toolkit.

mayavoz is a Pytorch-based opensource toolkit for speech enhancement. It is designed to save time for audio practioners & researchers. It provides easy to use pretrained speech enhancement models and facilitates highly customisable model training.

| Quick Start | Installation | Tutorials | Available Recipes | Demo

Key features 🔑

• Various pretrained models nicely integrated with huggingface hub 🤗 that users can select and use without any hastle.
• 📦 Ability to train and validate your own custom speech enhancement models with just under 10 lines of code!
• 🪄 A command line tool that facilitates training of highly customisable speech enhacement models from the terminal itself!
• ⚡ Supports multi-gpu training integrated with Pytorch Lightning.
• 🛡️ data augmentations integrated using torch-augmentations

Demo

Noisy speech followed by enhanced version.

mayavoz_demo.mp4

Quick Start 🔥

from mayavoz.models import Mayamodel

model = Mayamodel.from_pretrained("shahules786/mayavoz-waveunet-valentini-28spk")
model.enhance("noisy_audio.wav")

Recipes

Model	Dataset	STOI	PESQ	URL
WaveUnet	Valentini-28spk	0.836	2.78	shahules786/mayavoz-waveunet-valentini-28spk
Demucs	Valentini-28spk	0.961	2.56	shahules786/mayavoz-demucs-valentini-28spk
DCCRN	Valentini-28spk	0.724	2.55	shahules786/mayavoz-dccrn-valentini-28spk
Demucs	MS-SNSD-20hrs	0.56	1.26	shahules786/mayavoz-demucs-ms-snsd-20

Test scores are based on respective test set associated with train dataset.

See tutorials to train your custom model

Installation

Only Python 3.8+ is officially supported (though it might work with Python 3.7)

• With Pypi

pip install mayavoz

• With conda

conda env create -f environment.yml
conda activate mayavoz

• From source code

git clone url
cd mayavoz
pip install -e .

Support

For commercial enquiries and scientific consulting, please contact me.

Acknowledgements

Sincere gratitude to AMPLYFI for supporting this project.