This repo contains model and training code for extracting multiple speaker emebddings from noisy mixture. This doesn't require number of speakers information prior, but instead computes speaker emebdding recursively until a stop criterion.

Installing requirements

pip install -r requirements.txt

Changing parameters

You can change the model, dataset, loss and the training parameters in config.py

Training code

You can start training the model by running the following in the terminal python trainer.py