Open-source code for ICASSP 2025

In this project, we will introduce: 1) How to train the Target Speaker Extraction (TSE) models using our method. 2) How to use the Text-to-Speech (TTS) system to conduct short enrollment speech augmentation, as proposed in our paper. 3) How to evaluate the TSE models using enrollment speeches.

Prepare the environment

cd Aug4TSE
conda create --name Aug4TSE python=3.8.18
conda activate Aug4TSE
pip install -r requirements.txt

Follow the steps in ./TSE to create:

• {train,valid}.npy: enrollment speeches' speaker embeddings
• _ckpt_epoch_xxx.ckpt: a TSE model's checkpoint

Follow the steps in ./data_preparation to create:

• {augmented_speech}.wav: augmented speeches based on short enrollment speeches
• {augmented_speech}.npy: speaker embeddings of augmented speeches

Follow the steps in ./data_preparation to create:

• {test}.npy: the speaker embeddings of enrollment speeches that you want to use to evaluate TSE models

Follow the steps in ./TSE to create:

• evaluation.log: the evaluation results