funcwj / aps

A personal toolkit for single/multi-channel speech recognition & enhancement & separation.

APS

My open-source toolkit for speech related tasks, e.g., single/multi-channel speech enhancement & separation & recognition. The goal is to simplify the training and evaluation procedure and make it easy and flexible for me to do experiments and verify neural network based methods.

Setup

git clone https://github.com/funcwj/aps
# set up the python environments
# 1) using "pip install -r requirements.txt" or
# 2) create conda enviroments based on requirements.txt (recommended, see docker/Dockerfile)
cd aps && pip install -r requirements.txt # the optional packages are not shown in requirements.txt

For developers (who want to make commits or PRs), continue to run

# set up the git hook scripts
pip install -r requirements-dev.txt && pre-commit install

to setup the development environments. To build C++ sources and demo commands, running

mkdir build && cd build
cmake .. && make -j

Introduction

• Feature
• Overivew
• Structure
• Instruction
• Q & A
• Issues
• Examples

Acknowledge

The project was started at early 2019 when the author was a master student of the Audio, Speech and Language Processing Group (ASLP) in Northwestern Polytechnical University (NWPU), Xi'an, China. Originally it was used to collect the source code of the experiments that the author did in the past.

aps's Issues

could you share the pretrain model for multi-channel speech enhancement? thanks!

could you share the pretrain model for multi-channel speech enhancement? thanks!

About PSD matrix

aps/aps/asr/filter/mvdr.py

Line 57 in 3f46e47

nominator = (spec * mask) @ spec.conj_transpose(-1, -2)

您好，我想问一下，这行代码是MVDR的分子部分，应该是mask之后的特征和其共轭转置矩阵乘，您的写法后面使用的是mask之前的特征，不会有问题嘛？