thanks to Ke Wang and awni's repo
git https://github.com/asappresearch/sru.git
cd sru
pip install -r requirements.txt
python setup.py install
git clone https://github.com/vBaiCai/python-pesq.git
cd python-pesq
python setup.py install
pip install pystoi
find ${aishell_dir}/train -iname "*.wav" > train.lst
find ${aishell_dir}/dev -iname "*.wav" > dev.lst
find ${aishell_dir}/test -iname "*.wav" > test.lst
find ${musan}/test -iname "*.wav" > noise.lst # Attention!!, please do not add musan/speech into noise.lst
bash tools/prepare_train.sh
bash tools/prepare_test.sh
bash run_sruc.sh
bash eval.sh
add clip_grad_norm in step/run_cldnn.py fix memory leak bug in ./tools/dataset.py's collat_fn: return numpy.array to torch.tensor, which can be lead to memory leak change mse in frames to mse in samples
- change the conv2d to like tf's 'same' padding
- add SRU: SRU
- add 1k6h train strategy: warmup
- add eval
- update data prepare pipeline
- add psa,psm