准备Python环境
- 建议安装anaconda3: anaconda3
- 创建ESPNET的环境: conda create -n espnet python=3.6
Install ESPNET相关依赖
- 下载espnet源代码 espnet github
- pip install torch==1.0.1 # Installation from anaconda is recommended for PyTorch
- pip install chainer==6.0.0
- pip install cupy==6.0.0 # Do not install cupy as default
- pip install setuptools>=38.5.1
- pip install scipy>=1.3.0
- pip install librosa>=0.7.0
- pip install soundfile>=0.10.2
- pip install inflect>=1.0.0
- pip install unidecode>=1.0.22
- pip install editdistance==0.5.2
- pip install h5py==2.9.0
- pip install tensorboardX==1.8
- pip install pillow>=6.1.0
- pip install nara_wpe>=0.0.5
- pip install museval>=0.2.1
- pip install pystoi>=0.2.2
- pip install kaldiio>=2.13.8
- pip install matplotlib>=3.1.0
- pip install funcsigs>=1.0.2 # A backport of inspect.signature for python2
- pip install configargparse>=0.14.0
- pip install PyYAML>=5.1.2
- pip install sentencepiece>=0.1.82
- pip install pysptk>=0.1.17
- pip install nltk>=3.4.5
- pip install nnmnkwii
- pip install jaconv
- pip install g2p_en
- pip install torch_complex@git+https://github.com/kamo-naoyuki/pytorch_complex.git
- pip install pytorch_wpe@git+https://github.com/nttcslab-sp/dnn_wpe.git
- pip install warpctc-pytorch10-cuda90==0.1.3 (GitHub源码不易安装,pip安装较为简单)
Install 其他依赖
- kaldi相关的依赖: sudo yum install atlas
- sudo yum install bc
- sudo yum install tree
- sudo yum install libsndfile
- sudo yum install flac (librispeech 语音转格式,需要安装flac)
install cuda
- 目前已支持cuda10,安装 sh /path/to/cuda10/cuda_10.0.130_410.48_linux.run
- export LD_LIBRARY_PATH=/path/to/cuda/lib64:$LD_LIBRARY_PATH
Install ESPNET
- conda activate espnet (或者recipe里面的 path.sh)
- pip install -e /path/to/espnet/
Run ESPNET
- conda activate espnet
- cd /path/to/espnet/egs/recipe
- sh run.sh