Counting 2,784 Big Data & Machine Learning Frameworks, Toolsets, and Examples...
Suggestion? Feedback? Tweet @stkim1

Last Commit
May. 24, 2018
Dec. 13, 2017

ESPnet: end-to-end speech processing toolkit

Build Status

ESPnet is an end-to-end speech processing toolkit, mainly focuses on end-to-end speech recognition. ESPnet uses chainer and pytorch as a main deep learning engine, and also follows Kaldi style data processing, feature extraction/format, and recipes to provide a complete setup for speech recognition and other speech processing experiments.

Key Features

  • Hybrid CTC/attention based end-to-end ASR
    • Fast/accurate training with CTC/attention multitask training
    • CTC/attention joint decoding to boost monotonic alignment decoding
  • Encoder: VGG-like CNN + BLSTM or pyramid BLSTM
  • Attention: Dot product, location-aware attention, variants of multihead (pytorch only)
  • Incorporate RNNLM/LSTMLM trained only with text data
  • Flexible network architecture thanks to chainer and pytorch
  • Kaldi style complete recipe
    • Support numbers of ASR benchmarks (WSJ, Switchboard, CHiME-4, Librispeech, TED, CSJ, AMI, HKUST, Voxforge, etc.)
  • State-of-the-art performance in Japanese/Chinese benchmarks (comparable/superior to hybrid DNN/HMM and CTC)
  • Moderate performance in standard English benchmarks


  • Python2.7+

  • Cuda 8.0 (for the use of GPU)

  • Cudnn 6 (for the use of GPU)

  • NCCL 2.0+ (for the use of multi-GPUs)

  • PyTorch 0.3.x (no support for PyTorch 0.4.x)

  • Chainer 4.x+


Install Kaldi, Python libraries and other required tools using system python and virtualenv

$ cd tools
$ make -j

or using local miniconda

$ cd tools
$ make -f -j

For higher version (>4.9) of gcc and cuda 9.1 use following command:

$ cd tools
$ make -j -f Makefile.cuda91.gcc6

You can compare Makefile and Makefile.cuda91.gcc6 to change makefile accordingly for other version of gcc/cuda.

To use cuda (and cudnn), make sure to set paths in your .bashrc or .bash_profile appropriately.



If you want to use multiple GPUs, you should install nccl and set paths in your .bashrc or .bash_profile appropriately, for example:


export CPATH=$NCCL_ROOT/include:$CPATH

Execution of example scripts

Move to an example directory under the egs directory. We prepare several major ASR benchmarks including WSJ, CHiME-4, and TED. The following directory is an example of performing ASR experiment with the VoxForge Italian Corpus.

$ cd egs/voxforge/asr1

Once move to the directory, then, execute the following main script with a chainer backend:

$ ./

or execute the following main script with a pytorch backend (currently the pytorch backend does not support VGG-like layers):

$ ./ --backend pytorch --etype blstmp

With this main script, you can perform a full procedure of ASR experiments including

Use of GPU

If you use GPU in your experiment, set --ngpu option in appropriately, e.g.,

# use single gpu
$ ./ --ngpu 1

# use multi-gpu
$ ./ --ngpu 3

# if you want to specify gpus, set CUDA_VISIBLE_DEVICES as follows
# (Note that if you use slurm, this specification is not needed)
$ CUDA_VISIBLE_DEVICES=0,1,2 ./ --ngpu 3

# use cpu
$ ./ --ngpu 0

Default setup uses CPU (--ngpu 0).

Note that if you want to use multi-gpu, the installation of nccl is required before setup.

Error due to ACS (Multiple GPUs)

When using multiple GPUs, if the training freezes or lower performance than expected is observed, verify that PCI Express Access Control Services (ACS) are disabled. Larger discussions can be found at: link1 link2 link3. To disable the PCI Express ACS follow instructions written here. You need to have a ROOT user access or request to your administrator for it.

Docker Container

To work inside a docker container, execute located inside the docker directory. It will build a container and execute the main program specified by the following GPU, ASR example, and outside directory information, as follows:

$ cd docker
$ ./ --docker_gpu 0 --docker_egs chime4/asr1 --docker_folders /export/corpora4/CHiME4/CHiME3 --dlayers 1 --ngpu 1 

The docker container is built based on the CUDA and CUDNN version installed in your computer. The arguments required for the docker configuration have a prefix "--docker" (e.g., --docker_gpu, --docker_egs, --docker_folders). accept all normal ESPnet arguments, which must be followed by these docker arguments. Multiple GPUs should be specified with the following options:

$ cd docker
$ ./ --docker_gpu 0,1,2 --docker_egs chime5/asr1 --docker_folders /export/corpora4/CHiME5 --ngpu 3

Note that all experimental files and results are created under the normal example directories (egs/<example>/).

Setup in your cluster

Change according to your cluster setup. If you run experiments with your local machine, please use default For more information about see It supports Grid Engine (, SLURM (, etc.

Error due to matplotlib

If you have the following error (or other numpy related errors),

RuntimeError: module compiled against API version 0xc but this version of numpy is 0xb
Exception in main training loop: numpy.core.multiarray failed to import
Traceback (most recent call last):
from . import _path, rcParams
ImportError: numpy.core.multiarray failed to import

Then, please reinstall matplotlib with the following command:

$ cd egs/voxforge/asr1
$ . ./
$ pip install pip --upgrade; pip uninstall matplotlib; pip --no-cache-dir install matplotlib

CTC, attention, and hybrid CTC/attention

ESPnet can completely switch the mode from CTC, attention, and hybrid CTC/attention

# hybrid CTC/attention (default)
#  --mtlalpha 0.5 and --ctc_weight 0.3 in most cases
$ ./

# CTC mode
$ ./ --mtlalpha 1.0 --ctc_weight 1.0 --recog_model

# attention mode
$ ./ --mtlalpha 0.0 --ctc_weight 0.0

The CTC training mode does not output the validation accuracy, and the optimum model is selected with its loss value (i.e., --recog_model About the effectiveness of the hybrid CTC/attention during training and recognition, see [1] and [2].


We list the character error rate (CER) and word error rate (WER) of major ASR tasks.

CER (%) WER (%)
WSJ dev93 5.3 12.4
WSJ eval92 3.6 8.9
CSJ eval1 8.5 N/A
CSJ eval2 6.1 N/A
CSJ eval3 6.8 N/A
HKUST train_dev 29.7 N/A
HKUST dev 28.3 N/A
Librispeech dev_clean 2.9 7.7
Librispeech test_clean 2.7 7.7

Chainer and Pytorch backends

Chainer Pytorch
Multi-GPU supported supported
VGG-like encoder supported no support
RNNLM integration supported supported
#Attention types 3 (no attention, dot, location) 12 including variants of multihead

References (Please cite the following articles)

[1] Suyoun Kim, Takaaki Hori, and Shinji Watanabe, "Joint CTC-attention based end-to-end speech recognition using multi-task learning," Proc. ICASSP'17, pp. 4835--4839 (2017)

[2] Shinji Watanabe, Takaaki Hori, Suyoun Kim, John R. Hershey and Tomoki Hayashi, "Hybrid CTC/Attention Architecture for End-to-End Speech Recognition," IEEE Journal of Selected Topics in Signal Processing, vol. 11, no. 8, pp. 1240-1253, Dec. 2017