KoSpeech: Open Source Project for Korean End-to-End Automatic Speech Recognition in PyTorch
Soohwan Kim1,2, Seyoung Bae1, Cheolhwang Won1, Suwon Park1*
1Elcomm, Kwangwoon Univ. 2Spoken Language Lab (of Sogang Univ.)
* author is advisor to this work.
End-to-end (E2E) automatic speech recognition (ASR) is an emerging paradigm in the field of neural network-based speech recognition that offers multiple benefits. Traditional “hybrid” ASR systems, which are comprised of an acoustic model, language model, and pronunciation model, require separate training of these components, each of which can be complex.
For example, training of an acoustic model is a multi-stage process of model training and time alignment between the speech acoustic feature sequence and output label sequence. In contrast, E2E ASR is a single integrated approach with a much simpler training pipeline with models that operate at low audio frame rates. This reduces the training time, decoding time, and allows joint optimization with downstream processing such as natural language understanding.
KoSpeech
is project for End-to-end (E2E) automatic speech recognition implemented in PyTorch.
KoSpeech
has modularized and extensible components for las models, training and evalutaion, checkpoints, parsing etc.
We appreciate any kind of feedback or contribution.
We used KsponSpeech
corpus which containing 1000h of Korean speech data.
At present our model has recorded an 87.99% CRR, and we are working for a higher recognition rate.
Also our model has recorded 92.0% CRR in Kaldi-zeroth corpus
-
Multi-headed (location-aware / scaled dot-product) Attention
-
Scheduled Sampling (Teacher forcing scheduling)
-
Inference with batching
-
Multi-GPU training
We have referred to many papers to develop the best model possible. And tried to make the code as efficient and easy to use as possible. If you have any minor inconvenience, please let us know anytime. We will response as soon as possible.
Our architecture based on Seq2seq with Attention.
Attention mechanism
helps finding speech alignment. We apply (location-aware
/ multi-head
) attention which you can choose. Location-aware attention proposed in Attention Based Models for Speech Recognition
paper and Multi-headed attention proposed in Attention Is All You Need
paper. You can choose between these two options as attn_mechanism
option. Please check this page.
We mainly referred to following papers.
「Attention Based Models for Speech Recognition」
「State-of-the-art Speech Recognition with Sequence-to-Sequence Models」
「SpecAugment: A Simple Data Augmentation Method for Automatic Speech Recognition」.
If you want to study the feature of audio, we recommend this papers.
「Voice Recognition Using MFCC Algirithm」.
Our model architeuture is as follows.
ListenAttendSpell(
(listener): Listener(
(extractor): VGGExtractor(
(cnn): MaskCNN(
(sequential): Sequential(
(0): Conv2d(1, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
(1): Hardtanh(min_val=0, max_val=20, inplace=True)
(2): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(3): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
(4): Hardtanh(min_val=0, max_val=20, inplace=True)
(5): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
(6): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(7): Conv2d(64, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
(8): Hardtanh(min_val=0, max_val=20, inplace=True)
(9): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(10): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
(11): Hardtanh(min_val=0, max_val=20, inplace=True)
(12): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
)
)
)
(rnn): LSTM(2560, 256, num_layers=3, batch_first=True, dropout=0.3, bidirectional=True)
)
(speller): Speller(
(embedding): Embedding(2038, 512)
(input_dropout): Dropout(p=0.3, inplace=False)
(rnn): LSTM(512, 512, num_layers=2, batch_first=True, dropout=0.3)
(attention): MultiHeadAttention(
(query_proj): Linear(in_features=512, out_features=512, bias=True)
(value_proj): Linear(in_features=512, out_features=512, bias=True)
)
(fc1): Linear(in_features=1024, out_features=512, bias=True)
(fc2): Linear(in_features=512, out_features=2038, bias=True)
)
)
kospeech
module has modularized and extensible components for las models, trainer, evaluator, checkpoints etc...
In addition, kospeech
enables learning in a variety of environments with a simple option setting.
- Options
usage: main.py [-h] [--mode] [--sample_rate]
[--window_size] [--stride] [--n_mels]
[--normalize] [--del_silence] [--input_reverse]
[--feature_extract_by] [--time_mask_para] [--freq_mask_para]
[--time_mask_num] [--freq_mask_num]
[--use_bidirectional] [--hidden_dim]
[--dropout] [--num_heads] [--label_smoothing]
[--listener_layer_size] [--speller_layer_size] [--rnn_type]
[--extractor] [--activation]
[--attn_mechanism] [--teacher_forcing_ratio]
[--dataset_path] [--data_list_path]
[--label_path] [--init_uniform] [--spec_augment]
[--noise_augment] [--noiseset_size]
[--noise_level] [--use_cuda]
[--batch_size] [--num_workers]
[--num_epochs] [--init_lr]
[--high_plateau_lr] [--low_plateau_lr] [--valid_ratio]
[--max_len] [--max_grad_norm]
[--rampup_period] [--decay_threshold] [--exp_decay_period]
[--teacher_forcing_step] [--min_teacher_forcing_ratio]
[--seed] [--save_result_every]
[--checkpoint_every] [--print_every] [--resume]
We are constantly updating the progress of the project on the Wiki page. Please check this page.
This project recommends Python 3.7 or higher.
We recommend creating a new virtual environment for this project (using virtual env or conda).
- Numpy:
pip install numpy
(Refer here for problem installing Numpy). - Pytorch: Refer to PyTorch website to install the version w.r.t. your environment.
- Pandas:
pip install pandas
(Refer here for problem installing Pandas) - Matplotlib:
pip install matplotlib
(Refer here for problem installing Matplotlib) - librosa:
pip install librosa
(Refer here for problem installing librosa) - torchaudio:
pip install torchaudio
(Refer here for problem installing torchaudio) - tqdm:
pip install tqdm
(Refer here for problem installing tqdm)
Currently we only support installation from source code using setuptools. Checkout the source code and run the
following commands:
pip install -r requirements.txt
python bin/setup.py build
python bin/setup.py install
you can preprocess KsponSpeech corpus
refer here.
Or refer this page. This documentation contains information regarding the preprocessing of KsponSpeech
.
- Default setting
$ ./run.sh
- Custom setting
python ./bin/main.py --batch_size 32 --num_workers 4 --num_epochs 20 --use_bidirectional \
--input_reverse --spec_augment --noise_augment --use_cuda --hidden_dim 256 \
--dropout 0.3 --num_heads 8 --label_smoothing 0.1 \
--listener_layer_size 5 --speller_layer_size 3 --rnn_type gru \
--high_plateau_lr $HIGH_PLATEAU_LR --teacher_forcing_ratio 1.0 --valid_ratio 0.01 \
--sample_rate 16000 --window_size 20 --stride 10 --n_mels 80 --normalize --del_silence \
--feature_extract_by torchaudio --time_mask_para 70 --freq_mask_para 12 \
--time_mask_num 2 --freq_mask_num 2 --save_result_every 1000 \
--checkpoint_every 5000 --print_every 10 --init_lr 1e-15 --init_uniform \
--mode train --dataset_path /data3/ --data_list_path ./data/data_list/xxx.csv \
--max_grad_norm 400 --rampup_period 1000 --max_len 80 --decay_threshold 0.02 \
--exp_decay_period 160000 --low_plateau_lr 1e-05 --noiseset_size 1000 \
--noise_level 0.7 --attn_mechanism loc --teacher_forcing_step 0.05 \
--min_teacher_forcing_ratio 0.7
You can train the model by above command.
If you want to train by default setting, you can train by Defaulting setting
command.
Or if you want to train by custom setting, you can designate hyperparameters by Custom setting
command.
- Default setting
$ ./eval.sh
- Custom setting
python ./bin/eval.py -dataset_path dataset_path -data_list_path data_list_path \
-mode eval -use_cuda -batch_size 32 -num_workers 4 \
-use_beam_search -k 5 -print_every 100 \
-sample_rate 16000 --window_size 20 --stride 10 --n_mels 80 -feature_extract_by librosa \
-normalize -del_silence -input_reverse
Now you have a model which you can use to predict on new data. We do this by running beam search
(or greedy search
).
Like training, you can choose between Default setting
or Custom setting
.
Checkpoints are organized by experiments and timestamps as shown in the following file structure.
save_dir
+-- checkpoints
| +-- YYYY_mm_dd_HH_MM_SS
| +-- trainer_states.pt
| +-- model.pt
You can resume and load from checkpoints.
We introduce incorporating external language model in performance test.
If you are interested in this content, please check here.
If you have any questions, bug reports, and feature requests, please open an issue on Github.
For live discussions, please go to our gitter or Contacts [email protected] please.
We appreciate any kind of feedback or contribution. Feel free to proceed with small issues like bug fixes, documentation improvement. For major contributions and new features, please discuss with the collaborators in corresponding issues.
We follow PEP-8 for code style. Especially the style of docstrings is important to generate documentation.
Ilya Sutskever et al. Sequence to Sequence Learning with Neural Networks arXiv: 1409.3215
Dzmitry Bahdanau et al. Neural Machine Translation by Jointly Learning to Align and Translate arXiv: 1409.0473
Jan Chorowski et al. Attention Based Models for Speech Recognition arXiv: 1506.07503
Wiliam Chan et al. Listen, Attend and Spell arXiv: 1508.01211
Dario Amodei et al. Deep Speech2: End-to-End Speech Recognition in English and Mandarin arXiv: 1512.02595
Takaaki Hori et al. Advances in Joint CTC-Attention based E2E ASR with a Deep CNN Encoder and RNN-LM arXiv: 1706.02737
Ashish Vaswani et al. Attention Is All You Need arXiv: 1706.03762
Chung-Cheng Chiu et al. State-of-the-art Speech Recognition with Sequence-to-Sequence Models arXiv: 1712.01769
Anjuli Kannan et al. An Analysis Of Incorporating An External LM Into A Seq2seq Model arXiv: 1712.01996
Daniel S. Park et al. SpecAugment: A Simple Data Augmentation Method for ASR arXiv: 1904.08779
Rafael Muller et al. When Does Label Smoothing Help? arXiv: 1906.02629
Jung-Woo Ha et al. ClovaCall: Korean Goal-Oriented Dialog Speech Corpus for ASR of Contact Centers arXiv: 2004.09367
@github{
title = {KoSpeech},
author = {Soohwan Kim, Seyoung Bae, Cheolhwang Won, Suwon Park},
publisher = {GitHub},
docs = {https://sooftware.github.io/KoSpeech/},
url = {https://github.com/sooftware/KoSpeech},
year = {2020}
}