KoSpeech: End-to-End Korean Speech Recognition Toolkit
Soohwan Kim1,2, Seyoung Bae1, Cheolhwang Won1
1Kwangwoon University. 2Kakao Brain.
We present KoSpeech, an open-source software, modular, extensible end-to-end Korean automatic speech recognition (ASR) toolkit based on the deep learning library PyTorch. Several automatic speech recognition open-source toolkits are released, but all of them deal with English speech recognition. (e.g. ESPnet, Espresso) Also AI Hub opened 1,000 hours of Korean speech corpus, KsponSpeech, there is no established preprocessing method and baseline model released to compare model performance, yet. For this reason, we proposes preprocessing methods for KsponSpeech corpus and baseline model to make benchmarks. By KoSpeech, we hope this will be a guideline for those who research Korean speech recognition. Our baseline model which is based on Listen, Attend and Spell (LAS) architecture achieved 10.31% character error rate (CER) only using acoustic model.
| Description | Feature | Dataset | Model | CER |
|---|---|---|---|---|
| seq2seq_vgg_multihead_epoch0 | kaldi_fbank_80 | KsponSpeech | download | 16.6 |
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.
-
Attention (Multi-Head / Location-Aware / Additive / Scaled-dot)
-
Various Feature (Spectrogram / Mel-Spectrogram / MFCC / Filter-Bank)
-
Scheduled Sampling (Teacher forcing scheduling)
-
Inference with batching
-
Multi-GPU training
We have referred to several 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.
Sequence-to-Sequence can be trained with serveral options. You can choose the CNN extractor from (ds2 /vgg),
You can choose attention mechanism from (location-aware, multi-head, additive, scaled-dot) attention.
Our architecture based on Listen Attend and Spell.
We mainly referred to following papers.
Wiliam Chan et al.「Listen, Attend and Spell」 ICASSP 2016
Ashish Vaswani et al 「Attention Is All You Need」 NIPS 2017
Chiu et al 「StateOf-The-Art Speech Recognition with Sequence-to-Sequence Models」 ICASSP 2018
Daniel S. Park et al 「SpecAugment: A Simple Data Augmentation Method for ASR」 Interspeech 2019
Our Seq2seq architeuture is as follows.
Seq2seq(
(encoder): Seq2seqEncoder(
(conv): VGGExtractor(
(conv): 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, 512, num_layers=3, batch_first=True, dropout=0.3, bidirectional=True)
)
(decoder): Seq2seqDecoder(
(embedding): Embedding(2038, 1024)
(input_dropout): Dropout(p=0.3, inplace=False)
(rnn): LSTM(1024, 1024, num_layers=2, batch_first=True, dropout=0.3)
(attention): AddNorm(
(sublayer): MultiHeadAttention(
(query_proj): Linear(in_features=1024, out_features=1024, bias=True)
(key_proj): Linear(in_features=1024, out_features=1024, bias=True)
(value_proj): Linear(in_features=1024, out_features=1024, bias=True)
)
(layer_norm): LayerNorm(1024)
)
(projection): AddNorm(
(sublayer): Linear(in_features=1024, out_features=1024, bias=True)
(layer_norm): LayerNorm(1024)
)
(generator): Linear(in_features=1024, out_features=2038, bias=False)
)
)The Transformer model is currently implemented, but the code for learning is not implemented.
We will implement as soon as possible.
SpeechTransformer(
(conv): Sequential(
(0): Conv2d(1, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
(1): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(2): Hardtanh(min_val=0, max_val=20, inplace=True)
(3): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
(4): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(5): Hardtanh(min_val=0, max_val=20, inplace=True)
(6): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
(7): Conv2d(64, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
(8): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(9): Hardtanh(min_val=0, max_val=20, inplace=True)
(10): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
(11): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(12): Hardtanh(min_val=0, max_val=20, inplace=True)
(13): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
)
(encoder): SpeechTransformerEncoder(
(input_proj): Linear(in_features=2560, out_features=512, bias=True)
(input_norm): LayerNorm()
(input_dropout): Dropout(p=0.3, inplace=False)
(layers): ModuleList(
(N): SpeechTransformerEncoderLayer(
(self_attention): AddNorm(
(sublayer): MultiHeadAttention(
(query_proj): Linear(in_features=512, out_features=512, bias=True)
(key_proj): Linear(in_features=512, out_features=512, bias=True)
(value_proj): Linear(in_features=512, out_features=512, bias=True)
)
(layer_norm): LayerNorm()
)
(feed_forward): AddNorm(
(sublayer): PositionWiseFeedForwardNet(
(feed_forward): Sequential(
(0): Linear(in_features=512, out_features=2048, bias=True)
(1): Dropout(p=0.3, inplace=False)
(2): ReLU()
(3): Linear(in_features=2048, out_features=512, bias=True)
(4): Dropout(p=0.3, inplace=False)
)
)
(layer_norm): LayerNorm()
)
)
)
)
(decoder): SpeechTransformerDecoder(
(embedding): Embedding(2038, 512, padding_idx=0)
(input_dropout): Dropout(p=0.3, inplace=False)
(layers): ModuleList(
(N): SpeechTransformerDecoderLayer(
(self_attention): AddNorm(
(sublayer): MultiHeadAttention(
(query_proj): Linear(in_features=512, out_features=512, bias=True)
(key_proj): Linear(in_features=512, out_features=512, bias=True)
(value_proj): Linear(in_features=512, out_features=512, bias=True)
)
(layer_norm): LayerNorm()
)
(memory_attention): AddNorm(
(sublayer): MultiHeadAttention(
(query_proj): Linear(in_features=512, out_features=512, bias=True)
(key_proj): Linear(in_features=512, out_features=512, bias=True)
(value_proj): Linear(in_features=512, out_features=512, bias=True)
)
(layer_norm): LayerNorm()
)
(feed_forward): AddNorm(
(sublayer): PositionWiseFeedForwardNet(
(feed_forward): Sequential(
(0): Linear(in_features=512, out_features=2048, bias=True)
(1): Dropout(p=0.3, inplace=False)
(2): ReLU()
(3): Linear(in_features=2048, out_features=512, bias=True)
(4): Dropout(p=0.3, inplace=False)
)
)
(layer_norm): LayerNorm()
)
)
)
)
(generator): Linear(
(linear): Linear(in_features=512, out_features=2038, bias=True)
)
)We mainly referred to following papers.
Ashish Vaswani et al 「Attention Is All You Need」 NIPS 2017
You can choose feature extraction method from (spectrogram, mel-spectrogram, mfcc, filter-bank).
In addition to this, You can see a variety of options here.
- Options
usage: main.py [-h] [--mode MODE] [--sample_rate SAMPLE_RATE]
[--frame_length FRAME_LENGTH] [--frame_shift FRAME_SHIFT]
[--n_mels N_MELS] [--normalize] [--del_silence]
[--input_reverse] [--feature_extract_by FEATURE_EXTRACT_BY]
[--transform_method TRANSFORM_METHOD]
[--time_mask_para TIME_MASK_PARA]
[--freq_mask_para FREQ_MASK_PARA]
[--time_mask_num TIME_MASK_NUM] [--freq_mask_num FREQ_MASK_NUM]
[--architecture ARCHITECTURE] [--use_bidirectional]
[--mask_conv] [--hidden_dim HIDDEN_DIM] [--dropout DROPOUT]
[--num_heads NUM_HEADS] [--label_smoothing LABEL_SMOOTHING]
[--num_encoder_layers NUM_ENCODER_LAYERS]
[--num_decoder_layers NUM_DECODER_LAYERS] [--rnn_type RNN_TYPE]
[--extractor EXTRACTOR] [--activation ACTIVATION]
[--attn_mechanism ATTN_MECHANISM]
[--teacher_forcing_ratio TEACHER_FORCING_RATIO]
[--num_classes NUM_CLASSES] [--d_model D_MODEL]
[--ffnet_style FFNET_STYLE] [--dataset_path DATASET_PATH]
[--data_list_path DATA_LIST_PATH] [--label_path LABEL_PATH]
[--spec_augment] [--noise_augment]
[--noiseset_size NOISESET_SIZE] [--noise_level NOISE_LEVEL]
[--use_cuda] [--batch_size BATCH_SIZE]
[--num_workers NUM_WORKERS] [--num_epochs NUM_EPOCHS]
[--init_lr INIT_LR] [--high_plateau_lr HIGH_PLATEAU_LR]
[--low_plateau_lr LOW_PLATEAU_LR] [--valid_ratio VALID_RATIO]
[--max_len MAX_LEN] [--max_grad_norm MAX_GRAD_NORM]
[--rampup_period RAMPUP_PERIOD]
[--decay_threshold DECAY_THRESHOLD]
[--exp_decay_period EXP_DECAY_PERIOD]
[--teacher_forcing_step TEACHER_FORCING_STEP]
[--min_teacher_forcing_ratio MIN_TEACHER_FORCING_RATIO]
[--seed SEED] [--save_result_every SAVE_RESULT_EVERY]
[--checkpoint_every CHECKPOINT_EVERY]
[--print_every PRINT_EVERY] [--resume]
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
- Command
$ ./run_pretrain.sh
- Output
아 뭔 소리야 그건 또
You can get a quick look of pre-trained model's inference, with a sample data.
You can preprocess KsponSpeech Corpus through this repo.
We recommended that you read README of this repository.
- Default setting
$ ./run_seq2seq.sh
- Custom setting
python ./bin/main.py --batch_size 32 --num_workers 4 --num_epochs 20 --spec_augmentYou 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
Now you have a model which you can use to predict on new data. We do this by running greedy search or beam 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.
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 Automatic Speech Recognition 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 Sequence-to-Sequence Model arXiv: 1712.01996
Daniel S. Park et al. SpecAugment: A Simple Data Augmentation Method for Automatic Speech Recognition 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 Automatic Speech Recognition of Contact Centers arXiv: 2004.09367
- IBM/Pytorch-seq2seq
- SeanNaren/deepspeech.pytorch
- kaituoxu/Speech-Transformer
- OpenNMT/OpenNMT-py
- clovaai/ClovaCall
This project is licensed under the Apache-2.0 LICENSE - see the LICENSE.md file for details
- Soohwan Kim, Seyoung Bae, Cheolhwang Won
- Contacts: [email protected]