The Dialpad ASR System for the Multilingual ASR Challenge for Low Resource Indian Languages 2021

Shreekantha Nadig, Riqiang Wang, Wang Yau Li, Jeffrey Michael, Frédéric Mailhot, Simon Vandieken, Jonas Robertson

Dialpad, Inc.

This paper describes the multilingual ASR systems developed at Dialpad, Inc. for the Multilingual ASR challenge for low resource Indian languages at Interspeech 2021. We participated in Sub-task 1, where the systems are trained on data of six Indic languages provided by the organizers. On this task, we experimented with both hybrid HMM-DNN and end-to-end ASR architectures and studied how fine-tuning techniques can help in this multilingual scenario. We also experimented with both multilingual and language-specific decoders by using a pre-trained encoder, as well as the use of appropriate RNN and n-gram language models. Furthermore, we present novel studies on transliteration-based pre-training of the encoder, and a joint LID and ASR architecture. We show that the multilingual end-to-end ASR models outperform both hybrid model and monolingual baselines. Also, we demonstrate that current methods of joint LID-ASR fail when there are confounding channel characteristics. We conducted studies and propose ideas on how to mitigate the effect of some of the channel characteristics on the task of language recognition. Our best submission to the challenge achieved an average WER of $22.95%$ on the development set and $31.87%$ on the held-out test set and contains language-specific decoders fine-tuned on the multilingual encoder, along with the use of language-specific RNNLMs and n-gram LMs.

Video Presentation

Models

All the end-to-end models in this work are trained using the ESPnet toolkit. Hence, the inference also follows the standard format of the toolkit. The features are extracted using torchaudio (as opposed to kaldi binaries) in the toolkit. We provide the feature extraction code as well. All of the models in this work can be used with the standard ESPnet decoding scripts as mentioned in the ESPnet toolkit: https://github.com/espnet/interspeech2019-tutorial

We make available the following pre-trained models for this work:

Name	Description
B0	Baseline encoder-decoder with combined vocabulary
B1	B0's encoder + monolingual decoder (Encoder frozen from B0)
B1 (unfreeze)	B0's encoder + monolingual decoder (Fine-tune after un-freezing Encoder)
B3	B0 but with transliterated latin script
C0	B0 + explicit LID subtask
C1	B3's encoder + explicit LID decoder
L0	LID trained from scratch
L1	LID with transliterated Encoder from B3
"lang"_RNNLM	Byte-level RNNLM for each language

You can find the pre-trained models in this Google Drive link: https://drive.google.com/drive/folders/1QlEZgzscznfPaVv_B62Ipz0grXdeDNIr?usp=sharing

Our data preparation recipe and inference scripts are under egs/mucs_2021/task1/

Extracting features for inference

For all experiments, we extracted 80-dimensional log Mel filterbank features with a window size of 25 ms computed at every 10 ms. The features are extracted using torchaudio.compliance.kaldi.fbank

lmspc = torchaudio.compliance.kaldi.fbank(
            waveform=torch.unsqueeze(torch.tensor(signal), axis=0),
            sample_frequency=8000,
            dither=1e-32,
            energy_floor=0,
            num_mel_bins=80,
        )

Performing inference with the pre-trained models

We give an example ipython notebook (inference_example.ipynb) to perform inference with various models with features extracted using torchaudio and with an appropriate RNNLM.

# ESPnet: end-to-end speech processing toolkit

system/pytorch ver.	1.3.1	1.4.0	1.5.1	1.6.0	1.7.1	1.8.1	1.9.0
ubuntu20/python3.8/pip
ubuntu18/python3.7/pip
debian9/python3.6/conda
centos7/python3.6/conda
doc/python3.8

ESPnet is an end-to-end speech processing toolkit, mainly focuses on end-to-end speech recognition and end-to-end text-to-speech. 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

Kaldi style complete recipe

Support numbers of ASR recipes (WSJ, Switchboard, CHiME-4/5, Librispeech, TED, CSJ, AMI, HKUST, Voxforge, REVERB, etc.)
Support numbers of TTS recipes with a similar manner to the ASR recipe (LJSpeech, LibriTTS, M-AILABS, etc.)
Support numbers of ST recipes (Fisher-CallHome Spanish, Libri-trans, IWSLT'18, How2, Must-C, Mboshi-French, etc.)
Support numbers of MT recipes (IWSLT'16, the above ST recipes etc.)
Support speech separation and recognition recipe (WSJ-2mix)
Support voice conversion recipe (VCC2020 baseline) (new!)

ASR: Automatic Speech Recognition

State-of-the-art performance in several ASR benchmarks (comparable/superior to hybrid DNN/HMM and CTC)
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 + BiRNN (LSTM/GRU), sub-sampling BiRNN (LSTM/GRU) or Transformer
Attention: Dot product, location-aware attention, variants of multihead
Incorporate RNNLM/LSTMLM/TransformerLM/N-gram trained only with text data
Batch GPU decoding
Transducer based end-to-end ASR
- Available: RNN-based encoder/decoder or custom encoder/decoder w/ supports for Transformer, Conformer, TDNN (encoder) and causal conv1d (decoder) blocks.
- Also support: mixed RNN/Custom encoder-decoder, VGG2L (RNN/Cutom encoder) and various decoding algorithms.
Please refer to the tutorial page for complete documentation.
CTC segmentation
Non-autoregressive model based on Mask-CTC
ASR examples for supporting endangered language documentation (Please refer to egs/puebla_nahuatl and egs/yoloxochitl_mixtec for details)
Wav2Vec2.0 pretrained model as Encoder, imported from FairSeq.

Demonstration

Real-time ASR demo with ESPnet2

TTS: Text-to-speech

Tacotron2
Transformer-TTS
FastSpeech
FastSpeech2 (in ESPnet2)
Conformer-based FastSpeech & FastSpeech2 (in ESPnet2)
Multi-speaker model with pretrained speaker embedding
Multi-speaker model with GST (in ESPnet2)
Phoneme-based training (En, Jp, and Zn)
Integration with neural vocoders (WaveNet, ParallelWaveGAN, and MelGAN)

Demonstration

Real-time TTS demo with ESPnet2
Real-time TTS demo with ESPnet1

To train the neural vocoder, please check the following repositories:

NOTE:

We are moving on ESPnet2-based development for TTS.

If you are beginner, we recommend using ESPnet2-TTS.

SE: Speech enhancement (and separation)

Single-speaker speech enhancement
Multi-speaker speech separation
Unified encoder-separator-decoder structure for time-domain and frequency-domian models
- Encoder/Decoder: STFT/iSTFT, Convolution/Transposed-Convolution
- Separators: BLSTM, Transformer, Conformer, DPRNN, Neural Beamformers, etc.
Flexible ASR integration: working as an individual task or as the ASR frontend
Easy to import pretrained models from Asteroid
- Both the pre-trained models from Asteroid and the specific configuration are supported.

Demonstration

Interactive SE demo with ESPnet2

ST: Speech Translation & MT: Machine Translation

State-of-the-art performance in several ST benchmarks (comparable/superior to cascaded ASR and MT)
Transformer based end-to-end ST (new!)
Transformer based end-to-end MT (new!)

VC: Voice conversion

Transformer and Tacotron2 based parallel VC using melspectrogram (new!)
End-to-end VC based on cascaded ASR+TTS (Baseline system for Voice Conversion Challenge 2020!)

DNN Framework

Flexible network architecture thanks to chainer and pytorch
Flexible front-end processing thanks to kaldiio and HDF5 support
Tensorboard based monitoring

ESPnet2

See ESPnet2.

Indepedent from Kaldi/Chainer, unlike ESPnet1
On the fly feature extraction and text processing when training
Supporting DistributedDataParallel and DaraParallel both
Supporting multiple nodes training and integrated with Slurm or MPI
Supporting Sharded Training provided by fairscale
A template recipe which can be applied for all corpora
Possible to train any size of corpus without cpu memory error
ESPnet Model Zoo
Integrated with wandb

Installation

If you intend to do full experiments including DNN training, then see Installation.

If you just need the Python module only:

pip install espnet
# To install latest
# pip install git+https://github.com/espnet/espnet

You need to install some packages.

pip install torch
pip install chainer==6.0.0 cupy==6.0.0    # [Option] If you'll use ESPnet1
pip install torchaudio                    # [Option] If you'll use enhancement task
pip install torch_optimizer               # [Option] If you'll use additional optimizers in ESPnet2

There are some required packages depending on each task other than above. If you meet ImportError, please intall them at that time.

(ESPNet2) Once installed, run wandb login and set --use_wandb true to enable tracking runs using W&B.

Usage

See Usage.

Docker Container

go to docker/ and follow instructions.

Contribution

Thank you for taking times for ESPnet! Any contributions to ESPNet are welcome and feel free to ask any questions or requests to issues. If it's the first contribution to ESPnet for you, please follow the contribution guide.

Results and demo

You can find useful tutorials and demos in Interspeech 2019 Tutorial

ASR results

expand

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

Task	CER (%)	WER (%)	Pretrained model
Aishell dev/test	4.6/5.1	N/A	link
ESPnet2 Aishell dev/test	4.4/4.7	N/A	link
Common Voice dev/test	1.7/1.8	2.2/2.3	link
CSJ eval1/eval2/eval3	5.7/3.8/4.2	N/A	link
ESPnet2 CSJ eval1/eval2/eval3	4.5/3.3/3.6	N/A	link
HKUST dev	23.5	N/A	link
ESPnet2 HKUST dev	21.2	N/A	link
Librispeech dev_clean/dev_other/test_clean/test_other	N/A	1.9/4.9/2.1/4.9	link
ESPnet2 Librispeech dev_clean/dev_other/test_clean/test_other	0.7/2.2/0.7/2.1	1.9/4.6/2.1/4.7	link
Switchboard (eval2000) callhm/swbd	N/A	14.0/6.8	link
TEDLIUM2 dev/test	N/A	8.6/7.2	link
TEDLIUM3 dev/test	N/A	9.6/7.6	link
WSJ dev93/eval92	3.2/2.1	7.0/4.7	N/A
ESPnet2 WSJ dev93/eval92	2.7/1.8	6.6/4.6	link

Note that the performance of the CSJ, HKUST, and Librispeech tasks was significantly improved by using the wide network (#units = 1024) and large subword units if necessary reported by RWTH.

If you want to check the results of the other recipes, please check egs/<name_of_recipe>/asr1/RESULTS.md.

ASR demo

expand

You can recognize speech in a WAV file using pretrained models. Go to a recipe directory and run utils/recog_wav.sh as follows:

# go to recipe directory and source path of espnet tools
cd egs/tedlium2/asr1 && . ./path.sh
# let's recognize speech!
recog_wav.sh --models tedlium2.transformer.v1 example.wav

where example.wav is a WAV file to be recognized. The sampling rate must be consistent with that of data used in training.

Available pretrained models in the demo script are listed as below.

Model	Notes
tedlium2.rnn.v1	Streaming decoding based on CTC-based VAD
tedlium2.rnn.v2	Streaming decoding based on CTC-based VAD (batch decoding)
tedlium2.transformer.v1	Joint-CTC attention Transformer trained on Tedlium 2
tedlium3.transformer.v1	Joint-CTC attention Transformer trained on Tedlium 3
librispeech.transformer.v1	Joint-CTC attention Transformer trained on Librispeech
commonvoice.transformer.v1	Joint-CTC attention Transformer trained on CommonVoice
csj.transformer.v1	Joint-CTC attention Transformer trained on CSJ
csj.rnn.v1	Joint-CTC attention VGGBLSTM trained on CSJ

SE results

expand

We list results from three different models on WSJ0-2mix, which is one the most widely used benchmark dateset for speech separation.

Model	STOI	SAR	SDR	SIR
TF Masking	0.89	11.40	10.24	18.04
Conv-Tasnet	0.95	16.62	15.94	25.90
DPRNN-Tasnet	0.96	18.82	18.29	28.92

SE demos

expand

You can try the interactive demo with Google Colab. Please click the following button to get access to the demos.

It is based on ESPnet2. Pretrained models are available for both speech enhancement and speech separation tasks.

ST results

expand

We list 4-gram BLEU of major ST tasks.

end-to-end system

Task	BLEU	Pretrained model
Fisher-CallHome Spanish fisher_test (Es->En)	51.03	link
Fisher-CallHome Spanish callhome_evltest (Es->En)	20.44	link
Libri-trans test (En->Fr)	16.70	link
How2 dev5 (En->Pt)	45.68	link
Must-C tst-COMMON (En->De)	22.91	link
Mboshi-French dev (Fr->Mboshi)	6.18	N/A

cascaded system

Task	BLEU	Pretrained model
Fisher-CallHome Spanish fisher_test (Es->En)	42.16	N/A
Fisher-CallHome Spanish callhome_evltest (Es->En)	19.82	N/A
Libri-trans test (En->Fr)	16.96	N/A
How2 dev5 (En->Pt)	44.90	N/A
Must-C tst-COMMON (En->De)	23.65	N/A

If you want to check the results of the other recipes, please check egs/<name_of_recipe>/st1/RESULTS.md.

ST demo

expand

(New!) We made a new real-time E2E-ST + TTS demonstration in Google Colab. Please access the notebook from the following button and enjoy the real-time speech-to-speech translation!

You can translate speech in a WAV file using pretrained models. Go to a recipe directory and run utils/translate_wav.sh as follows:

# go to recipe directory and source path of espnet tools
cd egs/fisher_callhome_spanish/st1 && . ./path.sh
# download example wav file
wget -O - https://github.com/espnet/espnet/files/4100928/test.wav.tar.gz | tar zxvf -
# let's translate speech!
translate_wav.sh --models fisher_callhome_spanish.transformer.v1.es-en test.wav

where test.wav is a WAV file to be translated. The sampling rate must be consistent with that of data used in training.

Available pretrained models in the demo script are listed as below.

Model	Notes
fisher_callhome_spanish.transformer.v1	Transformer-ST trained on Fisher-CallHome Spanish Es->En

MT results

expand

Task	BLEU	Pretrained model
Fisher-CallHome Spanish fisher_test (Es->En)	61.45	link
Fisher-CallHome Spanish callhome_evltest (Es->En)	29.86	link
Libri-trans test (En->Fr)	18.09	link
How2 dev5 (En->Pt)	58.61	link
Must-C tst-COMMON (En->De)	27.63	link
IWSLT'14 test2014 (En->De)	24.70	link
IWSLT'14 test2014 (De->En)	29.22	link
IWSLT'16 test2014 (En->De)	24.05	link
IWSLT'16 test2014 (De->En)	29.13	link

TTS results

ESPnet2

You can listen to the generated samples in the following url.

ESPnet2 TTS generated samples

Note that in the generation we use Griffin-Lim (wav/) and Parallel WaveGAN (wav_pwg/).

You can download pretrained models via espnet_model_zoo.

You can download pretrained vocoders via kan-bayashi/ParallelWaveGAN.

ESPnet1

NOTE: We are moving on ESPnet2-based development for TTS. Please check the latest results in the above ESPnet2 results.

You can listen to our samples in demo HP espnet-tts-sample. Here we list some notable ones:

You can download all of the pretrained models and generated samples:

Note that in the generated samples we use the following vocoders: Griffin-Lim (GL), WaveNet vocoder (WaveNet), Parallel WaveGAN (ParallelWaveGAN), and MelGAN (MelGAN). The neural vocoders are based on following repositories.

kan-bayashi/ParallelWaveGAN: Parallel WaveGAN / MelGAN / Multi-band MelGAN
r9y9/wavenet_vocoder: 16 bit mixture of Logistics WaveNet vocoder
kan-bayashi/PytorchWaveNetVocoder: 8 bit Softmax WaveNet Vocoder with the noise shaping

If you want to build your own neural vocoder, please check the above repositories. kan-bayashi/ParallelWaveGAN provides the manual about how to decode ESPnet-TTS model's features with neural vocoders. Please check it.

Here we list all of the pretrained neural vocoders. Please download and enjoy the generation of high quality speech!

Model link	Lang	Fs [Hz]	Mel range [Hz]	FFT / Shift / Win [pt]	Model type
ljspeech.wavenet.softmax.ns.v1	EN	22.05k	None	1024 / 256 / None	Softmax WaveNet
ljspeech.wavenet.mol.v1	EN	22.05k	None	1024 / 256 / None	MoL WaveNet
ljspeech.parallel_wavegan.v1	EN	22.05k	None	1024 / 256 / None	Parallel WaveGAN
ljspeech.wavenet.mol.v2	EN	22.05k	80-7600	1024 / 256 / None	MoL WaveNet
ljspeech.parallel_wavegan.v2	EN	22.05k	80-7600	1024 / 256 / None	Parallel WaveGAN
ljspeech.melgan.v1	EN	22.05k	80-7600	1024 / 256 / None	MelGAN
ljspeech.melgan.v3	EN	22.05k	80-7600	1024 / 256 / None	MelGAN
libritts.wavenet.mol.v1	EN	24k	None	1024 / 256 / None	MoL WaveNet
jsut.wavenet.mol.v1	JP	24k	80-7600	2048 / 300 / 1200	MoL WaveNet
jsut.parallel_wavegan.v1	JP	24k	80-7600	2048 / 300 / 1200	Parallel WaveGAN
csmsc.wavenet.mol.v1	ZH	24k	80-7600	2048 / 300 / 1200	MoL WaveNet
csmsc.parallel_wavegan.v1	ZH	24k	80-7600	2048 / 300 / 1200	Parallel WaveGAN

If you want to use the above pretrained vocoders, please exactly match the feature setting with them.

TTS demo

ESPnet2

You can try the real-time demo in Google Colab. Please access the notebook from the following button and enjoy the real-time synthesis!

Real-time TTS demo with ESPnet2

English, Japanese, and Mandarin models are available in the demo.

ESPnet1

NOTE: We are moving on ESPnet2-based development for TTS. Please check the latest demo in the above ESPnet2 demo.

You can try the real-time demo in Google Colab. Please access the notebook from the following button and enjoy the real-time synthesis.

Real-time TTS demo with ESPnet1

We also provide shell script to perform synthesize. Go to a recipe directory and run utils/synth_wav.sh as follows:

# go to recipe directory and source path of espnet tools
cd egs/ljspeech/tts1 && . ./path.sh
# we use upper-case char sequence for the default model.
echo "THIS IS A DEMONSTRATION OF TEXT TO SPEECH." > example.txt
# let's synthesize speech!
synth_wav.sh example.txt

# also you can use multiple sentences
echo "THIS IS A DEMONSTRATION OF TEXT TO SPEECH." > example_multi.txt
echo "TEXT TO SPEECH IS A TECHQNIQUE TO CONVERT TEXT INTO SPEECH." >> example_multi.txt
synth_wav.sh example_multi.txt

You can change the pretrained model as follows:

synth_wav.sh --models ljspeech.fastspeech.v1 example.txt

Waveform synthesis is performed with Griffin-Lim algorithm and neural vocoders (WaveNet and ParallelWaveGAN). You can change the pretrained vocoder model as follows:

synth_wav.sh --vocoder_models ljspeech.wavenet.mol.v1 example.txt

WaveNet vocoder provides very high quality speech but it takes time to generate.

See more details or available models via --help.

synth_wav.sh --help

VC results

expand

Transformer and Tacotron2 based VC

You can listen to some samples on the demo webpage.

Cascade ASR+TTS as one of the baseline systems of VCC2020

The Voice Conversion Challenge 2020 (VCC2020) adopts ESPnet to build an end-to-end based baseline system. In VCC2020, the objective is intra/cross lingual nonparallel VC. You can download converted samples of the cascade ASR+TTS baseline system here.

CTC Segmentation demo

ESPnet1

CTC segmentation determines utterance segments within audio files. Aligned utterance segments constitute the labels of speech datasets.

As demo, we align start and end of utterances within the audio file ctc_align_test.wav, using the example script utils/ctc_align_wav.sh. For preparation, set up a data directory:

cd egs/tedlium2/align1/
# data directory
align_dir=data/demo
mkdir -p ${align_dir}
# wav file
base=ctc_align_test
wav=../../../test_utils/${base}.wav
# recipe files
echo "batchsize: 0" > ${align_dir}/align.yaml

cat << EOF > ${align_dir}/utt_text
${base} THE SALE OF THE HOTELS
${base} IS PART OF HOLIDAY'S STRATEGY
${base} TO SELL OFF ASSETS
${base} AND CONCENTRATE
${base} ON PROPERTY MANAGEMENT
EOF

Here, utt_text is the file containing the list of utterances. Choose a pre-trained ASR model that includes a CTC layer to find utterance segments:

# pre-trained ASR model
model=wsj.transformer_small.v1
mkdir ./conf && cp ../../wsj/asr1/conf/no_preprocess.yaml ./conf

../../../utils/asr_align_wav.sh \
    --models ${model} \
    --align_dir ${align_dir} \
    --align_config ${align_dir}/align.yaml \
    ${wav} ${align_dir}/utt_text

Segments are written to aligned_segments as a list of file/utterance name, utterance start and end times in seconds and a confidence score. The confidence score is a probability in log space that indicates how good the utterance was aligned. If needed, remove bad utterances:

min_confidence_score=-5
awk -v ms=${min_confidence_score} '{ if ($5 > ms) {print} }' ${align_dir}/aligned_segments

The demo script utils/ctc_align_wav.sh uses an already pretrained ASR model (see list above for more models). It is recommended to use models with RNN-based encoders (such as BLSTMP) for aligning large audio files; rather than using Transformer models that have a high memory consumption on longer audio data. The sample rate of the audio must be consistent with that of the data used in training; adjust with sox if needed. A full example recipe is in egs/tedlium2/align1/.

ESPnet2

CTC segmentation determines utterance segments within audio files. Aligned utterance segments constitute the labels of speech datasets.

As demo, we align start and end of utterances within the audio file ctc_align_test.wav. This can be done either directly from the Python command line or using the script espnet2/bin/asr_align.py.

From the Python command line interface:

# load a model with character tokens
from espnet_model_zoo.downloader import ModelDownloader
d = ModelDownloader(cachedir="./modelcache")
wsjmodel = d.download_and_unpack("kamo-naoyuki/wsj")
# load the example file included in the ESPnet repository
import soundfile
speech, rate = soundfile.read("./test_utils/ctc_align_test.wav")
# CTC segmentation
from espnet2.bin.asr_align import CTCSegmentation
aligner = CTCSegmentation( **wsjmodel , fs=rate )
text = """
utt1 THE SALE OF THE HOTELS
utt2 IS PART OF HOLIDAY'S STRATEGY
utt3 TO SELL OFF ASSETS
utt4 AND CONCENTRATE ON PROPERTY MANAGEMENT
"""
segments = aligner(speech, text)
print(segments)
# utt1 utt 0.26 1.73 -0.0154 THE SALE OF THE HOTELS
# utt2 utt 1.73 3.19 -0.7674 IS PART OF HOLIDAY'S STRATEGY
# utt3 utt 3.19 4.20 -0.7433 TO SELL OFF ASSETS
# utt4 utt 4.20 6.10 -0.4899 AND CONCENTRATE ON PROPERTY MANAGEMENT

Aligning also works with fragments of the text. For this, set the gratis_blank option that allows skipping unrelated audio sections without penalty. It's also possible to omit the utterance names at the beginning of each line, by setting kaldi_style_text to False.

aligner.set_config( gratis_blank=True, kaldi_style_text=False )
text = ["SALE OF THE HOTELS", "PROPERTY MANAGEMENT"]
segments = aligner(speech, text)
print(segments)
# utt_0000 utt 0.37 1.72 -2.0651 SALE OF THE HOTELS
# utt_0001 utt 4.70 6.10 -5.0566 PROPERTY MANAGEMENT

The script espnet2/bin/asr_align.py uses a similar interface. To align utterances:

# ASR model and config files from pretrained model (e.g. from cachedir):
asr_config=<path-to-model>/config.yaml
asr_model=<path-to-model>/valid.*best.pth
# prepare the text file
wav="test_utils/ctc_align_test.wav"
text="test_utils/ctc_align_text.txt"
cat << EOF > ${text}
utt1 THE SALE OF THE HOTELS
utt2 IS PART OF HOLIDAY'S STRATEGY
utt3 TO SELL OFF ASSETS
utt4 AND CONCENTRATE
utt5 ON PROPERTY MANAGEMENT
EOF
# obtain alignments:
python espnet2/bin/asr_align.py --asr_train_config ${asr_config} --asr_model_file ${asr_model} --audio ${wav} --text ${text}
# utt1 ctc_align_test 0.26 1.73 -0.0154 THE SALE OF THE HOTELS
# utt2 ctc_align_test 1.73 3.19 -0.7674 IS PART OF HOLIDAY'S STRATEGY
# utt3 ctc_align_test 3.19 4.20 -0.7433 TO SELL OFF ASSETS
# utt4 ctc_align_test 4.20 4.97 -0.6017 AND CONCENTRATE
# utt5 ctc_align_test 4.97 6.10 -0.3477 ON PROPERTY MANAGEMENT

The output of the script can be redirected to a segments file by adding the argument --output segments. Each line contains file/utterance name, utterance start and end times in seconds and a confidence score; optionally also the utterance text. The confidence score is a probability in log space that indicates how good the utterance was aligned. If needed, remove bad utterances:

min_confidence_score=-7
# here, we assume that the output was written to the file `segments`
awk -v ms=${min_confidence_score} '{ if ($5 > ms) {print} }' segments

See the module documentation for more information. It is recommended to use models with RNN-based encoders (such as BLSTMP) for aligning large audio files; rather than using Transformer models that have a high memory consumption on longer audio data. The sample rate of the audio must be consistent with that of the data used in training; adjust with sox if needed.

References

[1] Shinji Watanabe, Takaaki Hori, Shigeki Karita, Tomoki Hayashi, Jiro Nishitoba, Yuya Unno, Nelson Enrique Yalta Soplin, Jahn Heymann, Matthew Wiesner, Nanxin Chen, Adithya Renduchintala, and Tsubasa Ochiai, "ESPnet: End-to-End Speech Processing Toolkit," Proc. Interspeech'18, pp. 2207-2211 (2018)

[2] 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)

[3] 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

Citations

@inproceedings{watanabe2018espnet,
  author={Shinji Watanabe and Takaaki Hori and Shigeki Karita and Tomoki Hayashi and Jiro Nishitoba and Yuya Unno and Nelson {Enrique Yalta Soplin} and Jahn Heymann and Matthew Wiesner and Nanxin Chen and Adithya Renduchintala and Tsubasa Ochiai},
  title={{ESPnet}: End-to-End Speech Processing Toolkit},
  year={2018},
  booktitle={Proceedings of Interspeech},
  pages={2207--2211},
  doi={10.21437/Interspeech.2018-1456},
  url={http://dx.doi.org/10.21437/Interspeech.2018-1456}
}
@inproceedings{hayashi2020espnet,
  title={{Espnet-TTS}: Unified, reproducible, and integratable open source end-to-end text-to-speech toolkit},
  author={Hayashi, Tomoki and Yamamoto, Ryuichi and Inoue, Katsuki and Yoshimura, Takenori and Watanabe, Shinji and Toda, Tomoki and Takeda, Kazuya and Zhang, Yu and Tan, Xu},
  booktitle={Proceedings of IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)},
  pages={7654--7658},
  year={2020},
  organization={IEEE}
}
@inproceedings{inaguma-etal-2020-espnet,
    title = "{ESP}net-{ST}: All-in-One Speech Translation Toolkit",
    author = "Inaguma, Hirofumi  and
      Kiyono, Shun  and
      Duh, Kevin  and
      Karita, Shigeki  and
      Yalta, Nelson  and
      Hayashi, Tomoki  and
      Watanabe, Shinji",
    booktitle = "Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics: System Demonstrations",
    month = jul,
    year = "2020",
    address = "Online",
    publisher = "Association for Computational Linguistics",
    url = "https://www.aclweb.org/anthology/2020.acl-demos.34",
    pages = "302--311",
}
@inproceedings{li2020espnet,
  title={{ESPnet-SE}: End-to-End Speech Enhancement and Separation Toolkit Designed for {ASR} Integration},
  author={Chenda Li and Jing Shi and Wangyou Zhang and Aswin Shanmugam Subramanian and Xuankai Chang and Naoyuki Kamo and Moto Hira and Tomoki Hayashi and Christoph Boeddeker and Zhuo Chen and Shinji Watanabe},
  booktitle={Proceedings of IEEE Spoken Language Technology Workshop (SLT)},
  pages={785--792},
  year={2021},
  organization={IEEE},
}

Name		Name	Last commit message	Last commit date
Latest commit History 11,616 Commits
.github		.github
ci		ci
doc		doc
docker		docker
egs		egs
egs2		egs2
espnet		espnet
espnet2		espnet2
test		test
test_utils		test_utils
tools		tools
utils		utils
.coveragerc		.coveragerc
.dockerignore		.dockerignore
.gitignore		.gitignore
.gitmodules		.gitmodules
.mergify.yml		.mergify.yml
CONTRIBUTING.md		CONTRIBUTING.md
LICENSE		LICENSE
README.md		README.md
setup.cfg		setup.cfg
setup.py		setup.py

License

dialpad/mucs_2021_dialpad

Folders and files

Latest commit

History

Repository files navigation

The Dialpad ASR System for the Multilingual ASR Challenge for Low Resource Indian Languages 2021

Shreekantha Nadig, Riqiang Wang, Wang Yau Li, Jeffrey Michael, Frédéric Mailhot, Simon Vandieken, Jonas Robertson

Dialpad, Inc.

Video Presentation

Models

Extracting features for inference

Performing inference with the pre-trained models

Key Features

Kaldi style complete recipe

ASR: Automatic Speech Recognition

TTS: Text-to-speech

SE: Speech enhancement (and separation)

ST: Speech Translation & MT: Machine Translation

VC: Voice conversion

DNN Framework

ESPnet2

Installation

Usage

Docker Container

Contribution

Results and demo

ASR results

ASR demo

SE results

SE demos

ST results

end-to-end system

cascaded system

ST demo

MT results

TTS results

TTS demo

VC results

CTC Segmentation demo

References

Citations

About

Topics

Resources

License

Stars

Watchers

Forks

Contributors 128

Languages