Mongolian Speech Recognition Save

An online demo trained with a Mongolian proprietary dataset (WER 8%): https://chimege.mn/.

In this repo, following papers are implemented:

• QuartzNet: Deep Automatic Speech Recognition with 1D Time-Channel Separable Convolutions
• An End-to-End Trainable Neural Network for Image-based Sequence Recognition and Its Application to Scene Text Recognition
  • speech recognition as optical character recognition

This repo is partially based on:

• decoder from SeanNaren/deepspeech.pytorch
• Jasper/QuartzNet blocks from NVIDIA/NeMo

Training

1. Install PyTorch>=1.3 with conda
2. Install remaining dependencies: pip install -r requirements.txt
3. Download the Mongolian Bible dataset: cd datasets && python dl_mbspeech.py
4. Pre compute the mel spectrograms: python preprop_dataset.py --dataset mbspeech
5. Train: python train.py --model crnn --max-epochs 50 --dataset mbspeech --lr-warmup-steps 100
   • logs for the TensorBoard are saved in the folder logdir

Results

During the training, the ground truth and recognized texts are logged into the TensorBoard. Because the dataset contains only a single person, the predicted texts from the validation set should be already recognizable after few epochs:

EXPECTED:

аливаа цус хувцсан дээр үсрэхэд цус үсэрсэн хэсгийг та нар ариун газарт угаагтун

PREDICTED:

аливаа цус хувцсан дээр үсэрхэд цус усарсан хэсхийг та нар ариун газарт угаагтун

For fun, you can also generate an audio with a Mongolian TTS and try to recognize it. The following code generates an audio with the TTS of the Mongolian National University and does speech recognition on that generated audio:

# generate audio for 'Миний төрсөн нутаг Монголын сайхан орон'
wget -O test.wav "http://172.104.34.197/nlp-web-demo/tts?voice=1&text=Миний төрсөн нутаг Монголын сайхан орон."
# speech recognition on that TTS generated audio
python transcribe.py --checkpoint=logdir/mbspeech_crnn_sgd_wd1e-05/epoch-0050.pth --model=crnn test.wav
# will output: 'миний төрсөн нут мөнголын сайхан оөрулн'

It is also possible to use a KenLM binary model. First download it from tugstugi/mongolian-nlp. After that, install parlance/ctcdecode. Now you can transcribe with the language model:

python transcribe.py --checkpoint=path/to/checkpoint --lm=mn_5gram.binary --alpha=0.3 test.wav

Contribute

If you are Mongolian and want to help us, please record your voice on Common Voice.