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
- Install PyTorch>=1.3 with conda
- Install remaining dependencies:
pip install -r requirements.txt - Download the Mongolian Bible dataset:
cd datasets && python dl_mbspeech.py - Pre compute the mel spectrograms:
python preprop_dataset.py --dataset mbspeech - 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
- logs for the TensorBoard are saved in the folder
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
If you are Mongolian and want to help us, please record your voice on Common Voice.