VietASR (Vietnamese Automatic Speech Recognition)

⚡ Some experiment with NeMo ⚡ Result

• Model: QuartzNet is a smaller version of Jaser model
• I list the word error rate (WER) with and without LM of major ASR tasks.

Model was trained with ~500 hours Vietnamese speech dataset, was collected from youtube, radio, call center(8k), text to speech data and some public dataset (vlsp, vivos, fpt). It is very small model (13M parameters) make it inference so fast ⚡

Installation

• ctcdecoder, kemlm for LM Decode
  pip install ds-ctcdecoder
• and some python libraries: torch, numpy, librosa, flask, flask_socketio, requests,...

Demo

• Vietnamese Model (pretrained): python flask_upload_record_vn.py
• Video demo in Youtube:
  • v1: https://youtu.be/P3mhEngL1us
  • v2: https://youtu.be/o9NpWi3VUHs

• English Model (pretrained): python flask_upload_record_en.py

TODO

• Conformer Model
• Transformer LM instead of kenlm
• Data augumentation: speed, noise, pitch shift, time shift,...
• FastAPI

Citation

  @article{kuchaiev2019nemo,
    title={Nemo: a toolkit for building ai applications using neural modules},
    author={Kuchaiev, Oleksii and Li, Jason and Nguyen, Huyen and Hrinchuk, Oleksii and Leary, Ryan and Ginsburg, Boris and Kriman, Samuel and Beliaev, Stanislav and Lavrukhin, Vitaly and Cook, Jack and others},
    journal={arXiv preprint arXiv:1909.09577},
    year={2019}
  }