VietASR (NVIDIA NeMo ToolKit)

⚡ 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.

Task	CER (%)	WER (%)	+LM WER (%)
VIVOS (TEST)	6.80	18.02	15.72
VLSP2018	6.87	16.26	N/A
VLSP2020 T1	14.73	30.96	N/A
VLSP2020 T2	41.67	69.15	N/A

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,...

Run Demo

• Vietnamese Model (pretrained): python flask_upload_record_vn.py

• Video demo in Youtube: https://youtu.be/P3mhEngL1us
  

• 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