Simple Khmer ASR Project

This repository contains a simple Khmer Automatic Speech Recognition (ASR) project from scratch. Feel free to fork this repository, submit pull requests, or send us suggestions on what should be improved! I just do it for fun and celebrating my bd :D

1. Data Collection and Preprocessing

Data Collection

• YouTube Videos

  • We crawl YouTube videos using yt_dlp and ffmpeg.
  • Instructions: Just follow the links to download the tools and refer to the given channel names for crawling.
  • yt_dlp: Download yt_dlp
  • ffmpeg: Download ffmpeg

• OpenSLR Dataset

  • Alternatively, you can use the OpenSLR dataset, which is open-source.
  • OpenSLR Dataset Link

Data Cleaning

• Background Noise Removal

  • We use Ultimate Vocal Remover for background noise removal.
  • Ultimate Vocal Remover
  • Separate code and model's files are provided in the folder Background_Noise.

• Chunking

  • Automated chunking is performed using Python. For non-stratified results, manual checking with Audacity is recommended.
  • Download Audacity

2. Transcription

• Transcribe_New.py
  • The script outputs three folders: 1_word, UNK (unknown), and non-transcript.
  • Manual checking is recommended for perfect transcription accuracy.

3. Data Training

Wav2Vec2

• About Wav2Vec2
  • Wav2Vec2 is a state-of-the-art model developed by Facebook AI (now Meta AI) for ASR. It converts raw audio waveforms into meaningful text.
  • 
  • The model was trained using connectionist temporal classification (CTC), so the output has to be decoded using Wav2Vec2CTCTokenizer.

Metrics

• WER (Word Error Rate)
  • WER is a metric used to evaluate the quality of transcriptions produced by ASR systems.
  • 
  • In many applications, it is of interest to estimate WER given a pair of a speech utterance and a transcript.

4. Simple StreamLit Application

• This application records voice or inputs a file and returns the transcript text!

References

• LJ Speech Dataset
• Wav2Vec2 on Hugging Face
• OpenSLR Resource
• Wav2Vec2 Khmer Model
• Separate Model

Problems

• If you face any problems about the model weight, this might help :
• Seanghay Yath
• Vituo Phy 1
• Vituo Phy 2