sudips413 / asr-nepali-using-cnn-bilstm-resnet

The research project for designing ML model to convert the Nepali speech to Devnagari text

Automatic Speech Recognition for the Nepali Language using CNN, Bidirectional LSTM and, ResNet

Keywords

Speech To Text, Nepali, CNN, ResNet, BiLSTM, CTC

Intorduction

This repo is a part of the research project for designing the automatic speech recogntion(ASR) model for Nepali language using ML techniques. The conference for this paper has been published in 2022 International Conference on Inventive Computation Technologies (2022 ICICT).

• DOI : 10.1109/ICICT54344.2022.9850832

Things to consider before

• You are free to use this research as a reference and make modifications to continue your own research in Nepali ASR.
• The trainer.py has been implemented to run on the sampled data for now. To replicate the result please replace dataset directory with original OpenSLR dataset.
• Please remove the (audio, text) pairs that include Devnagari numeric texts like १४२३, ५९२, etc from the dataset because they degrade the performance of dataset.

Our approach

0. Remove the (audio, text) pairs that include Devnagari numeric transcriptions
1. Data cleaning (clipping silent gaps from both ends)
2. MFCC feature extraction from audio data
3. Design Neural Network (optimal: CNN + ResNet + BiLSTM) model
4. Calculate CTC loss for applying gradient (training)
5. Decode the texts by using beam search decoding (infernce)

Architecture

Running the project

0. Initialize the virtual environment by installing packages from requirements.txt.
1. Run the training pipeline & evaluate authors model, which can be also be used to evaluate your own (audio,text) pairs.

python trainer.py   # For running the training pipeline
python eval.py      # For testing and evaluating the model already trained by the author

Results

Models and Their character error rate (CER) on Test Data (5% of Total Data.)

Model	Test Data CER	# Params
BiLSTM	19.71%	1.17M
1D-CNN + BiLSTM	24.6%	1.55M
1D-CNN + ResNet + BiGRU	29.6%	1.30M
✓ 1D-CNN + ResNet + BiLSTM	✓ 17.06%	✓ 1.55M
1D-CNN + ResNet + LSTM	30.27%	0.88M

For any queries email the author here.