Robust spoken word detection in assamese language using BiLSTM with data augmentation for noisy environments

Abstract

This study focuses on enhancing spoken word detection in the Assamese language using bidirectional long shor term memory (BiLSTM). The primary objective is to improve the model’s robustness in noisy environments by using various data augmentation methods. The research addresses the challenges of keyword detection in low-resource languages like Assamese. A BiLSTM model was trained and tested using a speech corpus sourced from the Indian Language Technology Proliferation and Development Center (ILTP-DC), comprising 32,335 utterances from 1,000 speakers and 262 unique Assamese words. The model was trained on 10 specific keywords. Feature extraction was conducted using 39 coefficients, including MFCC, ΔMFCC, and ΔΔMFCC. The model’s performance was evaluated on clean and augmented noisy datasets. The application of data augmentation techniques significantly improved the model’s performance in noisy environments. This model achieved an average accuracy of 98.01% and a word error rate (WER) of 19.94% on noisy data, showcasing the effectiveness of augmentation in enhancing keyword detection. This work introduces a novel approach to Assamese spoken word detection by integrating BiLSTM with data augmentation techniques, making the model more noise-resilient. This study sets a benchmark for Assamese speech recognition and showcases augmentation techniques’ effectiveness in low-resource languages.