Bangla speech recognition using 1D CNN and LSTM with different dimension reduction techniques

Abstract

In the area of machine learning, speech recognition was always a hot topic but as world's 8th most widely spoken language Bangla hasn't got the focus as much as she deserved. This research will be on speech recognition using Bangla language dataset. The training model to recognize consists of 1 dimensional Convolutional Neural Network (CNN) and Long-Short Term Memory (LSTM). For feature extraction Mel-frequency Cepstral Coe cient (MFCC) and Mel Spectrogram has been used as the key features for the recognition task. MFCC alone gave an accuracy of 98% for 1d CNN. MFCC when used with LSTM gave an accuracy of 82.35%. Next dimensionality reduction technique was implemented Principal Component Analysis (PCA), Kernel-PCA (k-PCA) and T-distributed Stochastic Neighbor Embedding (t- SNE) transformation on MFCC and Mel Spectrogram for dimensionality reduction technique in a hope to obtain better as e ciency as possible. This is the rst attempt to implement these feature reduction methods on Bengali speech. Dimensionality reduction is a technique that is used to reduce large number of features into fewer factors which holds several advantages like reducing time and required storage space. After transformation using PCA a high consistent accuracy was obtained compared to k-PCA and t-SNE transformation (lowest in t-SNE). With PCA applied on MFCC coe cient the accuracy obtained was 94.54% for 1D CNN and 82.35% for LSTM. With t-SNE the accuracy obtained was 49% with 1D CNN and 50% with LSTM. We have also computed the Mel Spectrogram of the audio data after feeding it to model we obtain an accuracy of 90.74% for 1D CNN and 91.6% for LSTM. With k-PCA applied on Mel Spectrogram coe cient the accuracy obtained was 73.95% for 1D CNN and 72.27% for LSTM.