Development of Robust Automatic Speech Recognition System for Children's using Kaldi Toolkit
Vivek Bhardwaj, Sashi Bala, Virender Kadyan, Vinay Kukreja
Abstract
In this paper, the Punjabi children speech recognition system is developed using Subspace Gaussian mixture models (SGMM) acoustic modeling techniques. Initially, the system is dependent upon Mel-frequency cepstral coefficients (MFCC) approach for controlling the temporal variations in the input speech signals. Here, SGMM is integrated with HMM to measure the efficiency of each state which carries the information of a short-windowed frame. For handling the children speaker acoustic variations speaker adaptive training (SAT), based on vocal-tract length normalization and feature space maximum likelihood linear regression is adopted. Kaldi and open-source speech recognition toolkit is used to develop the Robust Automatic Speech Recognition (ASR) System for Punjabi Children's speech. S GMM accumulate the frame coefficients and their posterior probabilities and pass these probabilities to HMM which systematically fit the frame and output have resulted from HMM states. Therefore, the achievement of SGMM has gotten a large performance margin in Punjabi children speech recognition. A remarkable depletion in the word error rate (WER) was noticed using SGMM by varying the feature dimensions. The developed children ASR system obtained a recognition accuracy of 83.66% while tested by varying the feature dimensions to 12.