In domain training data augmentation on noise robust Punjabi Children speech recognition

For building a successful automatic speech recognition (ASR) engine large training data is required. It increases training complexity and become impossible for less resource language like Punjabi which have zero children corpus. Consequently, the issue of data scarcity, and small vocal length of children speakers also degrades the system performance under limited data conditions. Unfortunately, Punjabi is a tonal language and building an optimized ASR for such a language is near impossible. In this paper, we have explored fused feature extraction approach to handle large training complexity using mel frequency-gammatone frequency cepstral coefficient (MF-GFCC) technique through feature warping method. The efforts have been made to develop children’s ASR engine using data augmentation on limited data scenarios. For that purpose, we have studied in-domain data augmentation that artificially combined noisy and clean corpus to overcome the issue of data scarcity in train set. The combined dataset is processed with a fused feature extraction approach. Apart, the tonal characteristics and child vocal length issues are also overcome by inducing pitch features and train normalization strategy using vocal tract length normalization (VTLN) approach. In addition to that, combined augmented and original speech signals are noted to reduce the Word error rate (WER) performance with larger relative improvement (RI) of 20.59% on noisy and 19.39% on clean environment conditions using hybrid MF-GFCC approach than that on conventional Mel Frequency Cepstral Coefficient (MFCC) and Gammatone Frequency Cepstral Coefficient (GFCC) based ASR systems.