Parallel convolutional neural network and hybrid architectures for accented speech recognition in Malayalam

This study investigates different approaches to recognizing accented speech for the Malayalam language, a language spoken in the southern region of India. A dataset was constructed for different language accents to conduct the study since there were no freely available datasets in the domain. The data collected has been preprocessed by applying band-pass filters and audio normalization. The speech dataset has been augmented using time-stretching, pitch shifting, and adding Gaussian noise. A total of 585 acoustic features have been extracted from the speech signals using adaptive fast Fourier transform (FFT) window size, spectral contrast, Tonnetz and polyfeatures, harmonic-to-noise ratio (HNR) and formants, zero-crossing rate (ZCR) and short-term Fourier transform, root mean square (RMS) and Mel spectrogram, and Mel-frequency cepstral coefficients (MFCC) and its deltas. Five accented models were constructed using a 2D parallel convolutional neural network (CNN), 4D parallel CNN without attention block, 4D parallel CNN with attention block, Bidirectional long short-term memory, and CNN–long short-term memory hybrid methods. The accented models constructed using 4D Parallel with attention block and hybrid CNN–long short-term memory architecture exhibited better performance with high accuracy and low error rates among all the five model architectures.