Learning Filterbanks from Raw Waveform for Accent Classification

Most of the applications in speech use mel-frequency spectral coefficients (MFSC) as features as they match the human perceptual mechanism, where the emphasis is given to vocal tract characteristics. But in accent classification, mel-scale distribution of filters may not always be the best representations, e.g., pitch accented languages where the emphasis should be on vocal source information too. Motivated by this, we use end-to-end classification of accents directly from waveforms which will reduce the effort of designing features specific to each corpus. The convolution neural network (CNN) model architecture is designed in such a way that the initial layers exhibit similar operation as in MFSC by initializing the weights using time approximate of MFSC. The entire network along with initial layers is trained to learn accent classification. We observed that learning directly from waveform improved the performance of accent classification when compared to CNN trained on hand-engineered features by 10.94% UAR on the test dataset of common voice corpus. Analyzing the filters after learning, we observed changes in distribution and bandwidths of center frequencies. We further observed the importance of appropriately initializing CNN filters.