Learning Filterbanks from Raw Waveform for Accent Classification

Abstract

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-toend 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.

Deep neural architectures for dialect classification with single frequency filtering and zero-time windowing feature representations

“…Instead of using fixed mel-scale spectral filters in feature representations for input to CNN, data-driven learnt spectral scale filters (as convolution layer) for dialect classification are investigated. Note that learnt spectral scale filters are well known and previously used for speech recognition (Seki et al, 2017), spoofing detection (Yu et al, 2017), and accent classification (Kethireddy et al, 2020b). As per our knowledge, this is the first attempt to propose to use learnt spectral scale filters for dialect classification.…”

“…From the early studies on accent classification (Kat and Fung, 1999;Arslan and Hansen, 1997), it was found that the favourable spectral scale depends on the language of dialects and sub-dialects contained in it. Furthermore, from the accent classification studies with neural networks (Kethireddy et al, 2020b), it was found that the distribution of learnt frequency bands are different from standard mel-scale distribution. It was observed that learnt scale showed an improvement of 10.94% UAR (relative) over mel-scale.…”

A review of social background profiling of speakers from speech accents

“…The fusion reported an accuracy of 86.05%. Kethireddy, Kadiri & Gangashetty (2020) have also used CNN with the raw audio as input and reported 81.26% accuracy over a subset of CV dataset comprising eight accents.…”

Recognition of Arabic Accents From English Spoken Speech Using Deep Learning Approach

“…The work of Kethireddy et al [22], focused on learning filterbanks that are initialized using customized features obtained from raw waveform, which are incorporated into the CNN network for English accent recognition. The experimental results demonstrate strong performance with an accuracy of 81.26%; these results were attained by using their techniques on a common dataset of 8 English accents.…”