Robust Multi-Band ASR Using Deep Neural Nets and Spectro-temporal Features

Abstract

Abstract. Spectro-temporal feature extraction and multi-band processing were both designed to make the speech recognizers more robust. Although they have been used for a long time now, very few attempts have been made to combine them. This is why here we integrate two spectrotemporal feature extraction methods into a multi-band framework. We assess the performance of our spectro-temporal feature sets both individually (as a baseline) and in combination with multi-band processing in phone recognition tasks on clean and noise contaminated versions of the TIMIT dataset. Our results show that multi-band processing clearly outperforms the baseline feature recombination method in every case tested. This improved performance can also be further enhanced by using the recently introduced technology of deep neural nets (DNNs).

Multi-band Approach to Deep Learning-Based Artificial Stereo Extension

“…Multi-band or sub-band representation of audio has been utilized in a variety of applications, such as audio coding [10], [11], audio enhancement [12], audio upmixing [13], and automatic speech recognition (ASR) [14]. For example, several audio coding standards, including MPEG Surround [10] and MPEG-H [11], have incorporated a quadrature mirror filter (QMF) to obtain uniformly distributed and oversampled frequency representations of audio signals.…”

“…In addition, multiband representation has been applied to audio enhancement and has improved the enhancement quality by variously performing noise attenuation according to the given sub-bands [12]. Moreover, a DNN was realized in a sub-band manner for ASR, which resulted in the reduction of the average word error rate compared to that in a full-band manner [14].…”