Yifei Zhao

2022

1. MLSP

   An Efficient Transformer-Based Model for Voice Activity Detection

   Yifei Zhao, and Benoit Champagne

   In 2022 IEEE 32nd International Workshop on Machine Learning for Signal Processing (MLSP), Aug 2022

   Abs HTML

   Voice Activity Detection (VAD) aims to distinguish, at a given time, between desired speech and non-speech. Although many state-of-the-art approaches for increasing the performance of VAD have been proposed, they are still not robust enough to be applied under adverse noise conditions with low Signal-to-Noise Ratio (SNR). To deal with this issue, we propose a novel transformer-based architecture for VAD with reduced computational complexity by implementing efficient depth-wise convolutions on feature patches. The proposed model, named Tr-VAD, demonstrates better performance compared to baseline methods from the literature in a variety of scenarios considered with the smallest possible number of parameters. The results also indicate that the use of a combination of Audio Fingerprinting (AFP) features with Tr-VAD can guarantee better performance.

2. ICASSP

   Complex IRM-Aware Training for Voice Activity Detection Using Attention Model

   Yifei Zhao, Yazid Attabi, Benoit Champagne, and 1 more author

   In ICASSP 2022 - 2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), May 2022

   Abs HTML

   Although many state-of-the-art approaches for improving the accuracy of Voice Activity Detection (VAD) have been proposed, their performance under adverse noise conditions with low Signal-to-Noise Ratio (SNR) remains limited. In this paper, we introduce a novel attention model-based deep neural network (DNN) architecture for VAD which takes advantage of complex Ideal Ratio Mask (cIRM). The proposed model, named AM-cIRM, consists of three sequential modules: extraction of cIRM features from the noisy speech using a DNN-based architecture; combination of cIRM with log-Mel spectrogram features along with temporal contextual extension; and VAD using an attention model that exploits the spectro-temporal information in the transformed features. Experimental results show that the proposed AM-cIRM achieves improved VAD performance when compared to state-of-the-art methods under different noise conditions.