Comparative experiments to evaluate a voiced-unvoiced-based pre-processing approach to robust automatic speech recognition in low-SNR environments

This paper presents an evaluation of a robust Voiced-Unvoicedbased large-vocabulary Continuous-Speech Recognition (CSR) system in the presence of highly interfering noise. Comparative experiments have indicated that the inclusion of an accurate Voiced-Unvoiced (V-U) classifier in our design of a CSR system improves the performance of such a recognizer, for speech contaminated by both additive Gaussian and uniform noises. Our results show that the V-U-based CSR system outperforms the CMS-based and the RASTA-PLP-based CSR systems in such environments for a wide range of SNRs.