On the Use of Forward Temporal Masking and Cumulative Distribution Mapping for Noisy Speech Recognition

Robustness in the presence of various types and levels of environmental noise remains an important issue for automatic speech recognition (ASR) systems. This paper describes a new noise-robust ASR front-end that employs a functional model of forward temporal masking combined with cumulative distribution mapping based on MFCC´s with c0. Recognition experiments on the Aurora II connected digits database reveal that the proposed front-end achieves an average digit recognition accuracy of 83.24% for a model set trained from clean data and 90.32% for a model set trained from data with multiple noise conditions. Compared with the ETSI standard Mel-cepstral front-end, the proposed front-end obtains a relative error reduction of around 57% for the clean model set and 21% for the multi-condition model set.