A generalized framework for compensation of mel-filterbank outputs in feature extraction for robust ASR

This paper describes a novel and efficient noise-robust frontend that utilizes a set of Mel-filterbank output compensation methods, together with cumulative distribution mapping of cepstral coefficients, for noisy speech recognition. The proposed compensation framework includes the use of noise spectral subtraction, spectral flooring and log Mel-filterbank output weighting. Recognition experiments on the Aurora II connected digit database have revealed that the proposed front-end achieves an average digit recognition accuracy of 83.46% for a model set trained from clean data. Compared with the recognition results obtained by using the ETSI standard Mel-cepstral front-end, these results represent a relative error reduction of around 58%.