Objective quality evaluation in blind source separation for speech recognition in a real room

Abstract

The determination of quality of the signals obtained by blind source separation is a very important subject fordevelopment and evaluation of such algorithms. When this approach is used as a pre-processing stage for automatic speechrecognition, the quality measure of separation applied for assessment should be related to the recognition rates of thesystem. Many measures have been used for quality evaluation, but in general these have been applied without priorresearch of their capabilities as quality measures in the context of blind source separation, and often they requireexperimentation in unrealistic conditions. Moreover, these measures just try to evaluate the amount of separation, and thisvalue could not be directly related to recognition rates. Presented in this work is a study of several objective qualitymeasures evaluated as predictors of recognition rate of a continuous speech recognizer. Correlation between qualitymeasures and recognition rates is analyzed for a separation algorithm applied to signals recorded in a real room withdifferent reverberation times and different kinds and levels of noise. A very good correlation between weighted spectralslope measure and the recognition rate has been verified from the results of this analysis. Furthermore, a good performanceof total relative distortion and cepstral measures for rooms with relatively long reverberation time has been observed