Performance estimation of noisy speech recognition using spectral distortion and recognition task complexity

Abstract

Previously, methods for estimating the performance of noisy speech recognition based on a spectral distortion measure have been proposed. Although they give an estimate of recognition performance without actually performing speech recognition, no consideration is given to any change in the components of a speech recognition system. To solve this problem, we propose a novel method for estimating the performance of noisy speech recognition, a major feature of which is the ability to accommodate the use of different noise reduction algorithms and recognition tasks by using two cepstral distances (CDs) and the square mean root perplexity (SMR-perplexity). First, we verified the effectiveness of the proposed distortion measure, i.e., the two CDs. The experimental results showed that the use of the proposed distortion measure achieves estimation accuracy equivalent to the use of the conventional distortion measures, the perceptual evaluation of speech quality (PESQ) and the signal-to-noise ratio (SNR) of noise-reduced speech, and has the advantage of being applicable to noise reduction algorithms that directly output the mel-frequency cepstral coefficient (MFCC) feature. We then evaluated the proposed method by performing a closed test and an open test (10-fold crossvalidation test). The results confirmed that the proposed method gives better estimates without being dependent on the differences among the noise reduction algorithms or the recognition tasks.