Evaluation of Speech Enhancement Based on Pre-Image Iterations Using Automatic Speech Recognition

Recently, we developed pre-image iteration methods for single-channel speech enhancement. We used objective quality measures for evaluation. In this paper, we evaluate the de-noising capabilities of pre-image iterations using an automatic speech recognizer trained on clean speech data. In particular, we provide the word recognition accuracy of the de-noised utterances using white and car noise at 0, 5, 10, and 15 dB signal-to-noise ratio (SNR). Empirical results show that the utterances processed by pre-image iterations achieve a consistently better word recognition accuracy for both noise types and all SNR levels compared to the noisy data and the utterances processed by the generalized subspace speech enhancement method.