Histogram equalization and noise masking for robust speech recognition

Abstract

Mismatch between training and test conditions deteriorates the performance of speech recognizers. This paper investigates the combination of parametric histogram equalization (pHEQ) and noise masking to compensate for the mismatch caused by additive noise. The proposed front-end maps the distribution of the observed power spectrum vectors to a target distribution. The target distribution matches the distribution of the noise free training data except for an artificially reduced signal-to-noise ratio. Different power spectrum estimation algorithms are used to estimate the noise distribution as used internally by pHEQ more reliably under nonstationary noise conditions. The proposed front-end is evaluated on the Aurora4 database and shows a significant improvement w.r.t. mean-normalized Mel-frequency spectral coefficients. Moreover, the performance could be further improved if better estimates of the instantaneous noise power spectrum were available.

Continuous Noise Masking Based Vocoder for Statistical Parametric Speech Synthesis

“…One simple method is presented in [13] as a small amount of artificial noise is added to the clean speech to improve the noise immunity of the model and reach the desired signal-to-noise ratio (SNR). Another method with similar goals is capable of lowering the statistical mismatch of acoustic features in the training and testing conditions [14]. Moreover, a good degree of noise robustness in both filter bank and Mel-frequency cepstral domains can be found in [15].…”