Predicting the effect of spectral subtraction on the speech recognition threshold based on the signal-to-noise ratio in the envelope domain

Digital noise reduction strategies are important in technical devices such as hearing aids and mobile phones. One well-described noise reduction scheme is the spectral subtraction algorithm. Many versions of the spectral subtraction scheme have been presented in the literature, but the methods have rarely been evaluated perceptually in terms of speech intelligibility.

This study analyzed the effects of the spectral subtraction strategy proposed by Berouti at al. [ICASSP 4 (1979), 208-211] on the speech recognition threshold (SRT) obtained with sentences presented in stationary speech-shaped noise. The SRT was measured in five normal-hearing listeners in six conditions of spectral subtraction.

The results showed an increase of the SRT after processing, i.e. a decreased speech intelligibility, in contrast to what is predicted by the Speech Transmission Index (STI). Here, another approach is proposed, denoted the speech-based envelope power spectrum model (sEPSM) which predicts the intelligibility based on the signal-to-noise ratio in the envelope domain.

In contrast to the STI, the sEPSM is sensitive to the increased amount of the noise envelope power as a consequence of the spectral subtraction operation, which leads to a decreased speech intelligibility in this model, in quantitative agreement with the experimental data.