Speech intelligibility prediction with the dynamic compressive gammachirp filterbank and modulation power spectrum

Abstract

The speech-based envelope power spectrum model (sEPSM) was developed to predict the speech intelligibility of sounds produced by nonlinear speech enhancement algorithms such as spectral subtraction. It is a linear model with a linear, level-independent gammatone (GT) filterbank as the front-end. Therefore, it seems difficult to evaluate speech sounds with low and high sound pressure levels (SPLs) consistently because the intelligibility of the speech is dependent on the SPL as well as the signal-to-noise ratio. In this study, the sEPSM was extended with the dynamic compressive gammachirp (dcGC) auditory filterbank and a ''common'' normalization factor of the modulation power spectrum component to improve the predictability of the model. For evaluating the proposed model, we performed subjective experiments on the intelligibility of speech sounds enhanced by spectral subtraction and a Wiener filter algorithm. We compared the subjective speech intelligibility scores with the objective scores predicted by the proposed dcGC-sEPSM, original GT-sEPSM, and other well-known conventional methods such as the short-time objective intelligibility measure (STOI), coherence speech intelligibility index (CSII), and hearing aid speech perception index (HASPI). The result shows that the proposed dcGC-sEPSM predicted the subjective results better did than the other methods.

The gammachirp auditory filter and its application to speech perception

“…The dcGC-FB was also used in models for speech intelligibility prediction [9][10][11][12][13]. A new model referred to as GEDI (the gammachirp envelope distortion index) [10,11] predicted the intelligibility of speech sounds processed with non-linear enhancement algorithms better than other recent indexes like STOI, CSII, and HASPI [12,13].…”

“…The GC architecture enables us to construct a hearing impairment simulator which allows normal hearing listeners to experience the difficulties of hearing impaired listeners [5,6]. The GC has also been used to model speaker size perception [7,8] and speech intelligibility [9][10][11][12][13].…”

GEDI: Gammachirp Envelope Distortion Index for Predicting Intelligibility of Enhanced Speech

“…To incorporate characteristics of a human auditory filter, Yamamoto et al (2019) extended sEPSM using a dynamic compressive gammachirp filterbank (dcGC-FB) (Irino & Patterson, 2006), in which the level-dependent frequency selectivity and gain of the auditory filter were reasonably determined by the data obtained from psychoacoustic masking experiments (Patterson et al, 2003). For OIMs, it is important to introduce the appropriate level dependency to incorporate the well-known fundamental knowledge that speech intelligibility is lower as sound level decreases and that peripheral hearing loss decreases the intelligibility.…”

“…A bank of modulation filters, defined in envelope frequency domain (f env ), is applied to the spectra. There are seven modulation filters whose power spectra are W f c env (f env ) for the modulation center frequency of f c env , as illustrated in Figure 2 and described in previous studies (Jørgensen & Dau, 2011;Yamamoto et al, 2019). The envelope power at the output of the modulation filter is calculated as…”

“…where the asterisk (*) represents either S or D, and E Ŝ (0) represents the 0-th order coefficient of the FFT (i.e., the direct-current (DC) component of the temporal envelope). Yamamoto et al (2019) reported that normalization of the E Ŝ (0) in dcGC-sEPSM was effective for speech intelligibility prediction of enhanced speech. The normalization has been inherited by GEDI which has the same filterbank structure as dcGC-sEPSM.…”