Chapter 4 - Spectral and temporal effects of signals on speech intelligibility

Abstract

Signal dynamics is representative of perceptual signatures of sound in the time domain. Narrowband envelopes expressing frequency-dependent signal dynamics are keys to estimating intelligibility of speech materials. On the other hand, waveforms can be decomposed into the magnitude and phase spectral components in the frequency domain. An interesting question from a perceptual frame of reference would be whether the magnitude- or phase-spectral component is dominant for speech samples to be intelligible. This chapter investigates the spectral dominance on the intelligibility through synthesizing spectrally (magnitude or phase) or temporally modified (reversed or random permutation) speech materials and listening tests. Results confirm that keeping narrowband envelopes is significant in order to preserve the intelligibility of speech materials. For the preservation of narrowband envelopes or the frequency-dependent signal dynamics in the time domain, the spectral dominance may be temporally dependent. Frame-wise (temporally dependent) spectral analysis reveals the magnitude-spectral dominance in the medium range of frame lengths around 30(ms) for speech to be intelligible, while the phase may be significant in other shorter or longer frame length conditions. The magnitude dominance region provides appropriate frame length conditions for frequency resolution for selecting narrowband frequencies such as 1/4-octave band and temporal resolution for envelope periods. In contrast, the spectral structure obtained in the longer frame length conditions is fine enough to reconstruct the envelopes from the phase. In the very shorter frame length conditions, the micro structure of the waveform such as zero-crossings in the time domain could be preserved so that the envelopes are preserved by the phase spectrum.