Intelligibility of interrupted meaningful and nonsense speech with and without intervening noise

Abstract

The insertion of noise in the silent intervals of interrupted speech has a very striking perceptual effect if a certain signal-to-noise ratio is used. Conflicting reports have been published as to whether the inserted noise improves speech intelligibility or not. The major difference between studies was the level of redundancy in the speech material. We show in the present paper that the noise leads to a better intelligibility of interrupted speech. The redundancy level determines the possible amount of improvement. The consequences of our findings are discussed in relation to such phenomena as continuity perception and pulsation threshold measurement. A hypothesis is formulated for the processing of interrupted stimuli with and without intervening noise; for stimuli presented with intervening noise, the presence in the auditory system of an automatic interpolation mechanism is assumed. The mechanism operates only if the noise makes it impossible to perceive the interruption.The insertion of broadband noise in the silent gaps of interrupted speech has a very striking overall auditory effect. While the interrupted speech sounds hoarse and raucous, the insertion of the noise restores fluency: the listener has the impression that he is hearing continuous speech with a superimposed pulsating noise. This phenomenon was first described by Miller and Licklider in 19'0. Whether or not the added noise does lead to a real increase in speech intelligibility has been a matter of discussion ever since. This debate is continued in the present paper.We became interested in the phenomenon after we had utilized the pulsation threshold technique (Houtgast, 1972(Houtgast, , 1974 for the measurement of frequency selectivity (Verschuure, 1977(Verschuure, , 1978(Verschuure, , 1981a(Verschuure, , 1981b. The technique is based on the finding that an alternation of two sine waves can be perceived as a superposition of one pulsating and one continuous tone, given certain relations between frequencies and levelsof the alternating sine waves. The perception of one tone as if it were continuous means that somewhere in the auditory system a gap in information is filled up.With interrupted speech and intervening noise, the gap in information must affect the intelligibility of The authors would like to thank all those who gave so willingly of their time to play the part of listener and S. Smit-Hermer for reading the speech material, Q. van den Brink for stimulating dlscuulons both during and after the Investiptlon and during the preparation of the paper. Also, the financial support received from the Helnslus Houbolt Fund Is Il'atefully acknowled,ed. The authors' mallin, address Is:

Comparing human and automatic speech recognition in a perceptual restoration experiment

“…Nonetheless, our data is in agreement with previous findings that additive noise either does not improve listener performance, or improves performance by a small amount when the additive-noise level is close to speech level. A similar observation has been reported by Verschuure and Brocaar (1983), where pilot experiments indicated that perceptual restoration is most effective when the additive-noise level exceeds the speech level by 10 dB.…”

“…Perceptual restoration studies indicate that the restoration effect depends on acoustic cues that relate to the additive noise (Powers and Wilcox, 1977;Samuel, 1981;Warren et al, 1997) and context cues that relate to the speech material (Warren and Sherman, 1974;Verschuure and Brocaar, 1983;Bashford et al, 1992). The current section describes the stimuli and evaluation metric used in the present work.…”

“…Cooke (2006) studied vowel-consonant-vowel data, whereas the speech stimuli in the current work consisted of read sentences with syntactic and semantic context. Semantic cues are expected to improve listener performance in a perceptual restoration task (Verschuure and Brocaar, 1983) and listeners can take advantage of semantic contexts spanning over sentences. In contrast, our ASR system uses a variable length n-gram language model, and can only benefit from rather short contexts (commonly no more than a few words).…”

“…A notable difference between human listeners and ASR systems, or between the spectral and temporal restoration tasks, is that human listeners, when presented with interrupted speech without additive noise, can perceive and locate what was removed, and can compensate for the information loss at a conscious level. Verschuure and Brocaar (1983) hypothesised that conscious restoration and perceptual restoration are separate processes, and that, unlike perceptual restoration, conscious restoration improves with practice.…”

“…It therefore appears that perceptual restoration only occurs if there is sufficient evidence that the speech has been masked by additive noise (Bregman, 1990). Furthermore, perceptual restoration is accompanied by a benefit in terms of speech intelligibility; listeners are better able to repeat speech content when additive noise is introduced in the removed portions (Warren, 1984;Verschuure and Brocaar, 1983;Powers and Wilcox, 1977;Warren et al, 1997). Repp (1992) proposed that perceptual restoration theories could be categorised into those based on segregation or top-down completion.…”