Effects of spectral smearing on phoneme and word recognition

Abstract

The principal goal was to measure the effects, on speech perception, of loss of spectral detail in the acoustic signal. Spectral smearing was produced by multiplying the speech waveform by low-pass filtered noise. Performance was measured in normal adults as the percentage of phonemes correctly repeated in lists of monosyllabic words. A smearing bandwidth of 250 Hz (i.e., each tonal component of the instantaneous spectrum replaced by a 250-Hz band of noise) had a small but significant effect on phoneme recognition. A smearing bandwidth of 8000 Hz was required to reduce phoneme recognition to a value that was indistinguishable from that produced by complete smearing. Vowels were somewhat more susceptible to the effects of spectral smearing than were consonants, and initial consonants were more susceptible than were final consonants. In an analysis of errors, place of consonant articulation was more susceptible than either manner of articulation or voicing. These findings are attributable to differences in the relative importance of spectral and temporal cues. Word recognition was more susceptible to the effects of spectral smearing than was phoneme recognition, but this finding was predictable on the basis of the known nonlinear relationship between the two measures. In a second experiment, smearing bandwidths of 707 and 2000 Hz increased phoneme recognition threshold by 12.9 and 16.4 dB, respectively, compared to that found without smearing. (Phoneme recognition threshold is defined, here, as the signal-to-noise ratio at which phoneme recognition is 50% of the value obtained in quiet.) The data are consistent with the hypothesis that reduced spectral resolution affects phoneme recognition to the extent that it reduces access to the formant patterns in the spectral envelope.