EFFECTS OF SPECTRAL SMEARING ON PHONEME AND WORD RECOGNITION

The principal goal was to measure the effects, on speech perception, o f loss of spectral detail in the acoustic signal. Spectral smearing wa s produced by multiplying the speech waveform by low-pass filtered noi se. Performance was measured in normal adults as the percentage of pho nemes correctly repeated in lists of monosyllabic words. A smearing ba ndwidth of 250 Hz (i.e., each tonal component of the instantaneous spe ctrum replaced by a 250-Hz band of noise) had a small but significant effect on phoneme recognition. A smearing bandwidth of 8000 Hz was req uired to reduce phoneme recognition to a value that was indistinguisha ble 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 consonan ts. In an analysis of errors, place of consonant articulation was more susceptible than either manner of articulation or voicing. These find ings are attributable to differences in the relative importance of spe ctral and temporal cues. Word recognition was more susceptible to the effects of spectral smearing than was phoneme recognition, but this fi nding 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 an d 16.4 dB, respectively, compared to that found without smearing. (Pho neme recognition threshold is defined, here, as the signal-to-noise ra tio at which phoneme recognition is 50% of the value obtained in quiet .) The data are consistent with the hypothesis that reduced spectral r esolution affects phoneme recognition to the extent that it reduces ac cess to the formant patterns in the spectral envelope. (C) 1996 Acoust ical Society of America.