Minimal spectral contrast of formant peaks for vowel recognition as a function of spectral slope

Abstract

In four experiments we investigated whether listeners can locate the formants of vowels not only from peaks, but also from spectral "shoulders"-features that give rise to zero crossings in the third, but not the first, differential of the excitation pattern-as hypothesized by Assmann and Summerfield (1989). Stimuli were steady-state approximations to the vowels [a, i, 3, U, :J] created by summing the first 45 harmonics of a fundamental of 100Hz. Thirty-nine harmonics had equal amplitudes; the other 6 formed three pairs that were raised in level to define three "formants." An adaptive psychophysical procedure determined the minimal difference in level between the 6 harmonics and the remaining 39 at which the vowels were identifiably different from one another. These thresholds were measured through simulated communication channels, giving overall slopes to the excitation patterns of the five vowels that ranged from -1 dB/erb to +2 dB/erb. Excitation patterns of the threshold stimuli were computed, and the locations of formants were estimated from zero crossings in the first and third differentials. With the more steeply sloping communication channels, some formants of some vowels were represented as shoulders rather than peaks, confirming the predictions of Assmann and Summerfield's models. We discuss the limitations of the excitation pattern model and the related issue of whether the location of formants can be computed from spectral shoulders in auditory analysis.When two people speak concurrently, a formant in one voice may be partially masked by a more intense formant in the competing voice. The partially masked formant may be displayed as only an irregularity on one of the skirts of the more intense formant in the internal representation of the spectrum. Assmann and Summerfield (1989) called such irregularities shoulders and offered a formal definition, which is restated in the next section. Assmann and Summerfield compared the ability of four models to predict the identification responses made by listeners to pairs of synthetic vowels with the same fundamental frequency (fO) that were presented concurrently to the same ear. The two models that performed best used shoulders in addition to peaks as evidence of the locations of formants. One explanation of this result is that listeners, like the models, can use shoulders to locate the positions of formants. An alternative explanaThe experiments described in this paper were performed at the Institute of Hearing Research while A.P.L. was supported by a studentship from the Medical Research Council. We wish to thank ATR Human Information Processing Research Laboratories, where this paper was written. Our thanks also go to Joanne Miller, Dianne Van Tasell, and two anonymous reviewers for helpful comments on an earlier version of this paper. Address correspondence to A. P. Lea, ATR Human Information Processing Research Laboratories, 2-2 Hikaridai, Seika-cho, Kyoto 619-02, Japan. tion is that listeners use only peaks to locate formants, but improve their performance...

The role of formant amplitude in the perception of /i/ and /u/

“…Lea and Summerfield ͑1994͒ found that the contrast threshold for a formant peak was 1.2-1.5 dB, which also agrees with results obtained in spectral enhancement studies ͑Summerfield et al, 1987;…”

Perceptual sensitivity to spectral properties of earlier sounds during speech categorization

“…While not a direct analog to context effects in speech perception, various psychophysical studies have reported sensitivity to even smaller increments in sound spectra. Normalhearing listeners detected spectral increments as small as 1 dB in studies of profile analysis (Green, 1988) and accurately recognized vowels based on spectral increments (corresponding to formant frequencies) only 1-2 dB in magnitude (Lea & Summerfield, 1994;Leek et al, 1987;Loizou & Poroy, 2001). Other studies reported high sensitivity to small spectral differences over time.…”

Modest, reliable spectral peaks in preceding sounds influence vowel perception

“…They measured the minimum intensity increment for these six sine waves that still produced robust vowel percepts. Normalhearing listeners required only 1-2 dB of spectral contrast to identify vowel sounds (see also Lea and Summerfield, 1994;Alcantara and Moore, 1995). Related studies examining enhancement effects and detection of spectral troughs (Turner and Van Tasell, 1984) also produced high levels of performance with only 1-2 dB of spectral contrast.…”