Concurrent vowel identification. I. Effects of relative amplitude and F difference

Abstract

Subjects identified concurrent synthetic vowel pairs that differed in relative amplitude and fundamental frequency (F 0). Subjects were allowed to report one or two vowels for each stimulus, rather than forced to report two vowels as was the case in previously reported experiments of the same type. At all relative amplitudes, identification was better at a fundamental frequency difference (⌬F 0) of 6% than at 0%, but the effect was larger when the target vowel amplitude was below that of the competing vowel ͑Ϫ10 or Ϫ20 dB͒. The existence of a ⌬F 0 effect when the target is weak relative to the competing vowel is interpreted as evidence that segregation occurs according to a mechanism of cancellation based on the harmonic structure of the competing vowel. Enhancement of the target based on its own harmonic structure is unlikely, given the difficulty of estimating the fundamental frequency of a weak target. Details of the pattern of identification as a function of amplitude and vowel pair were found to be incompatible with a current model of vowel segregation.

The auditory organization of speech and other sources in listeners and computational models

“…· A dierence in fundamental frequency between pairs of concurrent vowels leads to an absolute improvement of 10±15% in vowel identi®cation performance, the eect beginning at a dierence as small as a quarter of a semitone and asymptoting by 2 semitones. This basic ®nding of Scheers (1983) has been replicated by several researchers (Assmann and Summer®eld, 1990;Culling and Darwin, 1993;Lea, 1992;Meddis and Hewitt, 1992;de Cheveign e et al, 1997a). · A dierence in mode of excitation (voiced/whispered) between the constituent vowels leads to an identi®cation improvement of around 10% (Lea, 1992).…”

Recovering sound sources from embedded repetition

“…Studies of the cocktail party problem have traditionally used unnatural synthetic stimuli (9,27,28) or familiar real-world sounds such as speech (3,10,12,29). Generative models have the advantage of producing novel stimuli that lack the confounding effects of familiarity but that share properties of natural sounds.…”

“…For instance, listeners implicitly assume that frequency components that are regularly spaced (9,10), begin and end simultaneously (11), or have similar distributions of binaural spatial cues (12) belong to the same sound. Listeners also use knowledge of specific familiar sound classes, filling in masked syllable segments in ways that are consistent with known speech acoustics (13).…”

Spectral Processing and Sound Source Determination

“…In the case of two simultaneous vowels produced with slightly different f0, harmonics of one f0 will be interspersed with harmonics of the other f0. However, listeners are able to identify simultaneous vowels at levels well above chance Summerfield, 1989, 1990;Bregman, 1989, 1993;Culling and Darwin, 1994;de Cheveigné et al, 1997). To identify any single vowel presented in isolation, the listener makes use of the vowel's spectral profile.…”

“…Simultaneous periodic sounds that differ in f0, that is, that differ in harmonicity, are more easily segregated than sounds that have the same f0. The introduction of small differences in f0, on the order of 1% to 3%, leads to rapid improvement in identification accuracy, and f0 differences greater than 1 semitone usually produce only modest additional improvement Summerfield, 1989, 1990;Bregman, 1989, 1993;Culling and Darwin, 1994;de Cheveigné et al, 1997de Cheveigné et al, , 1999. Ongoing speech is more intelligible in the presence of a competing voice if the pitch of the second voice is different (Brokx and Nooteboom, 1982).…”