Perceptual compensation for speaker differences and for spectral-envelope distortion

Abstract

This study asks whether perceptual mechanisms that compensate for the spectral-envelope distortion of transmission channels also contribute to compensation for speaker differences. Subjects identified test words that were played after a carrier sentence. In some conditions the carriers were synthesized with F1 in low- and high-frequency ranges and in others they were distorted by filters whose frequency response is the spectral envelope of one vowel minus the spectral envelope of another. The filter /I/ minus /epsilon/ and its inverse were used. Test words were drawn from an /Itch/ to /epsilon tch/ continuum. Carriers filtered by /I/ minus /epsilon/ and its inverse give a phoneme boundary difference, indicating compensation for spectral envelope distortion. A phoneme boundary difference also occurs between carriers with F1 in low and high ranges, indicating compensation for speaker differences. Neither of these effects is reduced by playing the carrier backwards, even though a measurement of the perceived naturalness of carriers is sharply reduced by this manipulation. Analysis of carriers synthesized with low and high F1 showed that they have different long-term spectra, and subsequent experiments used time-stationary filters to alter this characteristic. The results showed that the long-term spectra of the carriers govern their influence on the identity of subsequent test sounds. However, measurements of perceptual confusions among the carriers and of perceived talker-differences between carriers revealed that other, time-varying factors are more important for voice identification.

The effect of language experience on perceptual normalization of Mandarin tones and non-speech pitch contours

“…Although context effects make phoneme recognition robust against speaker variability or speech co-articulation, the perceptual process may not require context information about speaker identity, vocal tract size, articulatory gesture, or phonetic space. For example, the contrastive context effects on consonant and vowel recognition reported by Mann (1980) and Ladefoged and Broadbent (1957) have been replicated by Lotto and Kluender (1998) and Watkins and Makin (1994), respectively, using non-speech contexts to model the formant structures of speech contexts. These results suggest that phonetic normalization may arise from general auditory processing of spectral contrast between context and target, but not necessarily from speaker-, phoneme-, or speech-specific processing.…”

Temporal properties of perceptual calibration to local and broad spectral characteristics of a listening context

“…Using synthesized vowels and frequency-modulated resonance filters, the current studies were conducted to better understand the temporal course of auditory perceptual calibration, in which relatively reliable filter characteristics of the environment or listening context are perceptually attenuated or "inverse filtered" (Watkins and Makin, 1994). Two filter characteristics, local spectral peaks and gross spectral tilt, were examined because of their codetermination in the phonetic identity of the vowels /u/ and /i/.…”

“…Despite multiple demonstrations incorporating both speech and non-speech signals (e.g., Holt, 2005Holt, , 2006Kiefte and Kluender, 2008;Stilp et al, 2010;Watkins, 1991;Watkins and Makin, 1994), relatively little is known about processes underlying perceptual calibration to listening context. Particularly with respect to applications of modern signal processing technology, natural biological mechanisms of perceptual calibration need to be better understood.…”

“…In each case, listeners were more likely to report hearing the vowel sound that shared the least in common with the precursor spectrum. Because spectral similarities to one vowel (e.g., /I/) in the long-term spectrum of the precursor were perceptually mitigated and spectral differences to the other vowel (e.g., /e/) were enhanced, Watkins and Makin (1994) characterize this calibration as "inverse filtering." This perceptual calibration and contrast were demonstrated across different speech contexts that varied in talker gender, spatial position, ear of presentation, forward or time-reversed, and even for speech-shaped, signal-correlated noise.…”

“…Listeners calibrated to the spectral consequences of the filter such that perception of the following vowel again matched that for unfiltered vowels heard in isolation. Watkins (1991) and Watkins and Makin (1994) demonstrated perceptual calibration to long-term spectral characteristics of the listening context by creating filters that captured the differences in spectral shapes between vowels /I/ and /e/ and applying these filters to a precursor sentence that preceded an "etch" to "itch" series. When the preceding context was passed through the /I/ minus /e/ difference filter, probability of listeners responding "etch" increased.…”