2Diebold, Inc., 5995 Mayfair Road, North Canton, OH 44720, U.S.A.

Moore and Jongman created three synthetic continua between Tones 2 and 3: the first continuum varied only in D F0 (the difference in Hz between F0 at onset and valley), a spectral parameter; the second varied only in the position of the 'Turning Point' (the point in time where the F0 valley occurs, a temporal parameter); and the third continuum varied in both D F0 and Turning Point. Each of these continua was then preceded by a precursor sentence with an F0 range appropriate for either a high F0 speaker or a low F0 speaker. Thus, listeners' responses to a

The continuum that varied only in terms of D F0 showed a significant shift in the identification functions: identification of a target tone, ambiguous between Tone 2 and Tone 3, shifted depending on whether the target was preceded by a precursor with a high or low F0 range. Moreover, similar to Ladefoged and Broadbent's results, the effect was contrastive such that the low F0 precursor caused a shift toward the higher Tone 2, while the high F0 precursor caused a shift toward the lower Tone 3.

A similar, significant, contrast effect was observed for the continuum varying in both D F0 and Turning Point. Finally, no shift was observed for the continuum that varied only in terms of Turning Point (Th). A high or low F0 precursor did not shift the perception of tone specified by only temporal parameters.

These results suggest that normalization only occurs if context and target vary along the same acoustic dimension. To further investigate this hypothesis, the present study evaluates whether normalization of temporal information also only occurs if context and target both vary along the temporal dimension. The temporal variable that we decided to investigate was speaking rate. Each of the three Tone 2-Tone 3 continua (D F0, TP, D F0/TP) was preceded by a precursor sentence with either a slow or fast speaking rate.

The following three experiments tested the hypothesis that listeners perceive tones in part by normalizing for speaking rate of a precursor. Subjects heard a precursor (fast or slow) plus a stimulus from the Tone 2-Tone 3 continuum described in [3]. The precursors (the segmental context was Zheige zi nian ____ ('This word is ___')), were modified from a natural utterance of a female speaker. The mean F0 range of the precursor was

The target words (lul) (Tone 2 'not', Tone 3 dance') were 11 members of the synthetic Tone 2-Tone 3 continuum varying in D F0, Turning Point, or both (as used in [3]), in which D F0 varied from 25 Hz to 75 Hz (stepsize: 5 Hz), and Turning Point varied from 20 ms to 220 ms into the tone (stepsize: 20 ms).

A total of 33 native speakers of Mandarin participated in this experiment (D F0: 10, TP: 11, D F0/TP: 12). The test consisted of a total of 220 trials (11 stimuli x 2 rates x 10 repetitions). Subjects were instructed to respond by pressing one of two buttons, which were labeled using the Chinese characters for either 'not', or 'dance', corresponding to the Tone 2 or Tone 3 lexical item, respectively.

For each subject, phoneme boundaries were determined by using a probit statistical analysis. Responses were averaged across subjects to obtain the identification functions. For the D F0/Turning Point continuum, a two-tailed, paired t-test revealed no significant effect of precursor rate [t( 11 )=- 1.5, p>.l 5]. For the stimuli varying only in D F0 no significant effect of precursor rate was observed [t(9)=- 1.5, p>.l 4]. Finally, for the stimuli varying only in Turning Point a significantly earlier shift toward Tone 3 responses was obtained in the fast precursor condition, [t( l0)=-2.5 1, p<.03]. A contrast effect was thus observed when precursor and target both varied exclusively along the temporal dimension.

A total of 33 native monolingual speakers of English served as subjects (D F0: 11, TP: 10, D F0/TP: 12). Materials and procedure were virtually identical to those described for Experiment I. Each experimental session included three tasks. In task I, subjects listened to the endpoints in pairs 10 times, where the first stimulus in each pair, the Tone 2 endpoint, was labeled

For the D F0/Turning Point continuum there was a significantly earlier shift toward Tone 3 responses in the fast precursor condition [t(11)=-2.43, p<.03]. For the D F0 continuum, no significant difference in responses between the two speaking rates was found [t(l0)=- 1.05, p>.3]. Finally, for the Turning Point continuum no significant effect of precursor speaking rate was observed [t(9)=-0.86, p>.4].

A total of 30 monolingual English speakers participated in the experiment (D F0: 8, TP: 9, D F0/TP: 13). Stimuli were identical to those in [3] for the Mandarin listeners.

For the D F0/Turning Point continuum a shift occurred toward low tone responses in the high precursor condition. The difference between responses in the two conditions was significant [t(12)= -2.29, p<.04]. For the D F0 continuum there was no significant shift in identification as a function of the precursor [t(7) = -.13, p>.9]. Finally, for the stimuli varying in Turning Point there was no significant shift as a function of the precursor [t(8)=l.12, p>.3].

Included in this study on tone perception were experiments on rate normalization by both Mandarin and English listeners, and on speaker normalization by English listeners. Each experiment tested perception of acoustic cues demonstrated to be relevant to distinguishing Tones 2 and 3 in Mandarin. These acoustic cues were tested independently and in combination. They are: D F0, a fundamental frequency parameter; Turning Point (TI)), a temporal D F0 and Turning Point together.

There are two major findings of this normalization study, summarized in Table 1. First, Mandarin listeners normalize for both speaker F0 range and speaking rate information in tone perception. The general pattern for the Mandarin listeners is that normalization only occurs when precursors and stimuli vary in the same acoustic dimension. Second, language background influences normalization. The different pattern of findings for the two language groups suggests that language background aids in disambiguating phonemic contrasts for Mandarin listeners, but that for English listeners the normalization effects are a consequence of acoustic discriminability.

Table 1: Summary of normalization findings. Precursors were manipulated either in terms of F0 (Speaker normalization) or speaking rate (Rate normalization). Target stimuli varied either in F0 (D F0), Turning Point (TP), or both (D F0/TP). Speaker normalization results for Mandarin listeners were reported in [3]. * indicates p<.05, ** indicates p<.00l, ns is nonsignificant.

5.1 Speaker and Rate Normalization by Mandarin Listeners

The results of Experiment I and those of Moore and Jongman [3] indicate that Mandarin listeners normalize for both F0 range and speaking rate in tone perception. Results also show that Mandarin listeners are sensitive to the degree of acoustic specification of the tone stimuli; the single-cue continuum contained less intrinsic information, elevating the importance of the extrinsic F0 range information. One hypothesis that emerged from the speaker normalization experiment [3] predicted that normalization occurs when both stimuli and precursors vary in the same acoustic dimension. This is because normalization effects were observed for stimuli varying in the F0 dimension (D F0, and D F0/TP), but not for stimuli varying only in the temporal dimension (Turning Point). This hypothesis was tested in Experiment I which investigated whether Turning Point is perceived relative to a temporal frame of reference established by speaking rate. Findings confirmed the prediction that Turning Point is perceived with respect to speaking rate, and also supported a corollary prediction, that no rate effect would be observed for the non-temporal, D F0 continuum. The lack of a rate effect for the D F0/TP continuum suggested that the temporal dimension was less informative than the F0 dimension. These results indicate that listeners refer to extrinsic acoustic information only when it is relevant to disambiguating intrinsic cues.

Despite differences in results for the single-cue continua, however, both language groups showed normalization effects in perception of the D F0/TP continua. Our hypothesis that normalization occurs only when listeners are making phonemic judgments predicted, for example, that only Mandarin listeners would compensate for speaker F0 range. Results showed that English listeners also take extrinsic acoustic information into account, even when identifying non-phonemic contrasts. The English effects in the D F0/TP continuum do not constitute phonemic normalization, as it was for the Mandarin listeners. Rather, the English results reflect a different perceptual process from the one employed by the Mandarin listeners, governed by the perceptual salience of the acoustic differences.

A role for perceptual salience is suggested by the finding that English listeners show different patterns of effects for identical acoustic manipulations; that is, no effect in the single-cue continua, but an effect when the manipulations of D F0or TP occur together in the D F0/TP continuum. To account for these results, it is proposed that listeners take the extrinsic pitch and timing information into account whenever possible, but that limitations on perceptual resources allow English listeners to attend to extrinsic information only when intrinsic acoustic differences become more perceptually salient. That is, we suggest that for English listeners, the cost of normalizing extrinsic acoustic information was too high in the single-cue continua.

1. Ladefoged, P., and Broadbent, D. E. "Information conveyed by vowels," J. Acoust Soc. Am. 29, 98-104, 1957.

2. Johnson, K. "The role of perceived speaker identity in FO normalization of vowels," J. Acoust Soc. Am. 88, 642-654, 1990.

3. Moore, C. B., and Jongman, A. 'Speaker normalization in the perception of Mandarin Chinese tones," J. Acoust Soc. Am. 102, 1864-1877, 1997.

4. Quatieri, T. F., and McAulay, R. J. "Speech transformations based on a sinusoidal representation," IEEE Trans. Acoust, Speech, Signal Process. (ASSP34) 449-1464, 1986.

5. Mullennix, J. W., Pisoni, D. B., Martin, C. S. "Some effects of talker variability on spoken word recognition," J. Acoust Soc. Am. 85, 365-378, 1989.

6. Sommers, M.S., Nygaard, L.C., and Pisoni, D. B. "Stimulus variability and the perception of spoken words: Effects of variations in speaking rate and overall amplitude," In ICSLP 92 Proceedings, edited by J. I. Ohala, T M. Nearey, B. L. Derwing, M. M. Hodge, and G. E. Wiebe (Priority Printing, Edmonton, Canada), 217-220, 1992.

7. Verbrugge, R. R., Strange, W., Shankweiler, D. P., and Edman, T. R. "What information enables a listener to map a talker's vowel space?" J. Acoust. Soc. Am. 60, 198-212, 1976.