Importance of tonal envelope cues in Chinese speech recognition.

Abstract

Recent studies have shown that temporal waveform envelope cues can provide significant information for English speech recognition. This study investigated the use of temporal envelope cues in a tonal language: Mandarin Chinese. In this study, the speech was divided into several frequency analysis bands; the amplitude envelope was extracted from each band by half-wave rectification and low-pass filtering and was used to modulate a noise of the same bandwidth as the analysis band. These manipulations preserved temporal and amplitude cues in each frequency band, but removed the spectral detail within each band. Chinese vowels, consonants, tones and sentences were identified by 12 native Chinese-speaking listeners with 1, 2, 3, and 4 noise bands. The results showed that the recognition score of vowels, consonants, and sentences increased monotonically with the number of bands, a pattern similar to that observed in English speech recognition. In contrast, tones were consistently recognized at about 80% correct level, independent of the number of bands. This high level of tone recognition produced a significant difference in the open-set sentence recognition between Chinese (11.0%) and English (2.9%) for the one-band condition where no spectral information was available. The data also revealed that, with primarily temporal cues, the falling-rising tone (tone 3) and the falling tone (tone 4) were more easily recognized than the flat tone (tone 1) and the rising tone (tone 2). This differential pattern in tone recognition resulted in a similar pattern in word recognition: words having either tone 3 or 4 were more likely to be recognized while words having tone 1 and 2 were not. The quantitative role of tones in Chinese speech recognition was further explored using a power-function model and found to play a significant role in relating phoneme recognition to sentence recognition.

Full text

Page 1

Importance of tonal envelope cues in Chinese speech recognition
Qian-Jie Fua)
Department of Auditory Implants and Perception, House Ear Institute, 2100 West Third Street, Los Angeles,
California 90057 and Department of Biomedical Engineering, University of Southern California,
Los Angeles, California 90089
Fan-Gang Zeng
Department of Auditory Implants and Perception, House Ear Institute, 2100 West Third Street, Los Angeles,
California 90057 and Department of Electrical Engineering, University of Southern California,
Los Angeles, California 90089
Robert V. Shannon
Department of Auditory Implants and Perception, House Ear Institute, 2100 West Third Street, Los Angeles,
California 90057 and Department of Biomedical Engineering, University of Southern California,
Los Angeles, California 90089
Sigfrid D. Soli
House Ear Institute, 2100 West Third Street, Los Angeles, California 90057
?Received 8 May 1997; revised 21 January 1998; accepted 31 March 1998?
Recent studies have shown that temporal waveform envelope cues can provide significant
information for English speech recognition. This study investigated the use of temporal envelope
cues in a tonal language: Mandarin Chinese. In this study, the speech was divided into several
frequency analysis bands; the amplitude envelope was extracted from each band by half-wave
rectification and low-pass filtering and was used to modulate a noise of the same bandwidth as the
analysis band. These manipulations preserved temporal and amplitude cues in each frequency band,
but removed the spectral detail within each band. Chinese vowels, consonants, tones and sentences
were identified by 12 native Chinese-speaking listeners with 1, 2, 3, and 4 noise bands. The results
showed that the recognition score of vowels, consonants, and sentences increased monotonically
with the number of bands, a pattern similar to that observed in English speech recognition. In
contrast, tones were consistently recognized at about 80% correct level, independent of the number
of bands. This high level of tone recognition produced a significant difference in the open-set
sentence recognition between Chinese ?11.0%? and English ?2.9%? for the one-band condition
where no spectral information was available. The data also revealed that, with primarily temporal
cues, the falling–rising tone ?tone 3? and the falling tone ?tone 4? were more easily recognized than
the flat tone ?tone 1? and the rising tone ?tone 2?. This differential pattern in tone recognition resulted
in a similar pattern in word recognition: words having either tone 3 or 4 were more likely to be
recognized while words having tone 1 and 2 were not. The quantitative role of tones in Chinese
speech recognition was further explored using a power-function model and found to play a
significant role in relating phoneme recognition to sentence recognition.
Society of America. ?S0001-4966?98?02607-1?
© 1998 Acoustical
PACS numbers: 43.71.Es, 43.71.Hw ?WS?
INTRODUCTION
Traditionally, spectral information in speech sounds has
been regarded as the most important cue for speech recogni-
tion, while the temporal waveform envelope of the speech
sounds has been considered largely a redundant cue. The
view that the temporal envelope plays an insignificant role in
speech recognition may originate in Licklider’s classic ex-
periment, in which speech sounds were infinitely clipped in
amplitude, resulting in a flat temporal envelope, but still
maintained a high degree of intelligibility ?Licklider and Pol-
lack, 1948?. However, the importance of temporal informa-
tion has not always been neglected; for example, Fletcher
?1995? attributed high intelligibility of whispered speech
sounds, at least partially, to their ‘‘manner of starting and
stopping.’’ More recent studies have found that, in the ab-
sence of spectral cues, temporal envelope cues alone can
produce significant consonant recognition ?Rosen, 1992; Van
Tasell et al., 1987?. Shannon et al. ?1995? systematically
studied the trade-off between the spectral and temporal en-
velope cues. In their study, temporal envelopes of speech
sounds were extracted from one to four broad frequency
bands, and used to modulate noises of the same bandwidths.
This manipulation preserved temporal envelope cues in each
band but restricted the listener to severely degraded spectral
information. Shannon et al. found that identification of con-
sonants, vowels, and sentences improved monotonically as
the number of bands was increased, and near perfect speech
recognition was achieved with only four bands of modulated
noise.
a?Electronic mail: qfu@hei.org
505505 J. Acoust. Soc. Am. 104 (1), July 19980001-4966/98/104(1)/505/6/$15.00© 1998 Acoustical Society of America

Page 2

The present study extends the Shannon et al. study to a
tonal language, i.e., Mandarin Chinese speech. Tones are im-
portant in Chinese speech recognition because the tonality of
a monosyllable is lexically meaningful ?e.g., Liang, 1963;
Wang, 1989; Lin, 1988?. There are four tonal patterns in
Mandarin Chinese, which are characterized by their funda-
mental frequency (F0) contours. Tone 1 has a flat F0 pat-
tern, tone 2 has a rising F0 pattern, tone 3 has a falling–
rising F0 pattern, and tone 4 has a falling F0 pattern. For
example, the same syllable /ma/ can mean ‘‘mother,’’
‘‘linen,’’ ‘‘horse,’’ or ‘‘scold’’ for the tone pattern 1, 2, 3, or
4, respectively.
Although the F0 pattern is the dominant cue for tone
recognition, other acoustic cues can contribute to tone recog-
nition. For example, Liang ?1963? found that 94.6% correct
tone recognition can be achieved with the speech high-pass
filtered at 300 Hz. He argued that this high level of tone
recognition in the high-pass filtered speech is due to the resi-
due pitch, extracted from the harmonic information and
termed as the ‘‘phenomenon of the missing fundamental’’
?Schouten et al., 1962?. Liang also found that 64.0% tone
recognition can be achieved in whispered speech in which
neither fundamental frequency nor the harmonic fine struc-
ture was present. The whispered speech results indicated that
the temporal envelope could also encode information for
tone recognition. However, other studies found that tonal
contrasts were not well preserved in whispered speech
?Abramson, 1972; Howie, 1976?. Whalen and Xu ?1992?
used signal-correlated-noise stimuli ?Schroeder, 1968? to fur-
ther study the contribution of both amplitude contour and
duration to tone recognition. To produce signal-correlated-
noise stimuli, a speech signal is digitized and the sign of
approximately half of the samples, chosen at random, is
flipped, resulting in a new stimulus that has a flat spectrum
but exactly the same temporal envelope as the original
speech signal. The original F0 pattern, spectral fine struc-
ture, and the spectral envelope are completely absent in the
signal-correlated-noise stimulus. Under these conditions,
Whalen and Xu found that about 70% recognition level of
tones can be achieved. These results clearly demonstrated
that the F0 pattern is not the only cue in tone recognition;
other acoustic cues, including amplitude contour and dura-
tion, can also play a significant role.
The major goal of this study was to investigate the use
of temporal envelope cues in a tonal language, specifically in
Mandarin Chinese at word and sentence levels. The same
approach as in the Shannon et al. study was used to manipu-
late the relative distribution of temporal and spectral infor-
mation. In the present study, the amount of spectral informa-
tion in speech was increased systematically by changing the
number of bands from one, two, and three to four. The
amount of temporal envelope information in speech was ma-
nipulated by changing the cutoff frequency of envelope ex-
traction filters. Rosen ?1992? defined three main temporal
features: envelope ?2–50 Hz?, periodicity ?50–500 Hz?, and
fine-structure ?500–10 000 Hz?. Two low-pass filters with
cutoff frequencies of 50 and 500 Hz were used for envelope
extraction to evaluate the relative contribution of both tem-
poral envelope and periodicity information to speech recog-
nition. Recognition of Chinese vowels, consonants, and sen-
tences was measured in 12 native Chinese-speaking listeners
as a function of the number of noise bands. In addition,
recognition of four tones in standard Chinese was measured
and correlated to recognition of Chinese sentence recogni-
tion.
I. METHODS
A. Subjects
Twelve native Chinese-speaking listeners, including
seven men and five women, ranging in age from 25 to 35
years old, participated in this study. All listeners were re-
cruited from the University of Southern California and were
paid for their services. All listeners had pure-tone thresholds
better than 15 dB HL at octave frequencies from 250 to 4000
Hz in both ears.
B. Stimuli
The tape-recorded test data were derived from the ‘‘Chi-
nese minimal auditory capability test’’ developed by Beijing
Union Hospital in P. R. China ?Zhang et al., 1988?. The test
stimuli were spoken by an adult male speaker. The test ma-
terials included 21 initial consonants, 35 final vowels, 4
tones, and 200 daily-life sentences. The letter and its associ-
ate IPA for all 21 initial consonants and 35 final vowels used
in the present study are listed in Appendix A. All materials
were divided into four test groups with each containing
vowel, consonant, tone, and sentence recognition. Similar to
the study of Shannon et al. ?1995?, the tape-played sound
was digitized at a 10-kHz sampling rate and passed through
a pre-emphasis filter to whiten the spectrum ?low pass below
1200 Hz, ?6 dB/octave?. Then the signal was split into sev-
eral analysis frequency bands ?24 dB/octave, elliptical band-
pass filter? and the amplitude envelope from each band was
extracted by half-wave rectification and low-pass filtering
?elliptical IIR filters with cutoff frequencies of 50 and 500
Hz, ?6 dB/octave?. The speech envelope was used to ampli-
tude modulate a wide-band white noise, which was then
spectrally limited by the same bandpass filter as in the origi-
nal analysis band. These manipulations preserved band-
specific temporal envelope cues, but removed totally the
spectral details within each band. The resulting modulated
noises from each band were summed, amplified ?CROWN
D75?, and then presented to the listeners through TDH-49
headphones. The overall levels were calibrated for each com-
bination of parameters to produce an average A-weighted
output level of 75 dB for continuous speech. All these ma-
nipulations were implemented on a real-time signal process-
ing system. In the present study, the one, two, three, or four
analysis bands, each combined with the 50- or 500-Hz enve-
lope filters, produced a total of eight conditions. The total
bandwidth was from 100 to 4000 Hz. The corner frequencies
for the one-band processor were 100 and 4000 Hz. The cor-
ner frequencies for the two-band processor were 100, 1500,
and 4000 Hz. The corner frequencies for the three-band pro-
cessor were 100, 800, 1500, and 4000 Hz. The corner fre-
quencies of the four-band processor were 100, 800, 1500,
2500, and 4000 Hz.
506506J. Acoust. Soc. Am., Vol. 104, No. 1, July 1998Fu et al.: Importance of tonal envelope cues

Page 3

C. Procedures
For the consonant, vowel, and tone tests, a four-
alternative, forced-choice procedure was used in which both
the pinyin1and its corresponding Chinese character were
shown on the choice list. One test session consisted of 4
blocks, each of which had 10 trials for consonant, vowel,
tone recognition, and 20 trials for sentence recognition. A
sample block is shown in Appendix B. The stimulus in con-
sonant, vowel and tone tests was a single syllable, consisting
of an initial consonant and a following vowel with a tone. In
each trial of the consonant test, four-alternative syllables,
which had the same following vowel and tone, were included
in the choice list. The listener was asked to identify the con-
sonant by marking the syllable containing the consonant that
was presented. Because not all combinations of the initial
consonant and the following vowel were lexically meaning-
ful, the vowel and tone varied from trial to trial to accom-
modate the meaning of Chinese words. The same was true
for consonants and tones in the vowel recognition test, and
for consonants and vowels in the tone recognition test. In the
daily-life ‘‘open-set’’ sentence recognition test, the listener
was asked to write down as many words as were recognized
in each sentence. Each test block included 20 sentences, and
each sentence had 2–8 key words, resulting in a total of 100
key words. Sentences were presented without repetition. The
recognition score was calculated based on the percentage of
the total number of key words correctly recognized. All sub-
jects received extensive training, including familiarization
with the testing environment and the speech quality of all
eight experimental conditions via 15-min casual conversa-
tion with experimenter through the real-time processing sys-
tem, and informal tests of sample materials. The sample ma-
terials were not included in the formal tests. The sequence of
these eight experimental conditions were randomized and
counterbalanced across subjects. No feedback was provided
regarding the correct answer in any test.
II. RESULTS AND DISCUSSION
Figure 1 shows the averaged recognition results from 12
subjects as a function of the number of noise bands. Results
from four different tasks, including consonant, vowel, tone,
and sentence recognition, are presented in panel A, B, C, and
D, respectively. The two temporal envelope filter cutoff fre-
quencies, 50 and 500 Hz, are represented by the short dashed
line and the solid line, respectively. As a comparison, panels
A, B, and D also show our previous results of English vowel,
consonant, and sentence recognition obtained with the enve-
lope filter frequency at 500 Hz ?the long dashed line, Shan-
non et al., 1995?. Figure 1A shows that, as the number of
bands was increased from one to four, the recognition score
of Chinese consonants increased monotonically from 50.1%
to 84.7% for the 500-Hz envelope filtering condition and
increased from 45.0% to 79.8% for the 50-Hz filtering con-
dition. A two-way, repeated-measures analysis of variance
?ANOVA? was performed, with the number of bands and the
low-pass envelope filters as within-subjects factors. This
analysis showed a significant effect of the number of bands
?F(3,57)?19.45, p?0.001] but no significant difference be-
tween the 500- and 50-Hz condition ?F(1,57)?1.34, p
?0.05?.
Figure 1B shows Chinese vowel recognition as a func-
tion of the number of bands. The recognition score of Chi-
nese vowels increased monotonically from 39.4% in the one-
band condition to 85.6% in the four-band condition for the
500-Hz condition and increased from 29.6% to 79.9% for the
50-Hz condition. A similar two-way ANOVA showed a sig-
nificant effect of the number of bands ?F(3,57)?35.09, p
?0.001? but no significant effect of the envelope filters
?F(1,57)?3.84, p?0.05?.
Figure 1C shows Chinese tone recognition as a function
of the number of bands. In contrast to the Chinese consonant
and vowel results, the two-way ANOVA showed no signifi-
cant effect of the number of bands ?F(3,57)?2.15, p
?0.05? but a significant effect of the envelope filters
?F(1,57)?26.54, p?0.001?. The percent correct score of
tone recognition averaged across four band conditions was
80.8% in the 500-Hz filtering condition and 66.9% in the
50-Hz filtering condition.
Figure 1D shows Chinese sentence recognition as a
function of the number of bands. Similar to consonant and
vowel recognition patterns, the recognition score of Chinese
sentences increased monotonically from 11.0% in the one-
band condition to 90.4% in the four-band condition for the
500-Hz condition and increased from 3.8% to 74.0% accord-
ingly for the 50-Hz condition. The two-way ANOVA
showed both significant effects of the number of bands
?F(3,55)?17.08,
p?0.001?
?F(1,55)?138.49, p?0.001?.
The present results showed several interesting similari-
ties as well as differences between Chinese and English
speech recognition with primarily temporal cues. First, both
Chinese and English consonant, vowel, and sentence recog-
nition improved monotonically as a function of the number
of bands. No significant effect was observed between the
and theenvelope filters
FIG. 1. Speech recognition as a function of the number of noise bands. A:
consonants; B: vowels; C: tones; D: sentences. The conditions included
English with the envelope filter frequency at 500 Hz ???, Chinese with the
envelope filter frequency at 500 Hz ??? and at 50 Hz ???.
507507J. Acoust. Soc. Am., Vol. 104, No. 1, July 1998Fu et al.: Importance of tonal envelope cues

Page 4

50-Hz and 500-Hz cutoff frequencies of the envelope filter in
Chinese consonant and vowel recognition, similar to English
consonant and vowel recognition ?Shannon et al., 1995?.
Second, different from consonant, vowel, and sentence re-
sults, tones were consistently recognized at an 80.8% correct
level in the 500-Hz filtering condition and 66.9% correct
level in the 50-Hz filtering condition, independent of the
number of bands. Third, a significant effect of the envelope
filters was observed in Chinese sentence recognition but not
in English sentence recognition. In particular, for the one-
band, 500-Hz filtering condition, where no spectral informa-
tion was available, Chinese produced a significantly higher
score in sentence recognition ?11.0%? than English ?2.9%?
?t(17)?4.71, p?0.001?.
The similar effect of the envelope filter cutoff frequency
on Chinese tone and sentence recognition suggested a possi-
bly important role for tonal envelope cues in Chinese speech
recognition. Two approaches were used to address the pos-
sible relationship between Chinese tone and sentence recog-
nition. First, a qualitative relationship between tone recogni-
tion and sentence recognition was investigated based on the
results from the 500-Hz condition. Results showed that dif-
ferent tones had different recognition scores. As shown in the
solid line of Fig. 2?A?, the recognition scores of the falling-
rising tone ?tone 3? and the falling tone ?tone 4? were almost
twice as high as those of the flat tone ?tone 1? and the rising
tone ?tone 2? in the one-band condition. This result was con-
sistent with an earlier finding ?Whalen and Xu, 1992? and
might be explained by the differences in the amplitude con-
tour ?Fu et al., 1995?. Our acoustic analysis showed that the
amplitude envelope was highly correlated with F0 contour
for the falling–rising tone and falling tone, and this correla-
tion was likely responsible for the high recognition score of
these two tones. However, tone 1 and tone 2 did not seem to
have amplitude envelopes that were highly correlated to their
respective F0 contours. If tone recognition had played an
important role in sentence recognition, then words with tone
3 and tone 4 would be more easily recognized than words
with tone 1 and tone 2. The dashed line in Fig. 2?A? shows
the distribution patterns of tones ?right y axis? for correctly
recognized words in sentence recognition for the one-band
condition. Indeed, most of the words that were correctly rec-
ognized in sentence recognition had either tone 3 or tone 4.
Figure 2?B? shows the recognition score of the individual
tones and the distribution tonal patterns of the correctly rec-
ognized words in sentence recognition for the four-band con-
dition. Although the difference was smaller between tones
for the four-band condition than the one-band condition, the
recognition scores of tone 3 and tone 4 were still signifi-
cantly higher than those of tone 1 and tone 2, resulting in a
similar tonal distribution pattern of the correctly recognized
words. These results indicated that tone played a greater role
in sentence recognition when no spectral information was
available, and became less important when more spectral in-
formation was available.
A quantitative relationship among Chinese vowel, con-
sonant, tone, and sentence recognition scores was also as-
sessed using a power-function model ?Boothroyd and Nit-
trouer, 1988; Rabinowitz et al., 1992?. The power-function
model, based on the original work by Fletcher ?1995?, could
account for the benefit of sentence context ?factor k? and the
relation between word and phoneme recognition ?factor j?.
First, the sentence context factor is represented in the follow-
ing equation:
ps?1??1?pw?k,
?1?
where psis the recognition probability for words in sen-
tences, pwis the recognition probability for isolated words,
and the exponent k is the sentence context factor.
Second, the recognition probability for an isolated CVC
word is represented as follows:
pw?pp
j,
?2?
where ppis the recognition probability for the individual
phonemes and pwis the recognition probability for the iso-
lated words, j has a value between 2 and 3, reflecting the fact
that, due to language constraints, only 2–3 phonemes must
be recognized for correct recognition of isolated CVC words.
FIG. 2. ?A? Percent correct of tone recognition in the one-band condition
???. The tonal distribution pattern of the correctly recognized words in
sentence recognition for the one-band condition ???. ?B? Percent correct of
tone recognition in the four-band condition ???. The tonal distribution pat-
tern of the correctly recognized words in sentence recognition for the four-
band condition ???.
508508 J. Acoust. Soc. Am., Vol. 104, No. 1, July 1998Fu et al.: Importance of tonal envelope cues

Page 5

Finally, we combine Eqs. ?1? and ?2? to obtain the rela-
tion between the recognition probability for words in sen-
tences and phonemes:
ps?1??1?pp
j?k.
?3?
Because no comparable results were available in Chi-
nese materials, the present study used the same k(4.5) and
j(2.67) values as in English materials ?Rabinowitz et al.,
1992?. Different from the previous study, the recognition
probability for phonemes was calculated by combing the rec-
ognition probability for vowels, consonants, and tones:
pp?pc
wcpv
wvpt
wt,
?4?
where pc, pv, and ptare the recognition probability of con-
sonants, vowels, and tones, and wc, wv, and wtare weight-
ing parameters. Figure 3 shows the best fitting power-
function
?thesolid line?
in
probability of phonemes to the recognition probability for
words in sentences2(r?0.92). The weights for consonants,
vowels, and tones were 0.64 (s.e.?0.13), 0.71 (s.e.?0.13),
and 0.79 (s.e.?0.18), respectively. A one-way ANOVA
showed no significant difference between the weights
?F(2,96)?0.26, p?0.05?. These results suggested an ap-
proximately equal contribution of consonants, vowels, and
tones to Chinese sentence recognition under the present con-
ditions.
relatingthe recognition
III. CONCLUSIONS
High-level performance of Mandarin Chinese speech
recognition was achieved with primarily temporal envelope
cues, extending previous results obtained with English
speech. Due to the tonal nature of Chinese speech, several
significant differences were also observed between English
and Chinese speech recognition. In contrast to the recogni-
tion of consonants, vowels, and sentences, which increased
monotonically as a function of the number of bands, tones
were recognized at an 80% correct level in the 500-Hz en-
velope filter condition, independent of the number of bands.
This high level of tone recognition produced a significant
difference in the open-set sentence recognition between Chi-
nese ?11.0%? and English ?2.9%? for the one-band condition
where no spectral information was available. The analysis of
tone recognition and the tone distribution pattern for words
in sentence recognition indicated a high correlation between
tone recognition and sentence recognition. The present study
also used a power-function model to reveal that, with tem-
poral envelope cues, tones contribute to Chinese speech rec-
ognition with temporal envelope cues in a similar way to
consonants and vowels.
ACKNOWLEDGMENTS
Portions of this paper were presented at the 129th Meet-
ing of the Acoustical Society of America. We are grateful to
all subjects for their participation in our experiments and to
Alena Wilson for editing the manuscript. We also thank Dr.
Winifred Strange, Dr. Adrian Fourcin, Dr. Xu Yi, and an
anonymous reviewer for their helpful comments. Research
was supported in part by NIH ?DC-02267 and DC-01526?.
APPENDIX A: A LIST OF CONSONANTS AND
VOWELS
A Chinese single syllable consists of an initial consonant
and a following vowel with a tone. There are 21 initial con-
sonants and 35 final vowels used in the present study. The
letter and its associated IPA symbol are shown as follows:
?1? 21 initial consonants
b?p?, p?ph?, m?m?, f?f?, d?t?, t?th?, n?n?, l?l?, g?k?, k?kh?,
h?x?, j?ti?, q?tih], x?i?, zh?t\?, ch?t\h], sh?\?, r?_?, z?ts?,
c?tsh?, s?s?
?2? 35 final vowels
?a? 6 simple vowels
a?a?, o?o?, e?$?, i?i?, u?u?, u ¨?y?
?b? 13 complex vowels
ai?a(?, ei?e(?, ao?a*?, ou?o*?, ia?ia?, ie?i}?, iao?iÄ*?,
iou?i.*?, ua?ua?, uo?uo?, uai?ua(?, uei?ue(?, u ¨e?y}?
?c? 16 compound nasal vowels
an?an?, en?.n?, ang?ÄG?, eng?.G?, ong?oG?, ian?i}n?,
in?in?, iang?iÄG?, ing?iG?, ion?i*G?, uan?uan?, uen?u.n?,
uang?uÄG?, ueng?*.G?, u ¨an?y}n?, u ¨n?yn?.
FIG. 3. The relationship between sentence and phoneme recognition based
on a power-function model. Symbols showed experimental data and the
solid line represented the predicted relationship between the recognition
probability of phonemes and sentences.
509509J. Acoust. Soc. Am., Vol. 104, No. 1, July 1998Fu et al.: Importance of tonal envelope cues