Acoustic Properties for Place of Articulation in Nasal Consonants

Acoustic properties for place of articulation in nasal consonants Kathleen Kurowski and Sheila E. Blumstein

Department of Cognitive and Linguistic Sciences, Brown University, Providence, Rhode Island 02912

(Received 1 July 1986; accepted for publication 16 February 1987)

The goal of this study was to determine whether acoustic properties could be derived for English labial and alveolar nasal consonants that remain stable across vowel contexts, speakers, and syllable positions. In experiment I, critical band analyses were conducted of five tokens each of [m] and In] followed by the vowels [i e a o u] spoken by three speakers. Comparison of the nature of the changes in the spectral patterus from the murmur to the release showed that, for labials, there was a greater change in energy in the region of Bark 5-7 relative to that of Bark 11-14, whereas, for alveolars, there was a greater change in energy from the murmur to the release in the region of Bark 11-14 relative to that of Bark 5-7. Quantitative analyses of each token indicated that over 89% of the utterances could be appropriately classified for place of articulation by comparing the proportion of energy change in these spectral regions. In experiment II, the spectral patterus of labial and alveolar nasals produced in the context of Is] + nasal ( [mn] ) + vowel ( [ie a o u] ) by two speakers were explored. The same analysis procedures were used as in experiment I. Eighty-four percent of the utterances were appropriately classified, although labial consonants were less consistently classified than in experiment I. The properties associated with nasal place of articulation found in this study are discussed in relation to those associated with place of articulation in stop consonants and are considered from the viewpoint of a more general theory of acoustic invariance.

PACS numbers: 43.70.Fq, 43.70.Aj

INTRODUCTION

Research on the acoustic cues for place of articulation in nasal consonants has investigated both the nasal murmur and formant transitions. Focusing on the nasal murmur, various acoustic analyses of nasal consonants have shown that small but systematic differences can be correlated with place of articulation. Many of these studies (House, 1957; Fant, 1960; Fujimura, 1962) have employed some form of vocal tract synthesis to determine the location of the zeros or antiformants that are said to characterize each nasal's place of articulation. Most researchers considered a single antiformant of [ m ] (around 800 Hz), [ n ] (around 1500-2000 Hz), and [ rj ] (above 3000 Hz) to be a kind of acoustic locus for place of articulation. Fujimura suggested looking at "cluster variables" of F 2 + F 3 + antiformant for [m] and F3 + F4 + antiformant for [n]. Nevertheless, results of these acoustic analyses of the murmur revealed considerable variability in the murmur itself, even within the same utterance (owing to the fact that the pole-zero pairs moved throughout the duration of the nasal murmur) (Fujimura, 1962). In addition, Fujimura noted large individual differences in his measures across speakers. Thus, while some place-of-articulation differences emerged, the results did not provide a clear-cut distinction for place of articulation in nasal consonants across various vowel environments.

In a pilot study, Blumstein and Stevens (1979) explored the acoustic characteristics of place of articulation for the nasal consonants [m n] across various vowel environments. Rather than focusing on the murmur, they investigated the gross shape of the spectrum sampled at the release of the nasal consonant, thus incorporating the early portions of the

formant transitions in their measure. Their results failed to

show clear-cut differences between labial and alveolar nasal

consonants. A more recent preliminary investigation by Repp (1986) also failed to provide systematic differences between the acoustic properties of [m] and In]. Repp focused on the spectral changes between the murmur and vo- ealic segments of nasal-vowel syllables. He determined the spectral differences between these two regions and compared derived spectral shapes for [m] and [n]. On the basis of his findings, Repp (1986, pp. 1993-1994) concluded, "There is no indication in these data of any invariant spectral change property distinguishing [ m ] and [ n] across all voca- lic contexts."

The acoustic properties contributing to the perception of nasal consonants have been investigated in early studies by focusing on the role of the second formant transition as a place cue. Results indicated that the starting frequency of the second formant transition provides a cue to the distinction between [m] and [n] (Liberman et al., 1954; Nakata, 1959; Hecker, 1962). In many of these studies, a neutral murmur was appended to the moving formant transitions (Cooper et al., 1952; Liberman et al., 1954; Miller, 1977; Miller and Eimas, 1977; Larkey et al., 1978). Because listeners were able to identify place of articulation in such stimuli, many researchers have considered the nasal murmur to be predominantly a manner cue, not a potential place cue (De- lattre, 1968; Pickett, 1965; Mermelstein, 1977). Neverthe- less, in tape-splicing experiments with natu.ral speech, Ma16- cot (1956) showed that, although inappropriate transitions in CV and VC syllables overrode the conflicting murmur cue, the murmurs were not perceptually neutral with respect

1917 J. Acoust. Sec. Am. 81 (6), June 1987 0001-4966/87/061917-11500.80 ¸ 1987 Acoustical Society of America 1917

Downloaded 19 Dec 2010 to 141.212.110.249. Redistribution subject to ASA license or copyright; see http://asadl.org/journals/doc/ASALIB-home/info/terms.jsp

to place of articulation. They detracted from perceptual sa- lienee of transitions mismatehed with them, and they were not judged randomly when presented in isolation. In a repli- cation of Malrcot's experiment, Recasens (1983) used synthetic speech stimuli to test the cueing effectiveness of murmurs versus transitions for Catalan [n, j•, tl ] in VC syllables. Although Reeasens' results generally corroborate Malrcot's finding that the transitions dominated conflicting murmurs, this was not true in the case of [n] transitions with [rj] murnlurs.

Kurowski and Blumstein (1984) assessed the role of both the nasal murmur and formant transitions as perceptual cues for place of articulation in nasal consonants across a number of vowel environments. Through the use of com- puter-edited natural speech stimuli, they presented listeners with five types of stimuli: full murmurs, the last six glottal pulses of the murmur, the entire transition plus vowel, six pulses of the transition plus vowel pulses starting from the beginning of the formant transitions, and the six pulses sur- rounding the nasal release, encompassing the last three pulses of the murmur and the first three pulses of the transitions. Results showed that the isolated murmur provided as much information for the perception of place of articulation as did the isolated transitions. Moreover, the highest performance scores for place of articulation were obtained in the six-puise condition containing both murmur and transition information. In a recent study, Repp (1986) confirmed these findings using edited speech stimuli from a larger variety of test utterances, conditions, and speakers. Although he found place of articulation information in the murmur to be somewhat less salient than that in the formant transitions, best identification performance occurred when both murmur and transition information were presented together. Moreover, performance was better than predicted by the "addition" of two separate cues, namely one for the murmur and the other for the transitions.

On the basis of these findings, both sets of researchers suggested that the perceptual cues to place of articulation in nasal consonants reside in the relation between spectral properties of the murmur and transitions immediately following the nasal release, rather than in either the murmur or the transitions alone. Kurowski and Blumstein (1984) further suggest that the combination of nasal murmur plus formant transitions forms a single integrated property for the perception of place of articulation. In this view, the acoustic information in the vicinity of the nasal release, incorporating both murmur and transitions, provides a single unitary property for place of articulation.

The notion of integrated properties has been explored most extensively with respect to place of articulation in stop consonants. Results of this research suggest that integrated properties encompassing the burst and some tens of milliseconds of the formant transitions can be used to characterize

place of articulation in stop consonants (Stevens, 1975; Ste- vens and Blumstein, 1978). Researchers investigating this issue have used different analysis procedures and have made different theoretical claims. Nevertheless, focusing on the burst and the onset of the transitions, stable acoustic patterns have been found across speakers (Blumstein and Ste-

vens, 1979; Searle, 1979, 1980; Kewley-Port, 1983), vowel contexts (Blumstein and Stevens, 1979; Searle, 1979, 1980; Kewley-Port, 1983), and syllable position (Blumstein and Stevens, 1979).

Two approaches have been taken in this research. The first is exemplified by the work of Stevens and Blumstein ( 1978, 1981; Blumstein and Stevens, 1979). They proposed static properties based on the gross shape of the spectrum sampled at the stop release. While their analysis procedures correctly classified place of articulation for labial, alveolar, and velar stop consonants over 83% of the time, as indicated above, this procedure failed to characterize place of articulation in nasal consonants. The second approach, originally proposed by Searle ( 1979; Searle et aL, 1980) and Kewley- Port (1983) and later by Lahiri et al. (1984), postulated time-varying properties for place of articulation in stop consonants. These researchers focused, as did Stevens and Blumstein (1978, 1981; Blumstein and Stevens, 1979), on the vicinity of the stop release in elaborating these properties. However, they explored the changes in spectral properties over time. Despite different analysis procedures, their results showed consistent patterns for place of articulation in syllable-initial position across speakers and vowel contexts.

The results of these studies investigating integrated acoustic properties for place of articulation in stop consonants, as well as the perceptual results for place of articulation in nasal consonants obtained by Kurowski and Blumstein (1984} and Repp (1986), suggest that acoustic properties may be derived for place of articulation in nasal consonants based on spectral changes in the vicinity of the nasal release. Such properties would presumably encompass spectral char- aeteristies of both the murmur and the transition portions of the nasal and would be reflected in the patterns of rapid spectral change charted from the murmur into the release.

The goal of the present study was to determine whether such acoustic properties could be derived for English labial and alveolar nasal consonants. We hoped to determine what the nature of such properties might be, and to determine whether they remain stable across vowel contexts, speakers, and syllable positions. It was also of interest to consider how, if at all, the properties associated with nasal place of articulation relate to those proposed for place of articulation in stop consonants. If it is the case that invariant acoustic properties correspond to phonetic features, as has been claimed by Ste- vens and Blumstein (1978, 1981; Blumstein and Stevens, 1979 ), then the same properties should characterize place of articulation across different manners of articulation.

The decision to focus on the vicinity of the nasal release as the potential area where the acoustic properties for place of articulation are located was based largely on the results of the perception experiments of Kurowski and Blumstein (1984) and Repp (1986). Consequently, we decided to use acoustic analysis procedures that reflected, in some way, properties of the auditory system. To that end, we employed critical band analyses of our stimuli. These analyses may more nearly reflect the auditory transformations that occur at the level of the auditory nerve, and thus provide one of the first transformations of the acoustic signal as it ascends the auditory pathway.

1918 J. Acoust. Sec. Am., Vol. 81, No. 6, June 1987 K. Kurowski and S. E. Blurnstein: Place of articulation 1918


A.A A._A A A A_A

/'V'V'V'

I I I I I

•h_A A.A

I I I I I

80-

v-v v- v- I I I I I

188

28'

t,,,, i, ,,, i , ,, , i , , ,, i,,,,

FIIEg KHZ

FIG. 1. The left panels show a portion of the waveform display of the [ ne] and [ mi] tokens sampled from the murmur through the transitions into the vowel. The solid vertical line represents the cursor position corresponding to the point of discontinuity between the nasal murmur and the glottal pulse containing the beginning of the formant transitions. The left-most dotted line and the middle solid line represent the edges of the window position for the two glottal pulses corresponding to the murmur, and the middle solid line and right-most dotted line represent the edges of the window position for the two glottal pulses corresponding to the release. The right panels show the LPC spectra corresponding to portions of the waveform depicted on the left. The dotted lines represent the LPC spectra of the three murmur pulses preceding the discontinuity, the solid line represents the LPC spectrum of the glottal pulse containing the discontinuity, and the dashed lines represent the LPC spectra of the two pulses immediately following the pulse containing the discontinuity.

In our analyses, we focused on likely areas of the spectrum where previous research had indicated potential differences between [m] and [n]. In particular, analysis of data across several languages, as reported by Recasens (1983), suggested systematic differences between [ m ] and [ n ] in the frequency of the first nasal resonance. The resonance occurs typically between 200 and 800 Hz. In addition, researchers investigating natural speech production found antiformants for [m] and [n] in two regions. For [m], they were located

around 800 Hz and for [n] between 1500 and 2000 Hz (House, 1957; Fant, 1960; Fujimura, 1962).

I. EXPERIMENT I

A. Methodology I. $t/mul/

The natural speech tokens analyzed in this study were uttered by three male speakers of English. Each of the origi-

1919 J. Acoust. Soc. Am., Vol. 81, No. 6, June 1987 K. Kurowski and S. E. Blumstein: Place of articulation 1919


nal utterances consisted of the nasal consonants [ m n ] followed by the vowels [ i e a o u ]. The subjects were individual- ly tape recorded in a sound-proofed booth. They read the entire list of utterances in citation form a total of five times

each. Each reading consisted of a different random order. There was thus a total of 150 stimulus tokens. Stimuli were

digitized using a 10-kHz sampling rate with a 4.5-kHz low- pass filter setting and a 10bit quantization.

2. Procedure

An important step in the analysis procedure was the determination of the point of release or discontinuity between the nasal murmur and the transitions into the fol-

lowing vowel. The procedures used for determining the discontinuity were the same as those described by Kurowski and Blumstein (1984). The point of release of the nasal was visually identified in the waveform as a break in the pattern of murmur pulses and the beginning of high-frequency components. Figure 1 shows an example of the waveform display of tokens of [ne] and [mi], with the cursor set, as indicated by the solid line, at the beginning of a pitch period that contains the first glottal pulse showing incipient high-frequency energy. In addition to using visual inspection of the waveform display, corroborative spectral analyses were conducted using linear predictive coding (LPC) with a 10-ms full Hamming window. The glottal pulse containing the discontinuity could be recognized by a change in the spectral characteristics of the stimulus, particularly in comparison to the spectral pattern of the preceding few murmur pulses. An example of the LPC analysis of the Inc] and [mi] tokens is provided in the right panels of Fig. 1. In both examples, there is an abrupt change in the spectral shape of the release compared to the preceding murmur. Finally, perceptual testing of the stimuli was used to double check the visual results. In

particular, subjects (including the authors and other lab per- sonnel) were informally asked to listen to tokens containing the transitions into the vowel to determine whether they per- ceived a nasal or stop consonant.

We then explored the spectral changes from the murmur to the vowel onset. To that end, we focused on the spectrum of the two glottal pulses of the murmur immediately preceding the release (henceforth, by convention to be called the murmur) and the first two glottal pulses of the transitions (by convention to be called the release). The decision to use two glottal pulses as the window of analysis was based in part on the results of the perception experiments of Kur- owski and Blumstein (1984) and Repp (1986), which showed that listeners could accurately perceive nasal consonants when presented with stimuli containing as little as two to three glottal pulses of the murmur, followed by two to three glottal pulses of the formant transitions. Moreover, our own pilot work on these stimuli indicated that edited CV stimuli containing two glottal pulses preceding the release were sufficient for the perception of a shortened nasal-vowel stimulus and for correct perception of place of articulation of the nasal consonant. A full Hamming window with a length of two pitch periods was centered over each of these two landmark positions: the first two complete murmur pulses preceding the release and the first two glottal pulses that

tø'l• [mi 1

s'ø 1 9JJ \., 3 o 5 IO 15 ;>o ;',5

BARK

tO a t - -

M '

I/ ..... ..--, .... , .... , .... 0 5 I0 15 •0 25

S I0 •

I/ .... , ..... .---, .... , .... - 0 5 I0 15 20 25

BARK

,o S ' •

0 5 I0 15 20 BARK

• x./ • ...... o 5 IO 15 2o 25

BARK

ß 40 dB O 5 IO 15 ZO 25

BARK

4OdB J' . ..... 12ø O 5 IO 15 20 25

BARK

60

dB

O 5 IO 15 20 z5 BARK

40

0 5 IO 15 20 25 BARK

dB

0 5 I0 15 20 25 BARK

FIG. 2. Critical band analyses of two glottal pulses of the murmur (represented by the dotted line ) and two glottal pulses of the release (represented by the solid line ) for ten stimulus tokens from the same speaker. The abscissa represents frequency plotted in Bark. The left scale of the ordinate represents rms energy and the right scale the associated dB values.

contained the transitions. It should be noted that there were

differences in the window size as a function of the duration of

the two pitch periods. Using a Patterson filter (Patterson, 1976) and no preemphasis, critical band analyses were then conducted at these two points.

Figure I shows examples of the cursor positions for the critical band analyses. The left-most dotted line and the middle solid line represent the edges of the window position for the two glottal pulses corresponding to the murmur, and the middle solid line and right-most dotted line represent the edges of the window position for the two glottal pulses corresponding to the release. The cursors marking the window positions were always placed pitch synchronously. As a consequence, the point of release corresponding to the onset of

1920 J. Acoust. Soc. Am., VoL 81, No. 6, June 1987 K. Kurowski and S. E. Blumstoin: Place of articulation 1920


TABLE I. Frequency and bandwidth values associated with each Bark.

Bark Center frequency Bandwidth

Prop dB

24 :30

I $0 80

2 150 !00 16 24

3 250 100

4 350 100

5 450 110

6 570 120

7 700 140 8 18 8 840 150

9 1000 160

10 1170 190

11 1370 210 4 12

12 1600 240

13 1 g50 280

14 2150 320

15 2500 380 2 6 16 2900 450

17 3400 550

18 4000 700

19 4800 900 I o 20 5800 1100

21- 7000 1300 22 8500 1800

the formant transitions was either at the leading edge of the window (as shown in the bottom panel of Fig. 1 ) or contained within the glottal pulse, including a portion of the murmur (as shown in the top panel of Fig. 1 ).

B. Results and discussion

In pilot work, we visually inspected the graphic displays obtained for each stimulus, comparing the change in the spectral patterns for the murmur to those of the release. Fig- ure 2 shows an example of one of the tokens of each CV syllable. The abscissa represents frequency plotted in Bark. [Table I provides a list of the frequency values associated with each Bark (Zwicker, 1961 ). ] The left scale of the ordinate shows the rms values and the right scale shows the associated dB values (in this case, dB = 20Xlog rms). Com- parison of the nature of the changes in spectral patterns from the murmur to the release suggested that the patterns of change were different for the labial and alveolar nasal consonants, particularly in the frequency regions encompassed in Bark 5-7 (395-770 Hz) and 11-14 (1265-2310 Hz). In particular, for labials, there is a greater change in energy from the murmur to the release in the region of Bark 5-7 relative to that of Bark 11-14. In contrast, for alveolars, there is a greater change in energy from the murmur to the release in the region ofBark 11-14 as compared to Bark 5-7. As Fig. 2 shows, these different patterns for labial and alveolar consonants emerge most clearly for [a], [o], and [u]. It is worth noting that these frequency ranges correlate with the ranges reported for the expected antiformants of [m] (around 800 Hz) and [n] (1500-2000 Hz) (House, 1957; Fant, 1960; Fujimura, 1962).

We then attempted to quantify these observations in order to determine how consistently the utterance tokens actu- ally met the qualitative changes described above. To this end, we compared the proportion of change in the energy

I I I i

ß ß

O ß O

eA o •

o o •O O Oo o • •

o o •o •oo o

1

I I I I

dB 0 6 12 18 24

Prop I 2 4 8 16

FIG. 3. A plot of the proportion of rms energy change from the murmur to the release in low frequencies (Bark 5-7) relative to high frequencies (Bark 11-14) for 150 tokens of [m] and [n] in the environment of [ie a o u] spoken by three speakers. The abscissa represents the proportion of change from the murmur to the release in the lower frequency range (Bark 5-7) and the ordinate represents the proportion of change from the murmur to the release in the higher frequency range (Bark 11-14). The dB values plotted on the abscissa and ordinate correspond to the difference (in riB) between the amplitude of the two pulses of the murmur and the two pulses of the transitions in these frequency regions. The open triangle, circle, and square represent the labial tokens produced by speakers 1, 2, and 3, respec- tively, and the closed triangle, circle, and square represent the alveolar tokens produced by these three speakers.

encompassing Bark 5-7 to the proportion of change encompassing Bark 11-14 from the murmur to the release. This was accomplished in the following manner. The cursor positions defining the murmur and the release, as discussed above, were determined. Using a rectangular filter, the rms energy encompassing the frequency range of Bark 5-7 was computed at the murmur and the release, as was the rms energy encompassing the frequency range of Bark 11-14. The proportion of change was then determined by dividing the rms value of the release by the rms value of the murmur for the lower frequencies (Bark 5-7) and by dividing the rms

TABLE II. Mean percent correct classification for place of articulation for nasal-vowel stimuli.

Speaker Labial Alvenlar Total

K 84 96 90 B 88 96 92 N 92 80 86

Total 88 91 89



TABLE IlL Total misclassification of place of articulation (out of 5) across syllable types.

Speaker [ma] [mel [rail [mo] [mul [no] Inc] [nil [nol [nu]

K 4 1 B 3 1 N i i 4 i

value of the release by the rms value of the murmur for the higher frequencies (Bark 11-14).

These proportions were calculated for each of the 150 stimuli. The results were then plotted and are shown in Fig. 3. The abscissa represents the proportion ofchange from the murmur to the release in the lower frequency range (Bark 5- 7) and the .ordinate represents the proportion of change from the murmur to release in the higher frequency range (Bark 11-14). The dB values plotted on the abscissa and ordinate correspond to the difference (in dB) between the araplitude of the two pulses of the murmur and the two pulses of the transitions in these frequency ranges. A proportion of one, represented by the diagonal line, indicates no difference in spectral change between the high and low frequencies.. A value below the diagonal indicates greater change in the lower frequencies (Bark 5-7) relative to the higher frequencies (Bark 11-14), and a value above the diagonal indicates greater change in the higher frequencies (Bark 11-14) relative to the lower frequencies (Bark 5-7). As Fig. 3 shows, the labials and alveolars fall into two clear- cut distributions, with a majority of the former falling below the diagonal and a majority of the latter falling above the diagonal.

Table II shows a summary of the distribution of the labial and alveolar stimuli. A token with a proportion falling below the diagonal was classified as a labial and one falling above the diagonal was classified as an alveolar. Tokens that fell on the diagonal were scored as misclassifications. As Ta- ble II shows, application of this metric resulted in the correct classification of over 89% of the utterances spoken by the three speakers. The range of classification scores was from a low of 84% for labials produced by speaker K and a high of 96% for alveolars produced by speakers K and B.

While these results suggest that there are distinct and consistent patterns of change from the murmur to the release for labials and alveolar nasals, it is also important to examine those CV syllables that were not correctly classified by these measurement procedures. Table III shows a summary of these results across the three speakers. As is indicated, there is a preponderance ofmisclassifications of labial nasals in the environment of the vowel [i]. Of a total of 15 [ mi ] syllables, eight were miselassified. Four of these were produced by speaker K, three by speaker B, and one by speaker N. More- over, out of 15 [na] syllables, 4, spoken by the same subject, were miselassified.

To summarize the results of experiment I, there are different patterns of change in the vicinity of the nasal release for labial and aircolor consonants. For labial nasals, there seems to be a greater relative change in the lower frequency regions encompassing Bark 5-7 (395-770 Hz) as compared

to the higher frequency regions encompassing Bark 11-14 ( 1265-2310 Hz), while, in alveolar nasals, there seems to be a greater change in these higher frequencies relative to the lower ones. While the correct classification of utterances is

high at 89%, there are clearly vowel context effects, particularly for labials preceding the vowel [i].

Before considering the implications of these results in detail, we felt it important to examine whether these same patterns of spectral change, which were shown for nasals in syllable-initial position, would emerge in a different syllabic context. To this end, we explored the spectral patterns of labial and alveolar nasals in a consonant cluster environ-

ment, in particular in an [s] + nasal + vowel context.

II. EXPERIMENT II

The purpose of this experiment was to test whether the acoustic properties for place of articulation for nasals noted in the first experiment would emerge in a different phonetic context. Moreover, we wanted to determine whether the measurement procedures derived for syllable-initial nasals would appropriately classify nasal place of articulation in the face of potential coartieulation effects of the preceding [s]. The presence of such context effects in this environment has been demonstrated both acoustically and perceptually, at least for place of articulation in stop consonants (Mann and Repp, 1981; Repp and Mann, 1981, 1982).

A. Methodology

L Stimuli

Two speakers, K and B from experiment I, recorded ten syllable types consisting of [s] followed by [m] or [n] and the vowels [i e a o u ]. The array of syllable tokens was uttered five times by each speaker. These tokens were blocked and read in a different random order. There was a total of

100 stimulus tokens. The recording procedures were the same as those used in experiment I.

2. Procedure

The analysis procedures paralleled those of experiment I. As in the case of syllable-initial nasal consonants, nasals produced in the cluster environments were first examined to determine the point of release or discontinuity between the nasal murmur and the transitions. We noted that the wave-

forms of these nasal consonants were not appreciably different from those of syllable-initial nasals, except that the murmur portion tended to be generally shorter. In fact, the murmur portion for the second speaker occasionally consisted of as few as two glottal pulses.

After the location of the release was ascertained, critical band analyses were conducted, as in the previous experiment, of the last two pulses of the murmur and the first two pulses of the release.

B. Results and discussion

The displays of the critical band analyses showed patterns similar to those obtained in experiment I. Figure 4 shows an example of the critical band analysis for the syllables [area sna smo sno]. As in experimen t I, the proportion

1922 d. Acoust. Soc. Am., Vol. 81, No. 6, June 1987 K. Kurowski and S. E. Blumstein: Place of articulation 1922


M

SiO

I o

BARK

10 3

102 R M

S IO

BARK

0 5 I0 15 20 25

BARK

,60

-40

. dB

'20

I o 5 5 ,o

BARK

60

SilO]

o

,40

dB

'20

FIG. 4. Critical band analyses of two glottal pulses of the murmur (represented by the dotted line) and two glottal pulses of the release ( represented by the solid line) for four stimulus tokens produced by the same speaker. The abscissa represents frequency plotted in Bark. The left scale of the ordinate represents rms energy and the right scale the associated dB values.

of change from the murmur to the release in Bark 5-7 and 11-14 was computed for the 100 stimuli, and the results are graphically displayed in Fig. 5. As Fig. 5 shows, the labials and alveolars again fall into two well-defined areas, with a

majority of the labials falling below the diagonal and a majority of the alveolars falling above the diagonal.

Table IV shows a summary of the classification of labials and alveolars with respect to this distribution. As in the previous analysis, a token was classified as a labial if its pro- ponion values fell below the diagonal, and it was classified as an alveolar if its proportion values fell above the diagonal. As Table IV indicates, although the overall classification scores at 84% are comparable to those obtained in experiment I (89%), the individual speaker's mean totals are lower (speaker K = 87%; speaker B = 80%) than those obtained for the same speakers in experiment I (speaker K = 90%; speaker B = 92%). Classification scores remained the same or improved for the alveolars. However, the metric misclas- sifted up to 28% more labials per speaker than in experiment I.

Table V shows the distribution of misclassificat[ons across the different syllable types. All of the misclassifiea- tiens occurred in the context of front vowels, with 15 of the 16 misclassifications occurring for labial nasals before [i] and [el, and one misclassifieation occurring for the alveolar nasal before [i]'. Thus, while the misclassifications of the nasal in experiment I were mainly limited to labials in the environment of [i l, the nasals produced in cluster environments were incorrectly classified in the context of the two front vowels [i] and [el. Apart from this, the stimuli were classified correctly across the three other vowel environments.

Prop dB 24 30

16 24

8 18

4 12

2 6

i o

I I I I

dB 0 6 12 18 24

Prop I 2 4 8 I 6

FIG. 5. A plot of the proportion ofrms energy change from the murmur to the release in low frequencies (Bark 5-7) relative to high frequencies (Bark 11-14) for 100 tokens of [m] and [n] spoken in the environment of [s] + nasal consonant + vowel ( [i e a o u] ) by two speakers (see the cap- tion for Fig. 3).

III. GENERAL DISCUSSION

The results of recent research on the perception of place of articulation in nasal consonants have suggested that perceptual cues to place of articulation reside in the integration of spectral properties in the vicinity of the nasal release incorporating both the murmur and the transitions '(Kur- owski and Blumstein, 1984; Repp, 1986). The results of the current study indicate that in this region stable acoustic properties can be found for the labial and alveolar places of articulation in nasal consonants. These properties correspond to the relative change in the distribution of energy from the murmur to the release encompassing two frequency areas (Bark 5-7 and 11-14). For labials, there is a rapid increase in spectral energy in the lower frequency range relative to the higher range, and, for alveolars, there is a rapid increase in spectral energy in the higher frequency range relative to the lower frequency range. These properties generalized across three speakers in syllable-initial position, and two speakers in syllable-medial position.

TABLE IV. Mean percent correct classification for Place of articulation for [ s ] -nasal-vowel stimuli.

Speaker Labial Alveolar Total

K 78 96 87

B 60 100 80 Total 69 98 '84

1923 d. Acoust. Sec. Am., Vol. 81, No. 6, June 1987 K. Kurowski and S. E. Blumstein: Place of articulation 1923


TABLE V. Total misclassification of place of articulation (out of 5) across syllable types.

Speaker [sinai [smel [smi] [stool [smul [sna] [sne] [sail [sno] [saul

K 4 I 1

B 5 5

Nevertheless, while the overall correct classification score across both environments was fairly high at 87%, clear vowel context effects were found. In particular, classification scores were poor for labials in initial position in the context of [i l, and for labials in syllable-medial position in the environment of [i] and [el. We explored whether these context-dependent effects could have been a function of the particular cutoff values used in developing the classification procedure. However, a review of the pattern of results as shown in Figs. 3 and 5 did not suggest any adjustment of the cutoff ratio that would incorporate these tokens into their appropriate category. An alternative possibility is that the cutoff ratio is appropriate, but the window size used in the analysis procedures was incorrect. In particular, in the environment of the vowel [i], the F 2 formant transitions are very short, and may be complete in only a few milliseconds. As a consequence, the two-pitch-period window size (approxi- mately 16 ms) used in the analyses may have included not only the formant transitions but also the steady-state portions of the vowel Ill. Thus the difference between the energy change in the formant transitions for [m] and [n] may not have emerged.

Other investigations of nasal consonants have also en- countered differences between [i] and other vowel environments in exploring the spectral characteristics for nasal place of articulation. Abramson et al. ( 1981 ), in their study of velar port size in relation to vowel height (Fig. 4, p. 333), indicated large differences in synthetic speech pole-zero positions within the transitions and vowel for [na] compared to [nil. According to their frequency tables, the pole-zero pair for the vowel [ a ] occurs between 550 and 750 Hz, which is exactly in the range of Bark 5-7 in our analyses. In contrast, the pole-zero pair for Ill spreads over the frequency range encompassing Bark 5-11. Fujitours (1962) also noted in his acoustic analysis that the frequency of the antiformant for [m] is "relatively high when the consonant precedes a front vowel such as [i] and is lower when the context is a back vowel" (p. 1872). He noted similar but smaller effects for [n]. Taken together, these results suggest that vowel context effects for [i] will work against a metric that focuses in part on the frequency range encompassed in Bark 5-7, since the crucial effects of the antiformants in this vowel

environment do not fully emerge in this frequency range. How to adjust the metric to incorporate the spectral characteristics in the context of the vowel [i] remains a problem.

It is worthwhile noting that the pattern of classification by these analysis procedures is similar to that obtained in recent perception experiments (Kurowski and Blumstein, 1984; Repp, 1986) and in discriminant analyses focusing on both the murmur and transitions (Kitazawa and Doshita, 1984). In these experiments, perception scores for place of

articulation were poorest in the environment of the vowel [ i]. These results suggest that the acoustic measures are consistent with the perceptual results for place of articulation in nasal consonants. Thus the failure to find stable acoustic

properties in the environment of the vowel [i] provides a potential explanation for the poorer perceptual performance noted in this vowel environment.

Results of experiment II indicated that the same acoustic property used to characterize place of articulation in initial position generalized to a consonant cluster environment and, in particular, to nasals preceded by [s]. Nevertheless, there was an increase in misclassifications of labials com-

pared to alveolars preceding [s]. The high identification scores for [n] in this cluster environment were presumably due to the fact that both the voiceless fricative and nasal

share the same place of articulation, and thus coarticulatory effects of the preceding fricative are minimal. In contrast, the lower scores for labials in the environment of [s] are consistent With results of Mann and Repp (1981), who found coarticulatory effects of [ s] on succeeding consonants with a different place of articulation. Nevertheless, the results of experiment II indicate that, despite these effects, similar patterns emerged characterizing labial place of articulation in initial position and following [ s].

Having demonstrated that spectral patterns for nasal consonants are similar in at least two syllable positions, we have begun work on nasals in syllable-final position. We have explored these patterns for both released and unreleased labial and alveolar nasals in both tense and lax vowel

environments. The released nasals pose no problems in analysis, and preliminary results are similar to those shown for initial nasal consonants. However, in the analysis of unreleased nasal consonants, our usual landmarks in the waveform have proved to be more difficult to determine. Similar problems of analysis in locating the point of closure have been reported elsewhere for nasals (Ma16cot, 1956) and for unreleased stops (Blumstein and Stevens, 1979). The sheer gradualhesS of the transitions from vowel to full murmur in VC position, which contributes to the ditfieulty of locating closure, also seems to have serious consequences for a metric that depends on rapid energy change over a relatively short (four glottal pulse) span in the waveform. At present, using all of the same parameters for the metric as in experiments I and II, the metric can correctly classify 75 % of the labials for one speaker. It did no better than 56% for a second speaker. All alveolars were problematic. We are presently pursuing the possibility of targeting noncontiguous areas of the waveform to see whether larger energy changes of the type described here can be found.

Finally, it is worth considering how the acoustic properties derived for labial and alveolar place of articulation in



nasals relate to those proposed for the same place of articulation in stop consonants. Direct comparisons are difficult due to differences in the acoustic analysis procedures used (Blumstein and Stevens, 1979; Kewley-Port, 1983), measurement procedures developed (Blumstein and Stevens, 1979; Kewley-Port, 1983; Lahiri etal., 1984), and landmark areas chosen to explore the spectral properties (Blumstein and Stevens, 1979; Kewley-Port, 1983; Lahiri et al., 1984). Nevertheless, the patterns proposed here for place of articulation in nasal consonants are more consistent with the view

that the acoustic properties for place of articulation are time varying (Kewley-Port, 1983; Lahiri etal., 1984) rather than static in nature (Blumstein and Stevens, 1979). Moreover, descriptively, the properties seem similar: Namely, there is a greater change in spectral energy in the higher frequencies for alveolars and in the lower frequencies for labials. Such descriptions are compatible with the diffuse-rising spectral patterns for alveolars versus diffuse-falling for labials.

Nevertheless, the frequency ranges over which these patterns were derived are clearly different for nasal and stop consonants. For nasals in this study, changes were compared in the frequency ranges 395-770 Hz and 1265-2310 Hz, whereas for stops (Blumstein and Stevens, 1979), comparisons focused on the region between 1200 and 3500 Hz. It is, of course, possible that while we found stable acoustic patterns for nasal consonants in the frequency ranges reported, we might have also found consistent patterns had we focused on frequency ranges comparable to stop consonants. In fact, in some preliminary research, Repp (1986) noted less relative energy increase from the murmur into the vowel for [ma] and [mu] at 2-2.5 kHz than at 1 kHz, whereas [na] and [ nu ] tended to have fiat or rising slopes. However, he was unable to find any distinctive patterns for nasals in the context of [i]. Although he did not attempt to quantify these observations, the frequency ranges he investigated were more in keeping with those explored for stop consonants. In the current study, we did not focus our analyses on these particular frequency regions. It is certainly possible that there are reliable differences in the spectral characteristics of labial and alveolar nasal consonants other than those de-

scribed in this study. For example, it has been suggested to us • that an alternative metric based solely on the changes in the high-frequency region (Bark 11-14) seems to separate labial and alveolar tokens. Referring to Figs. 3 and 5, he observed that a 9-dB difference in the frequency regions between Bark 11-14 might serve as a cutoff value, with changes less than 9 dB in this frequency region characterizing an [m], and changes greater than 9 dB characterizing an [n]. Reanalysis of the current data in experiments I and II indicated that the metric correctly classified 77% of the labials and 96% of the alveolars in experiment I, and 80% of the labials and 78% of the alveolars in experiment II. While the data are generally not categorized as well as those proposed in this article (of. Tables II and IV), the results are clearly better than chance and probably not significantly different from our own metric.

It is clearly an empirical issue which metric would be the preferred one. A metric focusing on relative changes might be able to withstand more sources of variability, such as dif-

ferent speakers and vowel contexts, than a metric in which absolute amplitude differences are required in a specific frequency region to distinguish labials from alveolars. In fact, the data analyses are consistent with this view. While the metric based on absolute amplitude values categorized 87% of the CV data, it failed to correctly categorize place of articulation in the environment of the front vowels [i] and [e]. In particular, it incorrectly classified 57% of the labial consonants in the environment of [i] and [e]. Further, recent cluster analyses exploring perceptual similarities among the consonants of English (Soli et al., 1986) indicated that the low-frequency regions, particularly between 600-1200 Hz, seemed to be critical for distinguishing [m] and [n]. These results suggest that a metric limited to the high-frequency regions would fail to capture these perceptual characteristics. Finally, it is not clear how a metric based on absolute amplitude differences would relate to properties described for place of articulation in stop consonants. In all of the current metrics (Blumstein and Stevens, 1979; Searle et al., 1979; Kewley-Port, 1983; Lahiri etal., 1984), the properties for place of articulation are derived from a consideration of the relative distribution of energy in the spectrum. None of these metrics focuses on absolute measures in a particular frequency region.

Experiments exploring the effects of filtering on the perception of consonant contrasts in English provide a means of addressing whether the frequency regions that were focused on in the current experiments are perceptually salient. The results of Miller and Nicely (1954), for example, suggest two important frequency regions for the perception of [ m ] and [n]: 600-1200 Hz and 1000-2000 Hz. Interestingly, subjects' performance was high when presented with stimuli filtered at either frequency range. These frequency ranges are similar (although not exactly the same) to those targeted in the current study (ef. 395-770 Hz and 1265-2310 Hz). At issue here is the fact that the metric requires two frequency regions for correct classification of place of articulation, whereas listeners seem to.be able to make perceptual deci- sions on the basis of either region. It is worth noting that Miller and Nicely's stimuli included consonant productions in a single vowel environment, [a]. If a greater variety of vo, wel environments were to be investigated, we would ex- pect that a single frequency region, or, at least, the frequency ranges delineated, would no longer be sufficient for identification of place of articulation in nasal consonants. It may be the case that in some vowel environments (other than [a] ), the relative change in both frequency regions will be necessary for correct perception (cf., for example, the spectra for [me] and [ne] versus [ma] and [na] in Fig. 2). Further research will be needed to determine whether perceptual results are consistent with the data obtained in this experiment.

The focus on the particular low-frequency regions taken in the current study is consistent not only with acoustic analyses of the location of antiformants as described earlier (House, 1957; Fujimura, 1962), but also with the physio- logical results obtained by Delgutte (1980) from the dis- charge patterns of auditory nerve fibers in the cat. With respect to the latter, Delgutte has shown that a nasal conso-

1925 J. Acoust. Soc. Am., Vol. 81, No. 6, June 1987 K. K•owski and S. E. Blumstein: Place of articulation 1925


nant context contains mostly low-frequency components. Moreover, he has shown that the spectra at the release of nasals have less intense low-frequency components than the spectra of stops with the same place of articulation. Thus the auditory representation of the same formant transitions is changed by the nature of the acoustic patterns preceding the transitions. Such auditory transformations suggest that it is highly unlikely that exactly the same frequency regions will contribute to properties for place of articulation in stop and nasal consonants.

Such results provide a challenge to the strong form of a theory of acoustic invariance, which claims that acoustic properties correspond to phonetic features, and that the same properties characterize place of articulation across different manners of articulation. While Jakobson was clearly a proponent of acoustic invariance for phonetic features, he viewed acoustic properties and phonetic features in particular as relational. That is, the properties or accompanying features were not "absolute" but varied in relation to the

presence of other properties or features and the linguistic behavior of the entire sound segment (Jakobson et al., 1963). Thus, for him, the fact that there is proportionately greater change in low-frequency energy for [m] and greater change in high-frequency energy for [n], and a similar relation holds for place of articulation in stop consonants, is sufficient evidence that the labials [m b p] share a common property and the alveolars [nd t ] also share a common property.

Further research will be needed to determine whether such relational invariance is indeed the correct characteriza-

tion of phonetic features or whether the same acoustic property can define a phonetic feature across different sound classes. In order to pursue this research, two strategies will need to be taken. First, it will be necessary to continue to explore place of articulation and other phonetic features across different manners of articulation. Second, it is necessary to develop and apply comparable measurement procedures in analyzing the stimuli. Nevertheless, whether the acoustic properties corresponding to phonetic features are relational or not, the results of this study suggest that stable acoustic properties can be derived for place of articulation in nasal consonants, and the acoustic properties derived are at least descriptively similar to those obtained for place of articulation in stop consonants. These properties turn on the presence of rapid spectral change in the vicinity of the consonant release and in the nature of that change. For both labial nasals and stops, there is a rapid spectral change with greater change in lower frequency regions relative to higher regions, and for alveolar nasals and stops there is greater change in high frequencies relative to lower frequencies.

ACKNOWLEDGMENTS

This research was supported in part by Grant NS 15123 to Brown University. Many thanks to John A. ' Mertus for his technical assistance and advice, to Kenneth N. Stevens for his comments on an earlier draft of this article and for many helpful discussions, and to Bruno Repp and Chris Darwin for their thorough and thoughtful reviews of this manuscript.

IThis suggestion was made to us by Bruno Repp in his review of this original manuscript.

Abramson, A. S., Nye, P. W., Henderson, J. B., and Marshall, C. W. (1981). "Vowel height and the perception of consonantal nasality," J. Acoust. Soc. Am. 70, 329-338.

Blumstein, S. E., and Stevens, K. N. (1979). "Acoustic invariance in speech production: Evidence from measurements of the spectral characteristics of stop consonants," J. Acoust. Soc. Am. 66, 1001-1017.

Cooper, F., Delattre, P., Liberman, A., Borst, J., and Gerstman, L. (1952). "Some experiments on the perception of synthetic speech sounds," J. Acoust. Soc. Am. 24, 597-606.

Delattre, P. (1968). "From acoustic cues to distinctive features," Phone- fica 18, 198-230.

Delgutte, B. (1980). "Representation of speech-like sounds in the dis- charge patterns of auditory-nerve fibers," J. Acoust. Soe. Am, 68, 843- 857.

Fant, G. (1960). Acoustic Theory of Speech Production (Mouton, The Hague, The Netherlands).

Fujimura, O. (1962). "Analysis of nasal consonants," J. Acoust. Soc. Am. 34, 1865-1875.

Hecker, M. (1962). "Studies of nasal consonants with an articulatory speech synthesizer," J. Acoust. Soc. Am. 34, 179-188.

House, A. S. (1957). "Analog studies of nasal consonants," J. Speech Hear. Disord. 22, 190-20•.

Jakobson, R., Fant, G., and Halle, M. (1963). Preliminaries to Speech A nal- ysis (MIT, Cambridge, MA).

Kewley-Port, D. (1983). "Time-varying features as correlates of place of articulation in stop consonants," J. Acoust. Soe. Am. 73, 322-335.

Kitazawa, S., and Doshita, S. (1984). "Nasal consonant discrimination by vowel independent features," Stud. Phonol. 18, 46-58.

Kurowski, K., and Blumstein, S. E. (1984). "Perceptual integration of the murmur and formant transitions for place of articulation in nasal consonants," J. Acoust. $oc. Am. 76, 383-390.

Lahiri, A., Gewirth, L., and Blumstein, S. E. (1984). "A reconsideration of acoustic invarianee for place of articulation in diffuse stop consonants: Evidence from a cross-language study," J. Acoust. Soc. Am. 76, 391-404.

Larkey, L. S., Wald, J., and Strange, W. (1978). "Perception of synthetic nasal consonants in initial and final syllable position," Percept. Psy- chophys. 23, 299-312.

Liberman, A., Delattre, P., Cooper, F., and Gerstman, L. (1954). "The role of consonant-vowel transitions in the perception of the •top and nasal consonants," Psychol. Monogr. 68, 1-13.

Ma16cot, A. (1956). "Acoustic cues for nasal consonants: An experimental study involving a tape-splicing technique," Language 32, 274-284.

Mann, V. A., and Repp, B. H. (1981). "Influence of preceding fricative on stop consonant perception," J. Acoust. Soc. Am. 69, 548-558.

Mermelstein, P. (1977). "On detecting nasals in continuous speech," J. Acoust. Soc. Am. 61, 581-587.

Miller, G. A., and Nicely, P. (1955). "An analysis of perceptual confusions among some English consonants," J. Acoust. Soc. Am. 27, 338-352.

Miller, J. (1977). !'Nonindependence of feature processing in initial consonants," J. Speech Hear. Res. 20, 519-528.

Miller, J., and Eimas, P. (1977). "Studies on the perception of place and manner of articulation: A comparison of the labio-alveolar and nasal- stop distinctions," J. Acoust. Soc. Am. 61, 835-845.

Nakata, K. (1989). "Synthesis and perception of nasal consonants," J. Acoust. Soc. Am. 31, 661666.

Patterson, R. D. (1976). "Auditory filter shapes derived with noise stimuli," J. Acoust. Soc. Am. 59, 640-654.

Pickett, J. M. (1965). "Some acoustic cues for synthesis of the/n-d/distinction," J. Acoust. Soc. Am. 38, 474-477.

Recasens, D. (1983). "Place cues for nasal consonants with special refer- ence to Catalan," J. Acoust. Soc. Am. 73, 1346-1353.

Repp, B. ( 1986 ). "Perception of the [ m ] - [ n ] distinction in CV syllables," J. Acoust. Soc. Am. 79, 1987-1999.

Repp, B., and Mann, V. (1981). "Perceptual assessment of fricative-stop coarticulation," J. Acoust. Soc. Am. 69, 1154-1163.

Repp, B., and Mann, V. (1982). "Fricative-stop coarticulation: Acoustic and perceptual evidence," J. Acoust. Soc. Am. 71, 1562-1567.

Searle, C. I., Jacobson, J. Z., and Raymerit, S. G. (1979). "Stop consonant discrimination based on human audition," J. Acoust. Soc. Am. 65, 799- 809.

Searle, C. L., Jacobson, J. Z., and Kimberly, E. (1980). "Speech as patterns in the three-space of time and frequency," in Perception and Production of

1926 J. Acoust. Sec. Am., Vol. 81, No. 6, June 1987 K, Kurowski and S. E. Blumstoin: Place of articulation 1926


Fluent Speech, edited by R. A. Cole (Erlbaum, Hillsdale, NIL pp. 7•- 102.

Soli, S. D., Arabie, P., and Carroll, I. D. (1986). "Discrete representation of perceptual structure underlying consonant confusions," $. Acoust. Soc. Am. 79, 826-837.

Stevens, K. N. (1975). "The potential role of property detectors in the perception of consonants," in ,4uditory Analysis and Perception of Speech, edited by G. Fant and M. A. A. Tatham (Academic, London), pp. 303- 330.

Stevens, K. N? and Blumstein, S. E. (1978). "Invariant cues for place of articulation in stop consonants," J. Acoust. Soc. Am. 64, 1358-1368.

Stevens, K. N., and Blumstein, S. E. (1981). "The search for invariant acoustic correlates of phonetic features," in Perspectives on the Study of Speech, edited by P. D. Eimas and 1. Miller (Erlbaum, Hillsdale, N J), pp. 1-38.

Zwicker, E. (1961). "Subdivision of the audible frequency range into critical bands (Frequenzgruppen)," $. Acoust. Soc. Am. 33, 248.



Acoustic Properties for Place of Articulation in Nasal Consonants

Documents

Transcript of Acoustic Properties for Place of Articulation in Nasal Consonants