Vowel reduction is a well-established phenomenon that has found its place in phonetics textbooks (e.g., Clark and Yallop, 1990; O'Shaughnessy, 1987). Briefly summarized, vowels are pronounced more "sloppily" and with less distinction when speaking style is informal, or when the vowels are part of unstressed syllables (Koopmans-van Beinum, 1980). Essentially, vowels become more centralized and/or more like the phonemes that surround them. Although there is an ongoing debate about the details, vowel reduction is generally considered to be a universal phenomenon of speech (Van Bergem, 1995).
There have been studies that investigated acoustic and articulatory consonant reduction in relation to the corresponding vowel reduction, but these were generally limited to only a few classes of consonants, with only limited speech material (e.g., Byrd, 1994; Duez, 1995; Farnetani, 1995; Keating et al., 1994; Schmidt and Flege, 1995). From these studies it is difficult to discern the general effects of consonant reduction in "normal" speech situations.
To study how consonants reduce acoustically, we decided to contrast speech from reading aloud with that of "spontaneous" story telling. It is known that vowels spoken informally or spontaneously are severely reduced with respect to vowels that are read aloud from text. Consonant reduction too can be expected to show itself when informal speech is compared with read speech.
At the moment, any understanding of the way reduction affects the spectro-temporal structure of consonants and the way it influences consonant identification is seriously lacking. Therefore, it is difficult to point to specific features of articulation where reduction will affect the phonemic distinction of consonants. In this paper, we will limit ourselves to an inventory of consonant acoustics that parallel the vowel characteristics that are affected by vowel reduction. One important question that we want to answer is whether acoustic consonant reduction is indeed similar to vowel reduction.
Table 1. Dutch consonants used in this paper. Columns: Place of articulation,
Rows: Manner of articulation (Plosives, Fricatives, Nasals and Vowel-like).
Table 2. Number of matched VCV pairs per consonant (ignoring voicing).
Four aspects of vowels and consonants are studied to characterise consonant
reduction:
Velar
Pal
Alve
Lab
Total
Plos
63
65
61
189
Fric
77
3
63
75
218
Nasal
14
72
63
149
V-like
60
21
94
60
235
Total
214
24
294
259
791
1. Formant values
2. Duration
3. Center of Gravity of the spectrum (i.e., the "mean" frequency)
4. Sound energy difference between vowels and consonants
To be able to compare realizations across both speaking styles, we will ignore the ultimate form of consonant reduction, i.e., complete deletion.
For this study we used speech material of an experienced newscaster who first told some stories and anecdotes to an interviewer (who he knew quite well). This speech was transliterated and after some time he was asked to read aloud the transcription. This way, we obtained 2 times 20 minutes of speech (spontaneous and read). The whole orthographic script was transcribed to phonetic symbols by the Grapheme-to-Phoneme conversion module of an experimental speech synthesizer developed at the Department of Language and Speech at the University of Nijmegen. One of the authors checked the transcription and marked words for sentence accent by listening. All speech was sampled with 16 bit precision and 48 kHz sampling rate.
From the phonetic transcription, all Vowel-Consonant-Vowel (VCV) segments were located in the speech recordings (also those crossing word boundaries). In total, 4814 pairs of VCV realizations from corresponding positions in the read and spontaneous utterances were identified. For 1847 VCV of these pairs, both members were realized with identical syllable structure, syllable boundary type, and sentence and word stress. Of these 1847 VCV-pairs, a random subset of 791 pairs has been analyzed in detail for this paper (see table 1 and 2) and will be used here to study consonant reduction in more detail.
Phoneme boundaries were placed using a waveform display with audio feedback (Boersma) combined with synchronized displays of the Harmonicity-to-noise ratio, total energy, and the spectral balance, i.e., energy in the high- (above 3 kHz) versus low- (below 750 Hz), high- versus mid- (between 750 and 3000 Hz), and mid- versus low-frequencies. In cases where none of the displays suggested a boundary, audio cues were used exclusively. The boundaries between vowels and consonants were placed preferably on waveform zero-crossings that corresponded to "visible" changes in the spectral composition of the waveform. If present, priority was given to spectral changes that indicated the start or end of a constriction (e.g., abrupt changes in the spectral balance). LPC formant tracks were extracted using the Split-Levinson algorithm (after down sampling to 10 kHz, using 5 pole zero pairs).
Vowel reduction is characterized by a centralization of the distribution of steady-state values in the F1/F2 plane. The vowels from the spontaneous VCV segments used in this study show such a centralization with respect to those from read VCV segments (Figure 1, see also an independent analysis of the same speech, Koopmans-van Beinum, 1992).
The formant transitions in the vowel off- and onset bordering a consonant, especially of the F2, are both sensitive to coarticulation and are important cues for consonant identification (e.g., Clark and Yallop, 1990; O'Shaughnessy, 1987). To quantify the extent of acoustic coarticulation we determined the difference between the F2 slopes at the CV- and the VC-boundaries (i.e., the F2 slope difference). We used formant track slopes normalized for vowel duration because formant track shapes are largely invariant with speaking rate (Pols and Van Son, 1993) and because in perception one also normalizes for speaking rate (Miller and Bear, 1983). The slopes were calculated from the coefficients of a 4th order polynomial fit of the F2 tracks of the vowels with the duration normalized to 1.
For the fricatives and plosives, as well as for all consonants pooled (not shown), there is a statistically significant difference between speaking styles with respect to the values of the F2 slope difference (p <= 0.001, two tailed Sign test). The F2 slope difference between CV- and VC-boundaries is generally lower in spontaneous speech than in read speech. The behaviour of individual phonemes is very erratic (Figure 2, none reaches statistical significance).
Another way to look at the "formant" values of consonants is to use the
correlation between the F2 frequency at the vowel onset and the vowel "target",
the so called F2 Locus Equations (Sussman et al., 1991, 1993, 1995). The vowel
target is defined here as the position in the vowel realization where the
formant frequencies reach their most extreme values. Examples of these Locus
Equations and the way they are calculated are given for both the read and
spontaneous realizations of the labial plosives (/
/ and /
/, Figure 3.a) and nasals (/
/, Figure 3.b) .
/ or /
/. The equations of the regression lines through the data points are the F2
Locus Equations. The equations and correlation coefficients (R) of the
regression lines are given in the lower parts of the graphs (p <= 0.001).
The correlation between F2-onset and F2-target frequencies of the following
vowel is generally quite strong and consonant specific. For the consonants /
/ the correlation explains more than half the variance for both speaking rates
and stressed and unstressed realizations alike (i.e., R >= 0.72, ignoring
voicing contrasts, e.g., /p/ and /b/ are combined). For half the consonants
examined, the correlation explains more than 75% of the variance (i.e., R >=
0.87). Furthermore, there is a robust (cor-)relation between the Y-axis
intercept value and the slope of the Locus equations which is linked to place
of articulation (Sussman et al., 1995).
For our data, we plotted the values of the Y-intercept and the slopes of the Locus Equations in Figure 4, split on speaking style and vowel stress. We selected only realizations that were not word-final and, furthermore, only those consonants and conditions for which the correlation coefficient between F2-onset and -target, i.e., the Locus Equations, had a level of statistically significance of p <= 0.05 (two-tailed). This gives us values that are reasonable stable. We must stress that we do not consider this a convincing level of significance to decide whether these correlations are genuine. The total number of coefficients calculated is too large for that. We only consider correlations convincing when a level of significance of p <= 0.001 (two tailed) is reached (the four points with p > 0.001 are indicated in the legend of Figure 4). It can be seen in Figure 4 that the correlation between the Y-intercept and the slope of the Locus equations is very strong indeed, R = 0.909, i.e., it explains 82% of the variance. The consonants seem to cluster according to place of articulation as was found by Sussman et al. (1995).
/. The equations of the regression lines through the data points are the F2
Locus Equations. The equations and correlations coefficients (R) of the
regression lines are given in the lower parts of the graphs
It can be seen in Figure 4 that there is no obvious systematic difference with
respect to the Locus equations between stressed and unstressed realizations or
between read and spontaneous realizations. For the /td/, /pb/, /fv/, and /
/ there were significant differences in correlation strength between stressed
and unstressed or read and spontaneous realizations (p <= 0.001). However,
although the correlation strength was larger for stressed than for unstressed
realization within the same speaking style, the correlation strength for read
speech could be larger or smaller than that for spontaneous speech, depending
on stress. All in all, consonant reduction seems to have no systematic effects
on the F2 Locus Equations.
/, the Read, Stressed realizations of the /
/, and the Spontaneous, Stressed realizations of the /
/, indicated by italic text). We already showed that spontaneously uttered vowels were markedly reduced with respect to read vowels (Figure 1). Stress too had a strong effect (not shown). The fact that the strong correlations between the vowel F2-onset and -target frequencies is preserved irrespective of stress and speaking style indicates that consonant reduction is strongly linked to vowel reduction. As the F2 "target" frequencies of the vowels change due to differences in stress or speaking style, the Locus Equations show that the F2 onset frequencies change too. Formant frequencies are markers for articulatory movements, so we can assume that articulatory vowel reduction is accompanied by correlated changes in the (preceding) consonants.
Duration is one of the strongest correlates of vowel reduction (Van Bergem, 1995; Van Son, 1993). As is to be expected, there is a decrease in vowel duration in the spontaneous members of each pair (Figure 5, pooled values, see also Koopmans-van Beinum, 1992). The consonant realizations too are shorter in spontaneous speech (Figure 5, C, pooled values). This holds for all individual consonantal categories (not all differences are statistically significant, see Figure 5), except for the vowel-like consonants where duration seems to remain constant or to increase slightly (not significant).
Both vowels and consonants become shorter when spoken spontaneously. Furthermore, they become shorter by the same amount. The relative duration of consonants in the VCV segments, i.e., as a fraction of the total duration, does not change with speaking style (not shown).
The center of gravity of a spectrum (COG) is in a sense, the "mean" frequency. It is calculated by dividing [integral]f.E(f).df by [integral]E(f).df. For sonorants, the COG is related to the spectral slope, the steeper the slope, the lower the COG. The steepness of the spectral slope, in its turn, is determined by the steepness of the glottal pulse which is a measure of speech effort. For turbulent noise, the COG is determined by the size of the quotient of (air flow speed) / (constriction area) which again is determined by speech effort.
For Dutch (and English), a more level spectral slope, i.e., a higher COG, strongly correlates with perceived sentence accent (Sluijter, 1995a, 1995b; Sluijter en Van Heuven, 1996). As the de-accentuation of vowels strongly correlates with vowel reduction, we can predict that reduction will show up as a lower COG. In Figure 6 this prediction bears out for the vowel realizations. For each vowel, spontaneous realizations have a lower COG than the read realizations (only shown for pooled data). For the sonorants and fricatives we see a similar picture (a lower COG for spontaneous realizations). For the release bursts of the plosives we see an erratic behaviour that does not seem to indicate a definite difference in the COG with respect to speaking style.
A subdivision of the phonemes in categories can be seen in Figure 6. Very high
absolute COG frequencies are found for most obstruents (plosives and
fricatives). For fricatives, the COG frequency is inversely related to the size
of the cavity in front of the noise source. For plosives the pattern is more
intricate. The COG frequencies for /
/ from spontaneous speech are indistinguishable or higher than those from read
speech (statistically not significant). The rather low COG frequencies for /pb/
(similar to that for vowels) show the influence of the open oral cavity behind
the sound source. The overall distribution of COG values of obstruents is
strongly bimodal due to the presence of approximants (not shown).
Quite low COG frequencies are found for sonorants (vowels and consonants) with vowels having higher values than nasals and vowel-like consonants. For the latter, the COG is dominated by the damping of the higher frequencies due to their closed articulation.
E
= Vmax - Cmax , and for nasals and vowel-like consonants:
E
= Vmax - Cmin.One of the most salient differences between vowels and consonants is in their respective sound energy level. Vowels generally have a much higher sound energy level than consonants. Vowel reduction decreases the maximal sound energy level of vowels. Whether the energy level of consonants changes by the same amount can be determined by measuring the sound energy, or the relative energy, of consonants with respect to their flanking vowels. The sound energy difference is measured as indicated in Figure 7.
Figure 8 displays the sound energy differences for read and spontaneous speech. For all consonants, except for the nasals, the intervocalic sound energy difference is smaller in spontaneous speech. Altogether, the effects of speaking style on the intervocalic sound energy differences seem to be small, on the order of 1 dB. Therefore, changes in the sound level of the vowels seem to be largely matched by corresponding changes in the intervocalic consonants.
Four correlates of reduction have been studied for consonants with respect to speaking style: 1) F2 slope differences and Locus Equations, 2) Duration, 3) Center of Gravity, and 4) Intervocalic sound energy difference.
In spontaneous speech, the nasal consonants "weaken" somewhat more than the neighbouring vowels whereas other consonants "weaken" somewhat less than the vowels (Figure 8).
The generally lower F2 slope differences in spontaneous speech indicate a decrease of coarticulation strength. This is equivalent to the spectral effect of articulatory reduction found in vowel space. We also found strong, consonant specific, correlations between the F2-onset and -target frequencies of vowel realizations following these consonants. These robust correlations, called F2 Locus Equations (Sussman et al., 1991, 1993, 1995), indicate that formant changes in consonants largely mirror those in vowels. The lack of any (systematic) differences between the F2 Locus Equations with respect to differences in stress or speaking style strongly suggests that the consonants also mirror any change in vowel F2 frequencies that is related to vowel reduction.
In spontaneous speech, consonant realizations shorten like vowels. The decrease in duration of consonants is such that the relative duration, as a fraction of total VCV segment duration, remains unchanged (not shown). Therefore, the change in duration seems to be a "global" feature of a change in speaking style.
Except for the plosives, all consonants and vowels showed a decrease in COG in the spontaneous speech condition. This indicates that both the vowels and the non-plosive consonants show a diminishing source strength in spontaneous speech. This in return, implies a decrease in vocal and articulatory effort. As the COG is strongly linked to the spectral slope at high frequencies, this lowering might be expected to correlate with a decrease in the perceived stress of the vowels and, if consonants contribute to stress perception, the consonants (Sluijter, 1995a, 1995b; Sluijter en Van Heuven, 1996).
When spoken in a more informal style, consonant realizations show reduction in terms of diminishing articulatory precision and global effort. Furthermore, consonant reduction resembles vowel reduction in both type and extent of the changes in the produced sounds. Details of the changes in spectral and sound energy level of consonants due to speaking style differences depend on the type of phoneme.
The authors want to thank Florien Koopmans-van Beinum for supplying the speech recordings and Noortje Blauw for her transliteration of the spontaneous speech. This research was made possible by grant 300-173-029 of the Dutch Organization of Research (NWO).