F0 CONTOURS OF UTTERANCES SUPERIMPOSED ON THE TEMPORAL VOICE RANGE PROFILE OF THE SPEAKER
Antti Iivonen · 1999
A set of fundamental frequency (F0) contours produced by one speaker can be superimposed on each other on the same time/frequency (log.) plot. If the number of contours is large enough, they together form a pragmatic approximation of the speaker's individual temporal range (TVRP). Single F0 contours or contour groups can be marked within the TVRP and indicate their acoustic characteristics in this way. Different utterance types and speech acts are illustrated with this method using data from two Finnish speakers. 1. TVRP AS A BACKGROUND OF SINGLE UTTERANCE OR UTTERANCE GROUP Superimposition on each other of a set of F0 contours of the same time/frequency (log.) plot (Fig.1) makes possible the display of a pragmatic approximation of t h e speaker's temporal range (TVRP). The description is better, if the database includes variable utterances representing different speech acts and emotional states. The more the single contours included in the profile, the more representative it is of the speaker's TVRP. The older notion of voice range profile with a different meaning is described in [1]. Isolated prolonged vowels (often [a]) have been used for the display of the speaker's range profile. The purpose is more to promote analysis and to reveal the speaker's resources than to study his speech behavior. The F0 contours have been used in the p r o s o d i c analysis of languages (e.g. [2]). Empirical work with F0 contours shows that individual speakers use different ranges when producing linguistically identical utterances. Therefore, it seems appropriate to combine some elements from research with those of prosodic research. A larger F0 range may be indicative of the speaker's more variable and vivacious or stronger marking of accents or both. It can be seen (Fig.1) that the accumulation of data points concentrates on the lower part of the temporal distribution area. This concentration area can be understood as the default, i.e. as the indifferent level of the speaker's intonation. It takes the form of a black declining bar. It can be assumed that those OtopsO and OvalleysO which indicate deviations from the default area, signal the most important information. It can be assumed that the more deviant the contours are compared to the default area, the more specific their communicative function. In superimposing the utterances on the same plot, i t is appropriate to apply the notion of intonation or tone unit, breath group or simply utterance. The length of the utterances and the speech rate can vary, but it is interesting to observe how they affect the form of the F0 contour. Figure 1. Background (green) consisting of 20 sentences produced by a female speaker NA. Sentence Nr. (11) marked more clearly: Minne Maija vei omenan? [Where did Maija bring the apple?] High beginning and following down step effect observable. 2. PROGRAMS UTILIZED The utterances were recorded by means of a AKG C 451E microphone and a TEAC W-850R cassette recorder. They were then digitalized by the SoundEdit 16 Version 2 (Macromedia) on the hard disk of a PowerMac 7100/66AV computer in IFF format. Single F0 contours can be obtained by the automatic F0 analysis option of commercial programs. I have applied the F0 plot option of the SoundScope program. The measurement values with 10 ms intervals are exported (copy wave to text) to another program called PsychoF0. It is also possible to add the time points indicating the segment, syllable or word boundaries to the analysis data manually and use them in the later phases of the procedure. The analysis range has been adjusted to the speaker. The autocorrelation option was applied. Rejection of all peaks with a 25% page 953 ICPhS99 San Francisco variation in adjacent pitch periods was applied. The reliability of the measurements was controlled by means of a zoomed (0-600 Hz) narrow band spectrogram and a FFT short time spectrum. When mistakes were observed on the computer screen, after-corrections were made. F0 contours were processed further by a special program called PsychoF0 (created in the FutureBasic II language). The measurement values are indicated as dots on the computer plot. The silent measurement points (occurring in unvoiced segments or during pauses) indicated as NAN15 by the SoundScope program were changed to the numerical form 0.1, which actually means their invisibility on the computer screen. The logarithmic scale according to the algorithm in [3] has been applied in order to simulate the human psychoacoustic response. As well as showing the total variation as a background, single F0 contours, comparison of two contours or comparison of contour groups can also be presented. A contour group can be shown on a background of total variation (Fig. 2Ð3). 3. FUNCTIONAL USE OF F0 CONTOURS Fundamental frequency (F0) is functionally linked to many linguistic factors [3]. The time window controlling the functional use of F0 varies from the segment to global intonation contours stretching over whole utterances. The F0contours representing different speech act classes (e.g. statements, questions, requests, answers, commands, apologies) may be grouped together and plotted by specific marking. 4. EXAMPLES OF FINNISH 4.1. Speech material Some examples of Finnish are illustrated in Figures 1Ð3. The 25 utterances were produced by a female speaker NA and a male speaker SL. The length of the utterances was 3 or 4 words. The utterances represent statements, wh-questions, answers, topicalizations, and commands. In the topicalizations, the syntactical form of the utterance implies that the first word must get a focused accent. The answers occurred in minidialogs. The following simple examples illustrate these classes: (1) Statements and answers : Maija vei omenan [Maija brought the apple.]; Kenelle Maija antoi omenan? Hn antoi sen Ollille. [To whom did Maija give the apple? She gave it to Olli.] (2) Wh-questions : Minne Maija vei omenan? [Where did Maija bring the apple?] (3) Commands: Anna Maijalle omena! [Give Maija an apple!] (4) Topicalizations: Juoman Olli joi. [It was a drink that Olli drank.] 4.2. Female speaker NA's utterances The illustrations in Fig. 2 show that wh-questions (2) are the most stable group: all 6 utterances indicate a recurrent high F0 peak at the beginning and they have the highest F0 values (a similar observation in [4]). After a high peak F0 lowers to the default level of the speaker's TVRP. One of the questions shows also a late accentuation peak. The topicalizations (4) follow the same pattern, but the initial peak is lower. After the peak, the F0 curve assumes the default level in the speaker's TVRP. In commands (3), no recurrent outstanding initial peak occurs, but one common feature can be observed: the later level of the F0 curves occupies the higher part of the distribution area. According to the results in [4], the commands more resemble questions in general. The group of statements and answers (1) is small and heterogeneous. The initial height varies greatly (probably owing to the degree of reference to the earlier utterance; cf. the anaphoric relation to the previous utterance). The later path of the F0 mainly occupies the default area of the TVRP, but some final cases get lower values compared to the default area. In one utterance a clear F0 peak occurs. The duration of NA's utterances was maximally about 1 second. 4.3. Male speaker SL's utterances The utterances of SL were about 20 % longer than those of NA, but there were cases stretching out to 1900 ms. SL also used heavier accents and hence his F0 curves show more frequent and more outstanding F0 peaks. SL's statements (1) show a moderate initial F0 peak and a later peak indicating the rhematic word. The result i s compatible with that in [4]. The F0 paths occupy the lowest level of the TVRP. SL's wh-questions (2) repeat the high intial F0 peak observed in NA's data, but the stronger accentuations later in the utterances show a different pattern. In general, F0 curves occupy SL's whole TVRP. In commands (3), the initial F0 peak is not marked as strongly as in wh-questions. Strongly accented syllables may be observed later in the commands, too. The F0 curves occupy predominantly the upper part of SL's TVRP. In top ica l i za t ions (4) high initial F0 peaks and later heavy accents occur, but the bottoms of the curves occupy the lower part of the TVRP. 5. CONCLUSIONS The experiments conducted were preliminary, but they do show clearly the usefulness of the TVRP for prosodic research. Although some differences were observed in the two speakers studied, rucurrent intra-speaker and inter-speaker features were found in the realizations of the speech act classes. The differences were influenced especially by the heavy accents used by the speaker SL. REFERENCES [1] Orlikoff, R. F., Kahane and J. C. 1996. Structure and function of the larynx. In Lass, N. (ed.), Principles of Experimental Phonetics, 112-181. St. Louis, Baltimore/ect.: Mosby. [2] Hirst, D, & Di Cristo, A. 1998. Intonation Systems. A Survey of Twenty Languages. Cambridge/ect.: Cambridge University Press. [3] Orlikoff, R.F., Baken, R.J. 1993. Clinical Speech and Voice Measurement Laboratory Exercises. San Diego: Singular Publishing Group. [4] Hirvonen, P. 1970. Finnish and English Communicative Intonation. Publications of the Department of Phonetics, University of Turku. page 954 ICPhS99 San Francisco