Is there a universal impact of duration on formant frequency values of oral vowels?
An automated analysis of speech from eight languages
Résumé
The present research addresses the question whether an automated analysis of large speech corpora can contribute to the study of the complex issue of phonetic and phonological universals. Ongoing investigations focus on cross-language patterns of variation of vowel quality (frequency of the first and second formant, hereafter F1 and F2) as a function of (phonetic) vowel length. We refer here to the Hyper- & Hypoarticulation model (Lindblom, 1990), for which hypoarticulated forms are produced with unreached targets (“undershoot”) while hyperarticulated phonemes are realised with a more peripheral articulation (“overshoot”); both being respectively related to shorter or longer segment duration. Speech data used come from annotated corpora of broadcast speech in eight languages: Arabic, Mandarin Chinese, American English, French, German, Italian, European Portuguese, and Spanish. All corpora correspond to radio and TV journalistic shows1: articulation, without being emphasized, remains quite distinct, so that speech can be understood by a broad audience. Reduced vowel phenomena are undoubtedly less salient in these corpora than in conversational-style spontaneous speech. The LIMSI speech transcription system was used for corpus alignment (Adda-Decker et al., 2005) and formant extraction makes use of the Burg algorithm implemented in the PRAAT software. F1 and F2 frequencies were extracted automatically (following a procedure set out in detail in Gendrot & Adda-Decker, 2005). The discussion focuses on cross-linguistic similarities in the relationship between the measured vowel length and the frequency of F1 (providing a cue to the degree of opening/closing) and F2 (related to the degree of frontness/backness and labialisation). The main objectives are to assess: (i) is there a general tendency of formant variations with segment duration in the eight languages? (ii) which vowels are most prone to variation? (iii) To what extent the observed patterns can be related to the prosodic system of these eight languages? The latter objective can be detailed as follows: in English (and generally in lexical stress languages such as German, Spanish, Italian and Portuguese) where the vowels of unstressed syllables typically undergo reduction (neutralization to //, with some exceptions, notably for /ɪ/), the prediction is that length should be more strongly correlated with vowel quality (as reflected in the frequencies of the first two formants) than in a language such as Chinese, where the differences in length across lexical tones (e.g. the brevity of the falling tone, tone 4, as opposed to the rising tone, tone 2) is not reported to have any salient effect on vowel quality. As for Arabic, word stress is not phonemically contrastive but bears a strong relationship to vowel length. In general, "heavy" syllables attract stress (i.e. syllables of longer duration - a closed syllable or a syllable with a long vowel). Since vowel length is 1 The present study was carried out in the framework of the MIDL (Modelisations for Identification of Languages) project. MIDL partners were LIMSI-CNRS, LPP Paris3, CTA/DGA/GIP, Télécom Paris, EA1483 Paris 3. phonologically contrasted, we expect fewer reduction phenomena. It is also expected that in French, which has neither lexical tones nor lexical stress, and is reputedly a ‘syllable-timed' language, there will be less influence of vowel length on vowel quality. Main emphasis has first been put on German and French so far with 25000 vowels (from 2 hours of speech per language. Results for German and French showed that only 4% of all the measured vowel formants of the corpus were to be discarded due to unlikely values. All the measured vowel formants of the corpus (except the 4%) occupy the vocalic space in an organised manner: for segments which are longer than 90 ms, the automatically measured values are very close to the reference values published for French and German in the literature. For shorter durations the vocalic triangle shrinks progressively resulting in “concentric” surfaces for both French and German: the vocalic surfaces undergo a centripetal movement towards some central vowel position. The formant values for central vowels remain somewhat stable whatever the duration variations. French high front vowels /i/ and /y/ seem to be less prone to variation, which could be explained by the pertinence of the third formant (F3) for these vowels in French. However important variations are still to be noticed in German for /i/ and /y/ (both tense and lax counterparts). For the other languages approximately 5000 vowels (~30 minutes of speech) have been gathered and analysed up to now per language. The corpus is being extended to 25000 vowels per language in the final contribution. Achieved results show that contrary to expectations, vowel reductions are observed with a comparable degree in 6 languages (French, Spanish, German, European Portuguese, Italian and English). This suggests that reduction is not an exclusively linguistic phenomenon, but admits also explanations of a physical or physiological nature. Variations have also been measured for Arabic and Mandarin although less important
Domaines
Linguistique
Loading...