Comparing praat and snack formant measurements on two large corpora of northern and southern French

Abstract

Abstract We compare formant frequency measurements between two authoritative tools (Praat and Snack), two large corpora (of face-to-face and telephone speech) and two French varieties (northern and southern). There are both an evaluation of formant tracking (as well as related filtering techniques) and an application to find out salient pronunciation traits. Despite differences between Praat and Snack with regard to telephone speech (Praat yielding greater F1 values), results seem to converge to suggest that northern and southern French varieties mainly differ in the second formant of the open /O/. /O/ fronting in northern French (with F2 values greater than 1100 Hz for males and 1200 Hz for females) is by far the most discriminating feature provided by decision trees applied to oral vowel formants. Index Terms : Praat and Snack formant tracking evaluation, northern and southern French accent identification, phonetics 1. Introduction Freely-available signal processing tools performing formant tracking are now widely used. The Praat software (www.fon.hum.uva.nl/praat) and the Snack Sound Toolkit (http://www.speech.kth.se/snack/) appear as the most widespread ones. They have both been developed to enable the creation of multi-platform applications with scripting languages. However, to our knowledge, they lack comparative evaluation on large databases [7]. Since the latter do not allow a manual labelling, the accuracy of this or that technique may depend on what is intended. This study focuses on the characterisation of two varieties of the same language: northern and southern French. Two large corpora both including northern (i.e. standard) and southern French speakers were analysed to assess the impact of the recording conditions. The first corpus comes from the PFC project (“Phonology of Contemporary French”) [3]: following a Labovian protocol [9], it comprises reading and spontaneous speech. The present study is based upon a dozen investigation points from this corpus: over 100 speakers who were firmly rooted in their places of residence. The second corpus is less controlled, but contains conversational telephone speech (CTS) of over 500 speakers from 7 broad French regions. CTS may cause specific problems for formant tracking, and there may be a need for parameter tuning. The two corpora, which have been segmented into phonemes and phonetically transcribed by automatic align-ment, are introduced in further detail in the next section. Various formant normalisation procedures have been proposed so far, to rule out speaker-specific physiological characteristics, but they have drawbacks [1]. We believe our corpora are large enough to handle formant measures in Hertz (Hz) without resorting to intrinsic or extrinsic normalisation. Other questions are raised in section 3: how many measurements per vowel are needed? How to cope with measurement errors? Results are reported in section 4, in terms of correlations and distances between formant values of males and females. The comparison between Praat and Snack on the two corpora is completed by a comparison between northern and southern French vocalic triangles. To try and discriminate these two varieties, data mining techniques are then used (section 5), and a linguistic discussion on a possible sound change in progress (namely // fronting in northern French) concludes this paper.