Describing voice in French media archives: age and gender effects on pitch and articulation characteristics
Abstract
Changes in voices and speech characteristics over time are widely commented on and deemed positive or negative. The assumed variation may rely on impressionistic feelings or hearsay rather than verifiable data. For example, some claim that women’s voices have lowered their pitch over the past century; such claims are generally sourced in a study based on two groups of carefully selected Australian female voices [16]. As many factors affect spoken voices (e.g., [11]), inter-speaker variability is the norm, and substantial changes are observed across situations for the same speaker. A difficulty with most diachronic studies is linked to the imbalance between genders in their corpora: the work of Pemberton et al. [16] typically does not study male voices, so it is difficult to understand whether change is general or gender-specific. Claiming a shift in spoken voice over time requires large datasets, which may represent part of the described population. In this respect, archives of audiovisual media now gather more than half a century of recordings and constitute invaluable resources to dive into our past and study spoken practices across time [5, 20, 23]. Meanwhile, there is a strong gender imbalance in media, both quantitatively [8] and qualitatively [6,7], sometimes leading researchers to remove female voices from the dataset to avoid bias [3]. This communication proposes an acoustic analysis of female and male voices, presented to public audiences through French TV and radio from the 1950s to the 2010s. Our long-term objective is to investigate potential changes that may be observed diachronically through such media archives in the way people spoke in French media, focusing on possible gender-specific, age-balanced, differences. The current presentation will focus on two characteristics of the voice’s perceived pitch: the voice fundamental frequency (F0) and the vocal tract resonances [4,10,14,19]. We’ll discuss our findings and what they mean for the understanding of voice and individual features, the strengths and limitations of the approach, and potential future works.
This study is based on a selection of 1,028 individual speakers, balanced for gender and age categories (20-35, 36-50, 51-65, >65 y.o.), found across the archives of the French National Audiovisual Institute (INA). The raw speech excerpts (representing about 60 hours of speech) were automatically transcribed using Whisper [17]; the obtained transcriptions were forced-aligned using MFA[15] and its pre-trained models for French. From this aligned dataset, only the oral vowels were retained (the set provided by MFA being: /i y e ø ɛ œ ə a ɑ ɔ o u/), removing unvoiced phones as determined by Reaper [21] and Praat [2] and phones for which F0 values estimated by Praat and FCN-F0 [1] differ by more than 20% (see [18] for a similar procedure). This leads to more than one million individual vowels. The first four formants (F1, F2, F3, F4) were estimated using Praat’s Burg algorithm and applying the method of optimal ceiling estimation presented in Escudero et al. [9]. A single value of each measurement was taken at the center of each vowel. From the formant values, the vocal tract length (VTL) was estimated following [12, 13]. On these measurements (F0, formants 1 to 3, and VTL), linear mixed models were fitted with each measurement as the dependent variable and the speaker’s Age and Gender, and the recording time Period as fixed factors; the Vowel and the Speaker were used as random factors. Results show some tendency of these pitch markers to cue lower voice pitch at more recent Periods, but with a complex link to Gender, possibly linked to the differential effect of Age on voice [22]. Potential differences in articulation and style and their impact on the outcome will be discussed, trying to weigh possible effects on formants and F0 linked to vocal effort.