Identifying speaker-dependent acoustic parameters in Spanish vowels

Abstract

In the frame of VILE Projects (Inter-and-Intra-Speaker-Variation-in-Spanish for automatic speaker identification), we try to identify what vowel's acoustic parameters depend more on the individual characteristics of the speaker and less of the linguistic variables. Variations on standard deviation (SD), when grouping together parameters by speaker or by phoneme, are analysed. 30 speakers (from AHUMADA database) read the same text in three sessions. Mean value (Hz) of four formants (F1-F2-F3-F4) and fundamental frequency (F0) are analysed in Spanish vowels (except /u/), surrounded by unvoiced stops or /s/ (1850 samples). Hypothesis: Individual parameters will show less SD when grouping by speaker/session; vowel quality parameters, when grouping by phoneme. F1 and F2 are timbre-dependent parameters. F3 and F4 are speaker-dependent parameters. F0 has characteristics of both. The most variable parameter is F2. The opposite is F4. -No significant differences grouping by session or by speakers in none of the parameters. -F0 has the highest variability between vowel qualities, even if stressed and unstressed vowels are separated. -When clustering data by speaker/sessión (all vowels together), by comparison with clustering by vowel (all speakers together), SD is *50% higher in F1-F2 *75% lesser in F4 *66% lesser in F0 *F3 shows no significant differences