An evaluation of noise on LPC-based vowel formant estimates: Implications for sociolinguistic data collection

Abstract

Current trends in sociophonetic data analysis indicate a shift to entirely automatic measurements of spectral properties using programs like Praat. While such practices are useful for the rapid collection of acoustic data from large corpora, they, by default do not permit human analysts to provide quality control or make hand corrected measurements when needed. Under ideal signal-to-noise conditions, such as in a sound-proof room, this may not be a problem. However, analysis of audio recordings made in acoustically-uncontrolled environments, like many standard sociolinguistic interviews, are arguably susceptible to spurious estimates using automated routines. This paper presents the results of a highly controlled noise-interference experiment designed to examine the effects of different types of noise at varying signal-to-noise levels on automated LPC-based vowel formant measurements made in Praat. Findings indicate that some noises are more detrimental than others, affect some formant frequencies more than others and that louder noises make it inappropriate to conduct an analysis of F1 and F2. Results are discussed and suggestions for better practices in recording sociolinguistic interviews for sociophonetic data collection are presented.

The beauty in a beast: Minimising the effects of diverse recording quality on vowel formant measurements in sociophonetic real-time studies

“…In such increased noise levels (reflected in lower SNR, see 2.3), formants often appear very faint or have larger bandwidths and are therefore less clearly defined (Plichta 2004); Plichta strongly advises against using such recordings for speech research. De Decker (2016), however, shows that not all types of background noise have an equally damaging effect on the accuracy of formant estimation. High levels of white noise (i.e.…”

“…A series of studies have been conducted in the context of forensic speaker identification (e.g. Byrne and Foulkes 2004;Künzel 2001); and only a few, mostly preliminary investigations have recently pointed out that technical issues of a recording may obscure the patterns of variation and change in sociophonetics, too (De Decker and Nycz 2011;De Decker 2016;Hansen and Pharao 2006;in progress).…”

“…Apart from the influence the format of a recording can have on its spectra and formant measurements taken using LPC, somewhat less obvious factors, such as ambient noise, room acoustics, microphone make and placement during the recording session, have also been shown to leave their spectral imprints and interfere with formant measurements (De Decker 2016;Hansen and Pharao 2006, in progress;Plichta 2004;Švec and Granqvist 2010). The quality of the recordings not controlled for such influencing factors will likely vary with respect to at least two technical specifications (cf.…”

“…De Decker 2016;De Decker and Nycz 2011;Rozborski 2007). Most of them compared recordings made simultaneously with variable equipment, and estimated the resulting differences in formant frequencies those recordings measured (Byrne and Foulkes 2004;De Decker and Nycz 2011;Hansen and Pharao 2006, in progress;Künzel 2001;Plichta 2004;van Son 2005).…”

Three steps forward for predictability. Consideration of methodological robustness, indexical and prosodic factors, and replication in the laboratory

“…The effects on mobile/cell lines are both stronger and more variable: Byrne & Foulkes (2004) found an average upshift of 29% for F1 compared to simultaneous clean recordings, with all vowel categories significantly affected to some extent. The effects of technical issues on data derived from a corpus are further investigated by De Decker (2016) and Rathcke et al (2017).…”