Assessing a speaker’s voice quality for forensic purposes: Using the example of creaky voice and breathy voice

The research project explores ways to improve the assessment of voice quality (VQ) for forensic voice comparisons. Until today, a speaker’s VQ is mainly assessed perceptually. However, the field has developed rapidly over the last two decades, prompting calls to objectify the analysis process by relying on voice acoustics instead. This poses a challenge as forensic audio recordings are degraded in several aspects.
The first study focuses on creaky voice (CV), which is particularly multifaceted in production and thus also in acoustics. Therefore, perceptually relevant categories must first be defined and tested before acoustic analysis can be conducted. A new CV classification scheme is conceptualised and tested. It is hypothesised that differences in speaker-specific CV spaces will facilitate speaker discrimination.
Using the example of breathy voice (BV), the second and fourth studies analyse the interplay between perception and acoustics. Spontaneous speech samples of BV speakers are compared with those of non-BV speakers under the studio condition and under the mobile phone condition. Under the studio recording condition, three parameters were found to correlate between perception and acoustics, i.e. H1*-H2*, H1*-A1*, CPP. Under the mobile recording condition, however, low frequency harmonics are attenuated and thus not meaningful. Therefore, the spectral tilt parameters of higher frequencies should be analysed instead.
The third study explores the suitability of f0 estimators with respect to recording condition, and VQ. Valid f0 estimation is required to obtain valid spectral slope measurements. The is explored using sustained cardinal vowels of one male and one female speaker in modal, breathy, and creaky VQ under two recording conditions (studio, mobile phone). Results allow for an informed decision which f0 estimator to use.
The research project sheds light on the needs and possibilities to refine VQ analysis for forensic application.