Mapping Across Feature Spaces in Forensic Voice Comparison: The Contribution of Auditory-Based Voice Quality to (Semi-)Automatic System Testing

Abstract

In forensic voice comparison, there is increasing focus on the integration of automatic and phonetic methods to improve the validity and reliability of voice evidence to the courts. In line with this, we present a comparison of long-term measures of the speech signal to assess the extent to which they capture complementary speaker-specific information. Likelihood ratiobased testing was conducted using MFCCs and (linear and Melweighted) long-term formant distributions (LTFDs). Fusing automatic and semi-automatic systems yielded limited improvement in performance over the baseline MFCC system, indicating that these measures capture essentially the same speaker-specific information. The output from the best performing system was used to evaluate the contribution of auditory-based analysis of supralaryngeal (filter) and laryngeal (source) voice quality in system testing. Results suggest that the problematic speakers for the (semi-)automatic system are, to some extent, predictable from their supralaryngeal voice quality profiles, with the least distinctive speakers producing the weakest evidence and most misclassifications. However, the misclassified pairs were still easily differentiated via auditory analysis. Laryngeal voice quality may thus be useful in resolving problematic pairs for (semi-)automatic systems, potentially improving their overall performance.

The use of the Vocal Profile Analysis for speaker characterization: Methodological proposals

“…nasality, creakiness) increased, decreased or remained the same when comparing high quality recordings and telephone-degraded recordings. We have investigated possible correlations between different long-term vocal tract output measures -including supralaryngeal settings -with the aim of finding how VQ analysis can complement longterm formant distributions (LTFDs) and Mel frequency cepstral coefficients calculated across entire speech samples (MFCCs; French et al 2015, Hughes et al 2017.…”

The Individual and the System: Assessing the Stability of the Output of a Semi-automatic Forensic Voice Comparison System

“…Recordings were drawn from the DyViS corpus of young male standard southern British English speakers [12]. Of the 100 available speakers, 97 were used, based on prior testing outlined in [13]. SASR testing was carried out under four different conditions according to the technical quality of the offender sample.…”

“…The suspect recording (Task1) and the four versions of the offender recording (Task2) were prepared for analysis in the same way as described in ¤2.2 of [13]. This involved manual editing of recordings to remove non-speech sounds and overlapping speech, removal of sections containing clipping, voice activity detection to remove silences of greater than 100ms (using the vadsohn function in the VOICEBOX toolkit [15]), and segmentation of the signal into consonants and vowels using stkCV [16].…”

“…Delta coefficients were also appended to the feature vector for each frame. Bandwidths and deltas were included in this SASR system as they have generally been shown to improve performance [6,13].…”

“…Mean formant values for each speaker based on pooled data for their HQ Task1 and Task2 samples were also used as independent variables Ð we predicted that low mean F1 would result in higher LLR variability as it is more susceptible to the Ôtelephone effectÕ which artificially alters F1 values [21,22]. Auditory-based judgments of supralaryngeal and laryngeal voice quality (using data described in [13]) were also used as independent variables. The best model fit was identified using model comparison based on ANOVAs.…”