There is growing interest in how speech technologies perform on L2 speech. Largely omitted from this discussion are tools used in the early data processing steps, such as forced aligners, that can introduce errors and biases. This study adds to the conversation and tests how well a model pre-trained for the alignment of L1 American English speech performs on L2 English speech. We test and discuss the impact of language variety, demographic factors, and segment type on the performance of the forced aligner. We also examine systematic errors encountered. Forty-five speakers representing nine L2 varieties were selected from the Speech Accent Archive and force aligned using the Montreal Forced Aligner. The phoneme-level boundary placements were manually corrected in order to assess differences between the automatic and manual alignments. Results show marked variation in the performance across language groups and segment types for the two metrics used to assess accuracy: Onset Boundary Displacement, a distance metric between the automatic and manual boundary placements, and Overlap Rate, which indicates to what extent the automatically aligned segment overlaps with the manually aligned segment. The highest accuracy on both measures was obtained for German and French, and lowest accuracy for Russian. The aligner's performance on all varieties was comparable to that on conversational American English and non-standard varieties of English. Furthermore, the percentage of boundary placements within 10 and 20ms of the corrected boundary was similar to that observed between transcribers. Apart from errors due to variety mismatch, most issues encountered in the alignment were due to issues not exclusive to L2 speech such as inaccurate orthographic transcriptions, hesitations, specific voice qualities, and background noise. The results of this study can inform the use of automatic aligners on L2 English speech and provide a baseline of potential errors and information to help the development of more robust alignment tools for further development of automatic systems using L2 English.

Automatic Speaker Recognition (ASR) involves a complex range of processes to extract, model, and compare speaker-specific information from a pair of voice samples.Using heavily controlled recordings, this paper explores the impact of specific vocal conditions (i.e.vocal setting, disguise, accent guises) on ASR performance.When vocal conditions are matched, ASR performance is generally excellent (whisper is an exception).When conditions are mismatched, as in most forensic cases, we see an increase in discrimination and calibration error in some cases.The most problematic mismatches are those involving whisper and supralaryngeal vocal settings; these produce the greatest phonetic changes to speech.Mismatches involving high pitch also produce poor performance, although this appears to be driven by speaker-specific differences in articulatory implementation.We discuss the implications of the findings for the use of ASR in forensic casework and the interpretability of system output.

In data-driven forensic voice comparison, sample size is an issue which can have substantial effects on system output.Numerous calibration methods have been developed and some have been proposed as solutions to sample size issues.In this paper, we test four calibration methods (i.e.logistic regression, regularised logistic regression, Bayesian model, ELUB) under different conditions of sampling variability and sample size.Training and test scores were simulated from skewed distributions derived from real experiments, increasing sample sizes from 20 to 100 speakers for both the training and test sets.For each sample size, the experiments were replicated 100 times to test the susceptibility of different calibration methods to sampling variability.The Cllr mean and range across replications were used for evaluation.The Bayesian model and regularized logistic regression produced the most stable Cllr values when the sample size is small (i.e.20 speakers), although mean Cllr is consistently lowest using logistic regression.The ELUB calibration method generally is the least preferred as it is the most sensitive to sample size and sampling variability (mean = 0.66, range = 0.21-0.59).

A significant question for forensic voice comparison, and for speaker recognition more generally, is the extent to which different input features capture complementary speakerspecific information.Understanding complementarity allows us to make predictions about how combining methods using different features may produce better overall performance.In forensic contexts, it is also important to be able to explain to courts what information the underlying features are actually capturing.This paper addresses these issues by examining the extent to which MFCCs and LPCCs can predict F0, F1, F2, and F3 values using data extracted from the midpoint of the vocalic portion of the hesitation marker um for 89 speakers of standard southern British English.By-speaker correlations were calculated using multiple linear regression and performance was assessed using mean rho (𝜌) values.Results show that the first two formants were more accurately predicted than F3 or F0.LPCCs consistently produced stronger correlations with the linguistic features than MFCCs, while increasing cepstral order up to 16 also increased the strength of the correlations.There was, however, considerable variability across speakers in terms of the accuracy of the predictions.We discuss the implications of these findings for forensic voice comparison.

ISCA grants each author permission to use the paper in that author's dissertation or in institutional and public repositories

In forensic voice comparison, it is essential to consider not only the similarity between samples, but also the typicality of the evidence in the relevant population. This is explicit within the likelihood ratio (LR) framework. A significant issue, however, is the definition of the relevant population. This paper explores the complexity of population selection for voice evidence. We evaluate the effects of population specificity in terms of regional background on LR output using combinations of the F1, F2, and F3 trajectories of the diphthong /aɪ/. LRs were computed using development and reference data which were regionally matched (Standard Southern British English) and mixed (general British English) relative to the test data. These conditions reflect the paradox that without knowing who the offender is, it is not possible to know the population of which he is a member. Results show that the more specific population produced stronger evidence and better system validity than the more general definition. However, as region-specific voice features (lower formants) were removed, the difference in the output from the matched and mixed systems was reduced. This shows that the effects of population selection are dependent on the sociolinguistic constraints on the feature analysed.

ISCA grants each author permission to use the paper in that author's dissertation or in institutional and public repositories