Elamur Research⌕
☰
home / papers / 2015 / Likelihood Ratio-based Forensic Voice Comparison on L2 speakers: A Case of Hong Kong native male production of English vowels

Likelihood Ratio-based Forensic Voice Comparison on L2 speakers: A Case of Hong Kong native male production of English vowels

Paper · 2015 · Daniel Frost and Shunichi Ishihara

Summary

Forensic voice comparison (FVC) research within the likelihood ratio (LR) framework has so far been built almost entirely on native speech. Offenders, however, are often L2 speakers. This pilot study asks whether an LR-based FVC system can work on non-native English. The authors recorded 15 male Hong Kong Cantonese speakers in two sessions held at least a week apart. Using an English-language map task, they drew out six frequently used target items from the speakers: "yes", "no", "left", "right", "hello" and "bye".

Formant trajectories (F1–F3) covering each vowel and its neighbouring transitions were measured in the EMU Speech Database System. After duration equalisation, the trajectories were modelled with discrete cosine transforms (DCTs). The DCT coefficients served as feature vectors for the multivariate-kernel-density (MVKD) LR procedure. Scores were then calibrated with logistic regression and evaluated by Cllr (validity) and 95% credible intervals (reliability).

Words with more dynamic trajectories, especially "bye", worked best. The best results were roughly on par with an earlier native Australian English system. This suggests L2 speech is not a fundamental obstacle for FVC.

Key points

  • The data are 15 HK Cantonese male speakers recorded in two non-contemporaneous sessions (mean gap about 12.9 days), with 5 tokens per word. Each word gave 15 same-speaker and 210 different-speaker comparisons.
  • The parameter grid covered 6 words × 3 formant sets (f12, f23, f123) × DCT coefficient counts 2–9, giving 144 outputs.
  • "bye" reached the lowest Cllr (0.158) but had a comparatively wide credible interval. "yes" performed worst.
  • Using all three formants (f123) beat the two-formant sets, and DCT curves of degree 2–5 tended to give the best outputs.
  • In a rough comparison with Morrison (2009b), native /aɪ/ scored Cllr 0.156 against 0.158 for "bye". Native /oʊ/ scored 0.129 against 0.318 for "no" and 0.392 for "hello".

Interesting ideas

  • I1. The LR splits evidence strength into two parts: similarity of the two samples (numerator) and their typicality against a background population (denominator). A match counts for little if the shared trait is common.
  • I2. The expert reports only how much more probable the evidence is under one hypothesis than its rival, and leaves the prior odds and the final verdict to others.
  • I3. Systems are judged on two separate axes, validity (Cllr) and reliability (credible interval). The paper shows the top system on one axis can be weak on the other.
  • I4. Raw MVKD outputs are passed through logistic-regression calibration so that the numbers behave like genuine likelihood ratios.
  • I5. A map task gathers many repetitions of chosen target words in fairly natural speech instead of a stilted word list.

For Mova

How these ideas could serve Mova, our small open language model — its files, stages and tests:

  • I1 (frames): In coreference, a link between two mentions could be scored by how well they match, discounted by how common that description is in the text or corpus.
  • I2 (answer): Readers could report how strongly the text supports an answer relative to alternatives, keeping priors from world knowledge explicit and separate for explainability.
  • I3 (learn): Evaluation could report both accuracy and spread across witnesses or folds, so that a high but unstable score is not mistaken for a robust one.
  • I4 (tree): Parser and gparse arc scores could be calibrated the same way, so confidence values are meaningful when witnesses are combined or thresholds are set.
  • I5 (answer): Test sets could use task-driven scenarios, such as giving directions in a story, to elicit repeated target constructions naturally while keeping control over what is being tested.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

 Fields, topics and concepts

📚 Shelves 6The field of science this work belongs to (OpenAlex fields).Computer Science · 18,614 worksArtificial Intelligence · 17,965 worksArts and Humanities · 17,450 worksLanguage and Linguistics · 17,296 worksPsychology · 1,649 worksExperimental and Cognitive Psychology · 930 works

🧭 Topics 1Research subjects and languages the work is about.English · 2,379 works

Cite

Cite as Daniel Frost and Shunichi Ishihara. 2015. Likelihood Ratio-based Forensic Voice Comparison on L2 speakers: A Case of Hong Kong native male production of English vowels. Proceedings of the Australasian Language Technology Association Workshop 2015.

The paper itself, not this page. BibTeX from the library's record of the work.

@inproceedings{frost2015likelihood,
  title = {{Likelihood Ratio-based Forensic Voice Comparison on L2 speakers: A Case of Hong Kong native male production of English vowels}},
  author = {Daniel Frost and Shunichi Ishihara},
  booktitle = {Proceedings of the Australasian Language Technology Association Workshop 2015},
  year = {2015},
  url = {https://aclanthology.org/U15-1005/}
}

References

27 references, 0 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Aitken, C. G. (1995). Statistics and the evaluation of evidence for forensic scientists. Chichester: J. Wiley. find
  2. Aitken, C. G., & Lucy, D. (2004). Evaluation of trace evidence in the form of multivariate data. Journal of the Royal Statistical Society: Series C (Applied Statistics), 53(1), 109-122. ↗
  3. BIPM, I., IFCC, I., IUPAC, I., & ISO, O. (2008). Evaluation of measurement data—guide for the expression of uncertainty in measurement. JCGM 100: 2008. find
  4. Brümmer, N., & du Preez, J. (2006). Application-independent evaluation of speaker detection. Computer Speech & Language, 20(2), 230-275. ↗
  5. Cox, F. (2012). Australian English pronunciation and transcription. Cambridge: Cambridge University Press. find
  6. Durou, G. (1999). Multilingual text-independent speaker identification. Proceedings of the Multi-Lingual Interoperability in Speech Technology, 115-118. find
  7. Evett, I. (1998). Towards a uniform framework for reporting opinions in forensic science casework. Science & Justice, 38(3), 198-202. ↗
  8. Harrington, J., Cox, F., & Evans, Z. (1997). An acoustic phonetic study of broad, general, and cultivated Australian English vowels. Australian Journal of Linguistics, 17(2), 155-184. ↗
  9. Ishihara, S., & Kinoshita, Y. (2008). How many do we need? Exploration of the population size effect on the performance of forensic speaker classification. Proceedings of the Interspeech 2008, 1941-1944. find
  10. Kajarekar, S. S., Scheffer, N., Graciarena, M., Shriberg, E., Stolcke, A., Ferrer, L., & Bocklet, T. (2009). find
  11. The SRI NIST 2008 speaker recognition evaluation system. Proceedings of the International Conference on Acoustics, Speech and Signal Processing (ICASSP 2009), 4205–4208. find
  12. Ladefoged, P., & Disner, S. F. (2012). Vowels and Consonants. Chichester: Wiley. find
  13. Li, J., & Rose, P. (2012). Likelihood ratio-based forensic voice comparison with F-pattern and tonal F0 from the Cantonese /ɔy/ diphthong. Proceedings of the 14th Australasian International Conference on Speech Science and Technology, 201-204. find
  14. Morrison, G. S. (2008). Forensic voice comparison using likelihood ratios based on polynomial curves fitted to the formant trajectories of Australian English /aɪ/. International Journal of Speech Language and the Law, 15, 247-264. find
  15. Morrison, G. S. (2009a). Forensic voice comparison and the paradigm shift. Science & Justice, 49(4), 298-308. ↗
  16. Morrison, G. S. (2009b). Likelihood-ratio forensic voice comparison using parametric representations of the formant trajectories of diphthongs. The Journal of the Acoustical Society of America, 125, 2387-2397. ↗
  17. Morrison, G. S. (2011). Measuring the validity and reliability of forensic likelihood-ratio systems. Science & Justice, 51(3), 91-98. ↗
  18. Morrison, G. S. (2013). Tutorial on logistic-regression calibration and fusion: Converting a score to a likelihood ratio. Australian Journal of Forensic Sciences, 45(2), 173-197. ↗
  19. Morrison, G. S., & Kinoshita, Y. (2008). Automatic-Type Calibration of Traditionally Derived Likelihood Ratios: Forensic Analysis of Australian English /o/ Formant Trajectories. Proceedings of the Interspeech 2008, 1501-1504. find
  20. Morrison, G. S., Ochoa, F., & Thiruvaran, T. (2012). Database selection for forensic voice comparison. Proceedings of the Odyssey 2012, 62-77. find
  21. Morrison, G. S., Zhang, C., & Rose, P. (2011). An empirical estimate of the precision of likelihood ratios from a forensic-voice-comparison system. Forensic science international, 208(1), 59-65. ↗
  22. Robertson, B., & Vignaux, G. A. (1995). Interpreting evidence. Chichester: Wiley. find
  23. Rose, P. (2002). Forensic speaker identification. London & New York: Taylor & Francis Forensic Science Series. ↗
  24. Rose, P. (2004). Technical forensic speaker identification from a Bayesian linguist's perspective. Proceedings of the Odyssey 2004, 3-10. find
  25. Rose, P. (2013). More is better: likelihood ratio-based forensic voice comparison with vocalic segmental cepstra frontends. International Journal of Speech Language and the Law, 20(1), 77-116. ↗
  26. Scheffer, N., Ferrer, L., Graciarena, M., Kajarekar, S., Shriberg, E., & Stolcke, A. (2011). find
  27. The SRI NIST 2010 speaker recognition evaluation system. Proceedings of the International Conference on Acoustics, Speech and Signal Processing (ICASSP 2011), 5292-5295. find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/frost-2015-likelihood-ratio-based-forensic-voice/