Phil Rose & Bruce Wang When experts do real-world forensic - - PowerPoint PPT Presentation

CANTONESE FORENSIC VOICE COMPARISON WITH HIGHER-LEVEL FEATURES:

likelihood ratio-based validation using F-pattern and tonal F0 trajectories over a disyllabic hexaphone

Odyssey 2016 (C) Phil Rose June 2016

← Phil Rose & Bruce Wang →

• When experts do real-world forensic voice

comparison case-work they need to know what is the best approach to use under the circumstances of the case.

• There is no more compelling reason for

forensic voice comparison research than this.

• You often get recurrent polysyllabic words

(alright, not too bad, hello, ok) in suspect and

• ffender speech samples.
• We usually model the F-pattern trajectory of

these on separate syllables (o - k, not – too – bad).

• I wanted to find out whether you get better

strength of evidence (aka LR) if you model the (usually more complex) trajectory of the word acoustics as a whole – superphthong.

• Tested with higher-level features (F-pattern and

tonal F0) in the Cantonese word 第一 daihyat L.H first using LR validation.

• I know there is nothing I can tell anyone here

about automatic speaker recognition …

• But I hope there might be a few aspects of the

simpler higher-level world you find interesting.

• More importantly, though, I DO have some things

to say about Likelihood Ratios in Forensic Voice Comparison based on my real-world case experience.

• And (more importantly still) I want to learn from

you.

• So please drop by.