Speech Science Forum - Vincent Hughes (York)

Abstract:

As part of our recent AHRC project, Humans and Machines: Novel Methods for Testing Speaker Recognition Performance, we created an immersive, jury-like data elicitation game. Our game provides a way of testing lay-listener performance at speaker recognition, whilst taking into account sources of cognitive bias that are common in legal contexts; e.g. the fact that the decision is being made in a criminal case, additional evidence in a case, expert opinion. In this talk, I will demo our game and present results from across all of the levels we have tested so far. I will also examine lay-listener performance relative to, and in combination with, a state-of-the-art automatic speaker recognition system. The implications for voice evidence in legal contexts will be considered.