When: Dec 9, 2020 – 11:00 – 11:45 AM
Where: google meet link
The problem of speaker verification, i.e., testing the speaker’s claim of identity using previously collected voice recordings, has a history of more than four decades. I decided to investigate the problem of speaker verification in the light of a recently established collaboration between DIISM and a company, which is interested in verifying caller’s identity through speaker verification.
Deep neural networks allow us to implement speaker verification systems with relatively good performance, typically exploiting a speaker recognition task or rather learning a suitable distance metric in the pattern space. I will discuss an alternative approach based on autoencoders for pattern verification and investigate the feasibility of an ensemble approach. In my thesis work, much attention has been devoted to practical challenges related to the real world setup, including integration with the enterprise workflow.