Mat. stat. seminarium 21 februari 2000

$KTH"$

Tid: 21 februari 2000 kl 1515-1600

Plats : Seminarierummet 3733, Institutionen för matematik, KTH, Lindstedts väg 25, plan 7. Karta!

Föredragshållare: Ylva Gustafsson

Titel: Hidden Markov Models with applications in speaker verification

Sammanfattning:

Speech recognition and speaker verification can be performed using a class of statistical models called hidden Markov models, HMMs. This report describes the field of HMMs, its use in speaker verification and some specific methods connected to parameter estimation in the models of the speakers.

Four methods are theoretically described:

maximum likelihood (ML)
maximum a posteriori (MAP)
maximum likelihood linear regression (MLLR)
minimum classification error (MCE)

A lot of research has already been done on the performance of models obtained through the ML method, and in this work the methods MAP and MLLR are investigated and compared to the ML method.

As far as the results of the tests performed in this work show, the methods MAP and MLLR are not better than ML, but there are still a lot of factors in the setup with need of further investigation.

Till seminarielistan
To the list of seminars