User Tools

Site Tools


asr:adaptation

Standard adaptation methods are:

  • MLLR
  • MAP
  • FMLLR/CMLLR
  • FMAPLR
  • Eigenvectors

More advanced ones

  • Structured MAP
  • Sparse MAP

The recent research recommends to consider factored MLLR where specific transforms correspond to speaker, channel and environment

For semi-continuous models only MAP adaptation make sense. If one needs to adapt semi-continuous model with less parameters he need to consider some other methods like this one

CROSSLINGUAL ADAPTATION OF SEMI-CONTINUOUS HMMS USING ACOUSTIC SUB-SIMPLEX PROJECTION Frank Diehl, Asuncion Moreno, Enric Monte

Adaptation is good together with SAT training.

For MAP we can't estimate tau smoothing parameter reliably yet.

asr/adaptation.txt · Last modified: 2012/04/29 07:16 by admin