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.