University Homepage School Homepage School Contacts School Search

Lecture 10 - Speaker Adaptation

In this lecture we discussed n-gram and neural network language models.

Basics

• Adaptation at acoustic, pronunciation, and language model level
• Supervised or unsupervised adaptation
• Static or dynamic adaptation
• Approaches
  • Model-based
  • Speaker normalisation
  • Speaker space
• Desirable properties
  • Compact
  • Efficient
  • Unsupervised
  • Flexible

Adapting GMM-based systems

• Maximim a posteriori - MAP
  • Basic formulation
  • Adapting the mean using MAP
  • Relationship to maximum likelihood
• Maximum likelihood linear regression - MLLR
  • Linear transform of mean parameters, and estimating the transform
  • Regression classes
• Constrained MLLR - cMLLR
  • Basic formulation
  • Equivalence to feature space transform
• Speaker-adaptive training

Adapting NN-based systems

• CMLLR feature transform using corresponding GMM system
• Linear input network - LIN
• Speaker codes
• Learning hidden unit contributions - LHUC

---

This page maintained by Steve Renals.
Last updated: 2018/04/30 20:59:23UTC