1. Optimal Adaptation for Statistical Classifiers Xiao Li

2. Motivation Problem  A statistical classifier works well if the test set matches the data distribution of the train set  It is difficult to get a large amount of matched training data A case study – vowel classification  Target test set – pure vowel articulation for specific speakers  Available train set – conversational speech with a great number of speakers

3. Adaptation Methodology 1. Extract vowel segments from conversational speech to form a train set 2. Feature extraction and class labeling 3. Train speaker-independent models on this train set 4. Ask a speaker to articulate a few seconds of vowels for each class 5. Adapt the classifier on this small amount of speaker- dependent, pure vowel data

4. Two Classifiers Gaussian mixture models (GMM)  Generative models  Training objective: maximum likelihood via EM Neural Networks (NN)  Multilayer perceptrons  Training objective: Least square error Minimum relative entropy

5. MLLR for GMM Adaptation Maximum Likelihood Linear Regression  Apply a linear transformation on the Gaussian mean  Same transformation for the mixture of Gaussians in the same class Adaptation Objective  Find the transformation matrices that maximizes the likelihood via EM

6. NN Adaptation Idea -- Fix the nonlinear mapping and update the last layer of linear classifier Two alternative methods with different objectives 1. Minimum relative entropy Optimization method – gradient descent 2. Optimal hyper-plane Optimization method – support vector machine

7. Vowel Classification Experiments Databases  Database A – speaker-independent conversational speech  Database B – sustained vowel recordings from 6 speakers, with different energy and pitch Method 1. Train speaker-independent classifiers Database A s 2. Adapt classifiers on a small set of Database B, 300- 500 samples per speaker 3. Test on the rest of Database B