1. An Alternative Approach of Finding Competing Hypotheses for Better Minimum Classification Error Training Mr. Yik-Cheung Tam Dr. Brian Mak

2.  Motivation  Overview of MCE training  Problem using N-best hypotheses  Alternative:1-nearest hypothesis  What?  Why?  How?  Evaluation  Conclusion Outline

3. MCE Overview  The MCE loss function:  Distance measure:  G(X) may be computed using the N-best hypotheses.  l(.) = 0-1 soft error-counting function (Sigmoid)  Gradient descent method to obtain a better estimate.

4.  When d(X) gets large enough,  It falls out of the steep trainable region of Sigmoid. Trainable region Problem Using N-best Hypotheses

5. What is 1-nearest Hypothesis?  d(1-nearest) <= d(1-best)  The idea can be generalized to N-nearest hypotheses.

6.  Keep the training data inside the steep trainable region. Trainable region Using 1-nearest Hypothesis

7.  Method 1 (exact approach)  Stack-based N-best decoder Drawback: N may be very large => memory problem Need to limit the size of N.  Method 2 (approximated approach)  Modify the Viterbi algorithm with a special pruning scheme. How to Find 1-nearest Hypothesis?

8. Approximated 1-nearest Hypothesis  Notation:  V(t+1, j) : accumulated score at time t+1 and state j  : transition probability from state i to j  : observation probability at time t+1 and state j  : accumulated score of the Viterbi path of the correct string at time t+1.  Beam(t+1) : beam width applied at time t+1

9.  There exists some “nearest” path in the search space (shaded area). Approximated 1-nearest Hypothesis (.)

10. System Evaluation

11. Corpus: Aurora  Aurora  Noisy connected digits derived from TIDIGIT.  Multi-condition training: (Train on noisy condition)  {subway, babble, car, exhibition} x {clean, 20, 15, 10, 5} (5 noise levels)  8440 training utterances.  Testing: (Test on matched noisy condition)  Same as above except with additional samples with 0 and –5 dB (7 noise levels)  28,028 testing utterances.

12. System Configuration  Standard 39-dimension MFCC (cep +  +  )  11 Whole-word digit HMM (0-9, oh)  16 states, 3 Gaussians per state  3-state silence HMM, 6 Gaussians per state  1-state short pause HMM tied to the 2 nd state of the silence model.  Baum-Welch training to obtain the initial HMM.  Corrective MCE training on HMM parameters.

13.  Compare 3 kinds of competing hypotheses:  1-best hypothesis  Exact 1-nearest hypothesis  Approx. 1-nearest hypothesis  Sigmoid parameters:  Various (control slope of Sigmoid)  Offset = 0 System Configuration (.)

14.  Learning rate = 0.05, with different  0.1 (best test performance)  0.5 (steeper)  0.02, 0.004 (more flat) Experiment I: Effect of Sigmoid slope Baseline: 12.71% 1-best: 11.01% Approx. 1-nearest: 10.71% Exact 1-nearest: 10.45%

15.  Soft error < 0.95 is defined to be “effective”.  1-nearest approach has more training data when the Sigmoid slope is relatively steep. Effective Amount of Training Data 1-best (40%) Approx. 1-nearest (51%) Exact. 1-nearest (67%)

16.  With 100% effective training data, apply more training iterations:  = 0.004, learning rate = 0.05  Result: Slow improvement compared to the best case. Experiment II: Compensation With More Training Iterations Exact 1-nearest with gamma = 0.1

17.  Use a larger learning rate (0.05 -> 1.25)  Fix = 0.004 (100% effective training data)  Result: 1-nearest approach is better than one-best approach after compensation. Experiment II: Compensation Using a Larger Learning Rate SystemBefore compensation After compensation Baseline12.71% 1-best12.07%11.55% Approx 1-nearest12.27%10.70% Exact 1-nearest12.16%10.79%

18. Using a Larger Learning Rate (.)  Training performance: MCE loss versus # of training iterations. 1-best Approx. 1-nearest Exact. 1-nearest

19. Using a Larger Learning Rate (..)  Test performance: WER versus # of training iterations. Approx. 1-nearest (10.70%) Exact. 1-nearest ( Exact. 1-nearest ( 10.79%) 1-best ( 1-best ( 11.55%)

20. Conclusion  1-best and 1-nearest methods were compared in MCE training.  Effect of Sigmoid slope.  Compensation on using a flat sigmoid.  1-nearest method is better than 1-best approach.  More trainable data are available in the 1-nearest approach.  Approx. and exact 1-nearest methods yield comparable performance.

21. Questions and Answers