Timeshrinking
From SpeechWiki
Fisher experiments
| train <math>\tau</math> | test <math>\tau</math> | dev 2k utt WER | frames dropped | comments |
|---|---|---|---|---|
| 1 | 1 | 51.6% | 0% | old baseline |
| 1 | 1 | TR | 0% | new baseline |
| .6 | .6 | 69.3% | ~35% | |
| .9 | .9 | 56.3% | ~5% | |
| 1 | .9 | 53.9% | ~5% | |
| .9 | .9 | 57.2 | ~5% | using the non-timeshrinking str file for test |
| .9 | 1 | 54.6 | ~5% |
I have to check for bugs. It could be that the threshold is too low or it could be something else too. We should probably rerun baseline too, just to make sure I didn't optimize it unfairly.
Things to try
- Test svitchboard with fisher-trained model to see if we still get good results
- Train and test on plp+mlp, like svitchboard timeshrinking was done.
- Do baseline train+test to see if something changed in going from baseline to timeshrink structure files.
Future Directions
- Can be viewed as a two-mode special case of best-first viterbi search. So make a real best-first lattice search. Mark mentioned some attempts in the 80'ies to do this.
