|
Open Source Toolkit For Speech Recognition Project by Carnegie Mellon University |
Description
The idea with Frame Skew Pruning is reduce the number of similar paths exiting from an HMM or a word over a small period of time. Typically, as the search moves through the search space, a number of paths will exit a particular HMM. These paths are similar but are skewed by a small number of frames. With FrameSkewPruning, when a path exits an HMM or a Word, it is compared to the most recent path that exited the HMM. The path is allowed to exit and continue if it: (a) Scores better than the best exiting path, or (b) the last exiting path is older than the frameSkew property (K).
For example, if K is set to 2 (meaning 2 frames), when path (a) exits HMM-1 in frame 1, it marks HMM-1 with itself. if Path (b) exits HMM-1 in frame 2, it only continues if its score is greater than the path(a) score. If path © exits HMM-1 in frame 5 it will become the best path since it is more than K frames newer than the last path out of HMM-1.
Results
The following table shows results for a number of tests. Frame Skew Pruning was applied in a couple of different ways:
- word exit - FSP was applied just at word exit states
- HMM exit - FSP was applied just at HMM exit states
- hmm/word exit - FSP was applied to both HMM and word exit states
Note that the current FSP implementation is an experiment and is not optimized for speed, but for ease of implementation. The implementation could be improved. TheAvg Beam Size column in the table below shows the average number of active states (tokens) per frame. This is probably the best guide to relative performance from the baseline, as opposed to RT.
| Test | WER | Avg Beam Size | RT |
| rm1_bigram - base | 4.20 | 1709 | 0.90 |
rm1_bigram - word exit - K1| 4.39 | 1556 | 0.86 |
| rm1_bigram - word exit - K2 | 4.66 | 1492 | 0.82 |
rm1_bigram - HMM exit- K1| 4.48 | 1545| 0.83 |
| rm1_bigram - HMM - K2 | 4.92 | 1470 | 0.84 |
rm1_bigram - hmm/word exit K1 | 4.52 | 1544 | 0.86 |
| rm1_bigram - hmm/word exit K2 | 4.92 | 1468 | 0.81 |
rm1_bigram - hmm/word exit K3 | 5.36| 1427 | 0.81 |
| hub4_trigram -base | 14.3| 40000 | 25 |
| hub4_trigram -word exit K2 | 15.12 | 39000 | 22 |
Conclusions
A modest improvement is seen here in terms of average beam size for RM1. The saving with the hub4 task, however is probably not worth the effort. Using standard beam pruning will achieve better results.
– Main.PaulLamere - 12 Nov 2003
