| Ideas for accelerating PTM computation | |
| -------------------------------------- | |
| First thing to note is that codebook computation now takes up the | |
| majority of the time spent evaluating PTMs. So speeding up Gaussian | |
| evaluation is suddenly important again. | |
| Using a tighter top-N beam will speed up Gaussian computation by | |
| imposing a higher floor on densities, but this effect isn't worth a | |
| whole lot, in contrast to SC models where mixture computation rather | |
| than density computation is the most expensive part. | |
| This means that we should probably bring back kd-trees, although the | |
| implementation should be tweaked to be faster loading. | |
| Also, maybe more importantly, we can do some form of CI-GMM selection | |
| on the codebooks. This won't actually work with the way the models | |
| are set up currently since the CI phones share the same codebook as | |
| the CD ones, and the goal is to prune codebooks rather than phones. | |
| But wait! It's okay, because we still have the same top-N mechanism | |
| as before. We can use those top-N scores to do early pruning of | |
| entire codebooks. This ought to give us the most bang for the buck. | |