I use HTK to train an acoustic model. My last step is splitting mixtures of the phone gaussians. Normally, I always split all phones (their inner states) in one step by one, then re-estimate and stop when the performance drops.
Now I want to try out splitting the phones one by one because this should lead to equal or better overall result. The way I do it is, try to split every phone, pick the one that led to the best result, keep it split, reset all others, and start over. This takes too long though. I thought of splitting all of those that brought an improvement, not just the best one, and then go to next iteration.
My question is: If splitting a phone lowers the performance, is there any point in trying to split it again at a later stage? Or can I just blacklist it and just try with those that brought an improvement in the last iteration?
Improvement from such schemes are usually tiny. You can get much better improvement simply moving to DNN (supported by HTK 3.5 by the way).
If splitting a phone lowers the performance, is there any point in trying to split it again at a later stage? Or can I just blacklist it and just try with those that brought an improvement in the last iteration?
You can blacklist