even less aggressive silence filtering and more data
1 minute read
Tensorboard data can be found here. This version used even less aggressive silence filtering (set to aggressiveness value to 0 this time) as well as data from two additional books in the anne of green gables series:
I had originally missed those books since they didn’t start with “Anne”
This model is definitely an improvement from the previous one, being much more consistent in its output. It is also more sensitive to punctuation, comparitively.
The best example of this model at work is probably its attempt at reading a page from Harry Potter and the Order of the Phoenix as part of the entire autoread system (camera and ocr included).
I modified the aggressiveness from 3 to 1. Training history can be found here for a run which used the LJSpeech mode I trained, and here for a model trained ...
Cheryl and I spent some time taking a bit of a break from training models to try and figure out what it is. One big thing could be an issue with long sentenc...
I decided to try yet another dataset, this time the Blizzard dataset. The tensorboard results can be seen here. With this dataset, it is again successful at ...
I suspect the most significant issue is the forced alignment being slightly off on two accounts:
It might sometimes just be plain wrong
It is common for...