Luise - DREU Research Intern
This is it. I can’t believe I just finished my last day here in the Speech Lab at Columbia University. It has been such a great summer. I have learned so much - both specific to speech recognition, and more general to computer science - while also having a lot of fun both at work, in NYC and around the country.
This week I did everything I could to finish up the last bunch recognizers. As I mentioned last week, each recognizer requires a lot of time for training, so I had to put in a lot of hours from home in order to finish on time - or at least try to. The recognizers I have trained this week were really the most interesting of them all. Instead of just using the standard corpora for training data, I have been synthesizing utterances from blogs and movie subtitles and included these in our the training sets. I have not been able to find any papers on previous research that did the same thing, so we were all very excited to see if it would be helpful.
Unfortunately the synthesized blog utterances did not improve the performance of our recognizers. I will finish up the recognizers with synthesized subtitles in training set when I am in the airport tonight. Hopefully that will give better results! However, in any case, bad results now don’t mean that the idea itself is bad. There is a lot more options to try out: different synthesized voices, “better voices” (whatever that would be - a question for research), different kinds of text data to synthesize (e.g. online forums) and much more. I could easily spend 10 more weeks working on all of this and much more :)