Week 4
Group meeting
For week 3 we had our third group meeting and talked about cataloging the CHILDES collection. Again, there are a lot of different corpora and perhaps too many for our speech recognition system. Additionally, Dr. Brumberg brought in one of his graduate students, Shadi to come listen. Her work is in line with our current project and is also a mix between EECS and SPLH. I think she’ll be a really interesting person to talk to as we make progress.
Readings
Automated Assessment of Child Vocalization Development Using LENA, Large Vocabulary Automatic Speech Recognition for Children, Markov Models for Pattern Recognition. I’ve also recently looked into the research surrounding the critical period hypothesis and other child language acquisition research.
Picking a framework
I initially wanted to learn how to use HTK, but after light experimenting with the different open source libraries I think Kaldi might be a better choice for learning how to build a speech recognition. HTK, Kaldi, and CMU Sphinx basically do the same thing (build speech recognition systems), but certainly they have different learning curves and niches. I think Kaldi is the more recent framework, which supports a lot of new technologies (RNN, DNN, etc.) and slightly more straightforward in their implementation. Still, the HTK handbook is really nice for getting an overview of the theoretical aspect of speech recognition.
For next week we plan on finishing the initial cataloging of the data and getting more comfortable with Kaldi.
Best,
EO