Having just fixed one stupid bug in Nitro (the credit allocation was confusing some peer devices like the Nokia 3650) and being a bit frustrated by another bug (reception from a Nokia 6310i fails, at least it's reproducible), I thought it was time for something different...
The idea of having a Newton that is at least capable of recognizing a small set of trained spoken words has haunted me long enough to give it a more serious shot now. The basic idea is to train a couple of dozens words and use them to control applications. There are quite many factors that could make this impossible, but anyway, how can you know when you haven't tried?
The basis for the recognition process is to filter and segment the incoming audio data. The first idea is to filter into ten frequency bands using FFT and record changes in these bands over time. Some trials using a very simple test program looked already quite interesting. At first, the data is confusing, but plotting it shows some patterns...
Posted by Eckhart at May 13, 2003 10:33 PM