Comparing two speech sounds

Question

I need to be able to determine if two sounds are very similar. The goal is to have a very limited vocabulary (10 or 15) of short one or two syllable words, then compare a captured sound to determine if it is one of those items with all the usual variability in environmental and capture conditions. The idea is that the user can issue a few simple commands by voice instead of keyboard or mouse.

Does anyone know the best approach to this? I don't want to do full blown speech recognition, just something much more limited.

Answer 1

I'd reconsider using a speech recognition library... like CMU's Sphinx software or Microsoft's speech recognizer. Unfortunately, its not a simple task to do this on your own. One approach that is somewhat typical of how to accomplish what you are trying to do is as follows:

1) Chop the sample into small segments (a few miliseconds)

2) Fourier transform on each segment, collect the principal coefficients

3) use a Hidden Markov Model to figure out the likely transition of phonemes given your sequence of coefficients

4) map to a dictionary depicting phonemes to words (you could look at the Sphinx dictionary as a guide)... a small set like yours should produce excellent results.

---

If you wanted to simplify this somewhat, you might try taking the coefficients at specific timesteps and feeding them into an SVM or neural network... I haven't tried this yet but bet you could get reasonable results with some tuning.