Our first feature vector

A key step in parameterising speech is to move from the time domain to a domain in which distances make more sense, and so where we can perform pattern matching.

There is no video here. Revise the following concepts from earlier in the course:

The spectrum, obtained by Fourier transform, is one possibility for a feature vector
A better option would be filterbank features, a bit like the cochlea produces
Yet another option could be the filter co-efficients of a source filter model

For now, you can imagine that any of the above are placed in the feature vector for each frame of speech. We’re going to come up with something better later on, but any of the above will be OK for the time being.