Exercise Feb 7: Speech preprocessing and a phonotopic map
Extract suitable features from raw digitized speech waveform data
and form a phonotopic map where each location represents one type
of a phoneme.
The mapping is fairly robust and straight-forward task,
if you apply Matlab SOM Toolbox
and the functions som_make, som_autolabel and
To extract the features use frequency analysis as in speech recognition.
E.g., apply FFT to short-time (Hamming) windows,
logarithmize the absolute value of the spectra,
and average the power on suitable frequency bands.
Description of the data is given in speech.des
Return a labeled SOM where the organization of phonemes
is clearly visible and your script to produce it
(Hint: Matlab has the basic preprocessing functions, too,
so use it for everything).