My proposal for a Thesis Proposal..

So far, all I know is that it is going to be on recognizing speech in the presense of music...

What I know of the problem:

• Music does not have as catastrophic an effect on recognition performance as noise does.
• Regular algorithms meant to handle speech in noise fail on speech in music
• The problems with music seem to be, variously, the harmonicity, the presence of transients and the colour (not necessarily in that order).
• The `nice' properties of music, that could potentially be taken advantage of, are the systematic nature (I dont call Heavy Metal music) and stationarity in parts.

Possible approaches:

• Characterize music with complex distributions; attempt regular ML/MAP/MMI algorithms on the model.
• Attempt to use the obvious spectral differences between music and speech to isolate part of the effects of music
• Different acoustic paramters for modelling speech in music??

Ideas, suggestions welcome. If you are within a 100 mile radius of Pittsburgh I'll even treat you to beer if you have a good suggestion for me.