Speech Recognition can categorize spoken words well, but what features of the time frequency characteristics of speech are most useful in making discrimination between words/speech sounds on average? Are there any papers out there that show what effective speech systems are 'attending to' in the signal? I'd really like to see a 2d (imagesc style) plot where the x-axis is word onset to 1 second or so, and the y-axis is frequency, and colors in the plot are feature importance. Perhaps this exact plot doesn't exist or isn't practical because features are happening at very different times, but it's usually helpful to give a concrete example of what I'm looking for. I'd like to learn why this isn't practical if so as well.
[link][2 comments]