Alternative to CMS for Short Utterances

[Raeldor]

Hi All,

Does anyone know if there is an alternative to CMS (or CMN) for short
utterances. I have been doing some testing, and it seems to work great for
sections of 2+ seconds in length, but for shorter 0.5-1 second utterances it
really distorts the signal a lot.

Thanks
Ray

[Raeldor]

Actually, I'm starting to wonder now if I'm doing my CMN correctly. I have the
mean calculated individually for each co-efficient over time, is that correct?
So, my (pseudo) code reads...

loop through each mfcc (1-12)
mean = 0
loop through each frame
mean = mean + {mfcc value for frame}
end loop
mean = mean / {frame count}
loop through each frame
{mfcc value for frame} = {mfccvalue for frame} - mean
end loop
end loop

should it really be...

mean=0
loop through each mfcc (1-12)
loop through each frame
mean = mean + {mfcc value for frame}
end loop
end loop
loop through each mfcc (1-12)
loop through each frame
{mfcc value for frame} = {mfcc value for frame} - mean
end loop
end loop

Otherwise it doesn't seem to make sense that i'm losing the relative strength
of each mfcc value to another within the same frame.

Thanks
Ray

[Nickolay V. Shmyrev]

So, my (pseudo) code reads...

This is correct

should it really be...

No

You'll be interested to read the following papers:

Reducing The Effects Of Linear Channel Distortion On Continuous Speech
Recognition (1996)
by Rebecca Anne Bates ,  Dr. Mari Ostendorf ,  Associate Professor ,  Dr. J.
Robin Rohlicek ,  Vice President ,  Dr. William ,  C. Karl ,  Assistant
Professor

http://citeseerx.ist.psu.edu/viewdoc/summary?doi=10.1.1.22.9450

and

The Use Of Cepstral Means In Conversational Speech Recognition (1997)
by Martin Westphal

http://citeseerx.ist.psu.edu/viewdoc/summary?doi=10.1.1.31.2342

[Raeldor]

Thank you very much for your reply! So here's my problem. If I am working on a
simple utterance of the word 'ah!', it's very heavy in MFCC (a value of around
28 in all frames, but quite low (between -10 to +2) in all other co-efficients
in all other frames. If I use this technique on each MFCC individually, I lose
the information about the relative strength of each MFCC to each other.

Is this a common problem, or am I going crazy? Is there another technique for
using on this kind of short, single vowel phrase?

Thank you!

[Nickolay V. Shmyrev]

If I use this technique on each MFCC individually, I lose the information
about the relative strength of each MFCC to each other.

This is not a problem. The problem is to estimate cepstrum mean for a
speaker/channel using just a short sample. Please read the papers first.

[Raeldor]

Aaahh... now I see the quote in the second paper. So by that design, couldn't
I keep a running average of the cepstral means for the last 'n' utterances
until 'n' reaches a point of diminishing returns?

[Nickolay V. Shmyrev]

So by that design, couldn't I keep a running average of the cepstral means
for the last 'n' utterances until 'n' reaches a point of diminishing returns?

This is just one of the methods, it's better than default though.