-1
$\begingroup$

I think female and male have total different MFCC feature for same word such as "one", but why the MFCC could be used for isolated word recognition which use the same HMM model for the same word without any consideration of gender?

$\endgroup$
1
$\begingroup$

Not only do female and male speakers produce different MFCCs, but each speaker will produce quite different MFCCs for the same word, depending on pitch, vocal tract, accent, and many more factors. The important thing is not that MFCCs be independent of gender or any other feature, but that the trained models are independent of that feature. And that is a matter of the training process. I.e., if you use a large number of female and male speakers to train the models, then you will end up with a gender independent model.

$\endgroup$
  • $\begingroup$ As the MFCC is gender dependent but the trained model is geneder independent, so there must be some feature inside the MFCC is gender independent that the model just extracts the feature from MFCC. So, what's the gender independent feature inside MFCC? $\endgroup$ – weifeng Dec 28 '18 at 8:52
  • $\begingroup$ @weifeng: Well, everything that the training sequences of a specific word have in common with each other, as opposed to the training sequences of a different word. And that's a lot, apart from the fundamental frequency. So you hope that the utterances of the word 'one' of a female speaker from Newcastle and a male speaker from Denver have more in common than utterances of the word 'two'. $\endgroup$ – Matt L. Dec 28 '18 at 9:16
0
$\begingroup$

The main difference between men and women voices is the fundamental frequency (f0). However what allows you to perceive a given phoneme is distinct of fundamental frequency. In vowels, for example, the formant frequencies can be very important to distinguish between different sounds. MFCCs, due to the use of filter banks, are used to capture the energy in certain frequency ranges (which introduces some "flexibility" for specific frequency values). If you apply HMM on top of this, which have statistical base, and have sufficient data for training, you will have a quite powerful model even if some frequencies can show slight variations.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.