I am not sure if this gives a complete answer to your query, but I would
like to put forward some of my views.
I feel cepstrum coeffcients are best features for discriminating speakers.
A techniques to reduce the feed for an NN is vector Quantise the cepstrum
features and combining them with Formant frequencies.
Hope this helps..
PS: Anyone: Please correct me if I am wrong.
> hi all ,
> I am a new member in this group and I work on SR for more than
> one year , now I want to use a neural net (MLP) for isolated word
> recognition in small size vocabulary that is user dependant , but I
> do not know which features are the best and disciminant and have less
> amount to feed to NN . if anyone know anything about this I will be
> glad to hear about it , and I will be glad if I help anyone in SR .
> scincerely Parisa >
Nandyala S Nagesh.
Ph +91 80 6585027/9 Ex 32