0
$\begingroup$

I am currently trying to understand what is involved to train a Hidden Markov Model (HMM) with Forced alignment. Forced alignment, as far I understand, is to align the audio file with the utterance at a word level or phonetic level.

What I don't understand is how a HMM is being trained by forced alignment.

Normally the emission probability would be given by the gmm described by the MFCC features, but what is the emission probability here?

$\endgroup$
2
$\begingroup$

For better understanding of the complex subject it is better to read the book instead of random sources on the web.

A book like Rabiner, Juang. Fundamentals of Speech Recognition will give you much more details.

In short, HMM training is iterative process and forced Viterbi alignment is only a part of it. First GMM distributions are initialized with a uniform values, then alignment aligns audio features to HMM states, then GMM distribution parameters are updated based on alignment. The process is repeated many times. You align, update GMM and align again. The process converges to a proper model after several iterations.

$\endgroup$

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.