I am working on isolated word recognition using HMMs|
and have a few basic questions on HMMs. 1. I have to recognize digits using HMMs. Hence, it is
an Isolated Word Recognition.
2. Say the digits are 1-7. I am having 1 class per
digit i.e., 5 utterances per class.. How many HMMs
will I be obtaining? Will it be 1 per utterance (i.e.,
5 per class) or 1 per class (1 per digit)?
3. What should be the input to these HMMs. How will I
relate my signals (training signals) to the HMMs?
4. I have extracted features for each signal. They are
Mel-Cepstrum Coefficients. I have been told that these
features would be the input to the HMMs but in vector
notation. So, I applied VQ to these features. Now, I
have 1 sequence of vectors per signal. i.e., 7*55
sequences in all.
5. How will I train my HMM with these sequences? The
method I used is as follows:
6. I have assumed the transition probabilities and the
initial probability to build a random HMM. Later I
will update the probabilities with correct values.
7. According to Rabiners paper:
www.ai.mit.edu/~murphyk/Bayes/rabiner.pdf , once I
build a model by assuming the initial probabilities, I
have to extract an observation sequence with the model
built. Why should a new observation sequence be
extracted when I have all training sequences ready (by
8. I have not found the new sequence with the model
prepared. I continued with 1 sequence among the 35.
Applied FB method, Viterbi method and found the most
likely state sequence for the i/p observation sequence
9. I tried to update the model parameters by
re-estimating the values, but couldnt do it
successfully. Am trying to use the Baum-Welch method.
10. Say it is done, should I find the ML state
sequence for all the sequences ( in the same class )
using the updated model?? I am assuming that Ill have
1 model per word.
11. Again, once the training part is done, for the
testing part, Ill take a sample sequence, and find
the ML state sequence using the updated model
parameters. Is it right? Or should I find another
model for the testing sequence as well , find the ML
state sequence and compare the two models??
12. What probabilities should be compared or what
parameters of the HMMs should be compared to perform
Please reply at the earliest. I am out of time.