Voice Recognition - MFCC

Question

I have been researching about Speech Recognition and I have decided to take the MFCC approach to solving this problem of detecting whether someone is saying either "Yes" or "No". So as mentioned before (My steps so far):

• Read in Audio File
• Split the Audio Signal into blocks (600 samples, 30msec long)
• Strip the blocks that do not warrant consideration (Total Energy / Zero-crossing)

So I am going to construct the MFCC based on this paper (http://arxiv.org/pdf/1003.4083.pdf) and it has the following steps:

1) Pre–emphasis

2) Framing

3) Hamming Windowing

4) FFT

5) Mel Filter Bank Processing

6) Discrete Cosine Transform

7) Delta Energy and Delta Spectrum

This makes sense to me (Kind of) and I am going to research into each of these steps. BUT should I perform the MFCC on the resulting blocks that I have already done with (steps 1, 2, 3) at the top of this question, or, should I not carry out these steps and just start from the beginning and compute the MFCC and will I still be able to implement a Hidden Markov Model?

The other question is, if I split the signal into "Frames" (2D vector) will the resulting MFCC be a 2D vector, or a 1D vector?

Hope someone can help :)!

Answer 1

Regarding your existing steps:

• Do you have some overlap between adjacent blocks? It is common in feature extraction systems to have some overlap between adjacent blocks, so that a short transient event can be correctly captured (if it is right at the end of a block, it will be right in the middle of the next one).
• Your idea of discarding blocks is dangerous. You can do this to roughly identify the beginning and end of a speech segment; but silence recognition can also be built into the recognition model.

Regarding the data size:

Your signal is split into frames. For each frame, you compute a 1-D MFCC vector. So in the end, you have a sequence of 1-D MFCC vectors. This is the data on which you train your HMMs.