|
The most popular features used in speech recognition systems are Mel Frequency Cepstral Coefficients (MFCCs). In this portion, we focus on converting our audio files into feature vectors. To do this we use a built in function in HTK, HCopy. Assuming you are using the data set, TI Digits Short, you already have all of your audio files. If you are not, and are recording your own data, use a program like Audacity and save them as either a .wav or .raw file. Procedure
1. We use the files train_list.scp and test_list.scp to indicate which files we are going to
convert into our features and the corresponding file name of the soon to be made .mfc feature files.
You should check your .scp files to make sure you are using the correct directories for both your raw
and feature files. We also need a configuration file, wav_config, to determine how to extract
the features. If you open wav_config you can see some of the different parameters that can be changed
depending on the experiment. For example you may need to change "SOURCE FORMAT," "SOURCE RATE," or
"TARGETKIND" depending on what type of data you use. Make sure that SOURCE RATE = 10^7/(Data Sampling
Rate). For the case of TI Digits Short the sampling rate is 8kHz so our source rate should be 1250.
Also be sure to check that in the directory isip/exp/htk_tutorial/data_preparation/data/train/raw
contains all your audio files and isip/exp/htk_tutorial/data_preparation/data/train/mfcc is an empty
directory. Check the raw files and mfcc directory for the test directory (i.e.
isip/exp/htk_tutorial/data_preparation/data/test) as well. Finally, you should check your .scp files
to make sure you are using the correct directories (otherwise your mfcc's will be saved elsewhere)
|