MFCC

C/C++ code to extract MFCC or FBank features from wav files.

masterCPLus should be used. The mater branch may not be updated in time.

Install

1. Download following code from my GitHub and put these three directories under the same root directory.
   1. https://github.com/hangtingchen/BasicAudioToolBox
   2. https://github.com/hangtingchen/MFCC
   3. https://github.com/hangtingchen/inih
2. Switch the branch from master to masterCPlus in BasicAudioToolBox and MFCC. (The master branch is okay but it is highly recommended to use masterCPlus)
3. For Linux, use make in MFCC directory to construct the program. The target is mfcc.
4. For Windows,
   1. Add these files to your visual studio.
   2. Add path of BasicAudioToolBox and inih to VC++ catalog.
   3. Add _CRT_SECURE_NO_WARNINGS to the preprocessor definition.
   4. Enable openmp.(if you don't want to use multiple threads, ignore this step)
   5. Generate the exe.

Example

Run an example first. Enter MFCC directory.

mfcc example/config.ini

The screen will display followings,

finish reading the config.ini
FFT passband 2 to 1024 out of 1 to 1024
Mel passband 15.986084 to 3923.357581
Thread ID 1
Convert example/a001_0_30.wav to example/a001_0_30.fbank
including : 
MFCCNum 40
energyFlag 1
zeroCrossingFlag 1
brightFlag 1
subBandEFlag 8
the frame feature dimension is 408
Sample Rate : 44100
Number of channels : 2
Each sample's size in byte : 3
Each container's size in byte : 3
Number of samples : 1323001
total coef size: 612000
post-processing...
writing the doc...

These message indicts following steps to extract MFCC,

1. The program will first read config.ini.
2. The Mel filter bands will be generated according to the setting in config.
3. Read WAV file.
4. Count and extract the features.
5. Write the feature file.

Usage

You should prepare WAV files first(.wav), then a config file and a list. The following introduces the specific format.

WAV

The WAV should be PCM encoding and has a standard 44-byte head. (It's okay if there exists additional information chunk between 44-byte head and data chunk.) Don't worry about this, for most of WAV files are satisfied. If the wav is transformed from mp3 or other format, have a look at the head first.

You should check WAV if No new information displayed in the screen for a long time while running the program.

config

The config is read using inih(https://github.com/benhoyt/inih). A standard config is example/config.ini under MFCC directory. The options listed should all be included. No promise for the program if you skip some options.

Website(https://github.com/hangtingchen/inih/blob/master/examples/test.ini) gives details about inih.

Frame section

This section sets options related to pre-process.

| Key | Value | Notes | | :------| :------ | :------ | | sampleRate | 8000/16000/44100/others | The sample rate should be set at first, which means the wav files should have same sample rate. | | lowpassfre | >0 && <hipassfre && <sampleRate | The min frequency | | hipassfre | >0 && >lowpassfre && <=sampleRate | The max frequency | | preemphasise | =0 no preemphasise; 0-1 preemphasise the signal | The coefficient of preemphasise | | zeroMeanSigFlag | 0/1 | whether to make input signal have a zero mean | | wlen | wlen=(wlenInTime(ms))*sampleRate/1000 | the number of samples of window | | inc | inc=(incInTime(ms))*sampleRate/1000 | the number of samples of window shift | | vecNum | =1(mono/double channel WAV);=2(double channel WAV);=4(double channel WAV) | the channels of the output feature |

MFCC section

This section controls the Mel filter settings.

| Key | Value | Notes | | :------| :------ | :------ | | fbankFlag | 0/1 | Extract MFCC or fbank | | bankNum | >0 | The number of Mel filters | | MFCCNum | >0 && <=bankNum | The number of MFCC; If fbankFlag=1, this option has no effect | | MFCC0thFlag | 0/1 | Whether to include MFCC0th; If fbankFlag=1, this option has no effect |

Others section

This section controls some other features.

| Key | Value | Notes | | :------| :------ | :------ | | energyFlag | 0/1 | Whether to include average energy | | zeroCrossingFlag | 0/1 | Whether to include average zero crossing rate | | brightFlag | 0/1 | Whether to include brightness | | subBandEFlag | =0, No subBand energy; >0 set the number of subband | | | fftLength | >=0 | Output fft. This is only for debugging. |

Regression section

This section controls post-process.

| Key | Value | Notes | | :------| :------ | :------ | | znormFlag | 0/1 | Whether to do z-norm in each dimension within the single audio | | regreOrder | =1,no diff;=2 first order diff; and so on | (The degree of diff) + 1 | | delwin | >0 | The context length of diff; If regreOrder=1, this option has no effect |

IO section

This section controls how to read file and store features.

| Key | Value | Notes | | :------| :------ | :------ | | fileList | | The position of list of files | | saveType | f/e/n/b | csv(double)/csv(scientific)/npy(numpy)/binary | | numThreads | >0 | The number of threads.(Only when openmp is supported.) |

file list

Each line in the list should include the source WAV and target feature, separated by tab. Please refer to example/fileList.txt.

Notes

• The main interface is MFCCWapper in mfcc.c in case you want to extract feature in your own code or you don't want to control so many options.
• The program has set up hmath and hsigProcess according to HTK.
• Recommend to use masterCPlus branch.
• The anomaly detection is not perfect. Please pay attention to your config and file list.
• Though the code in masterCPlus is C++, the style is still C,and so is the memory application and release.