14.5.1 Introduction

HARK provides static feature extraction modules such as MSLSExtraction and MFCCExtraction . HARK also provides following additional feature extraction modules and preprocessing modules:

  1. Dynamic feature (Delta term): The temporal change of the static features. Notated as $\Delta $MSLS in Tab. 14.13. Calculated by Delta .

  2. Power: The power of the input signal. Notated as Power in Tab. 14.13. Calculated by PowerCalcForMap .

  3. Delta power: The temporal change of power. Notated as $\Delta $Power in Tab. 14.13. Calculated by Delta 

  4. Preprocessing: Emphasizing of the high frequency range using PreEmphasis or mean normalization. Notated as Preprocessing in Tab. 14.13.

We provide six sample networks of acoustic feature extraction as shown in Tab. 14.13. From the left, the three columns of the table denotes: the network file name, the feature description, and the feature file name generated by the network file. You can run the samples by executing demo.sh. For example, if you want to execute demo1.n, run

> ./demo.sh 1

The samples calculate the features using the MSLS of 13 dimensions in offline. If you want to calculate features in online, replace AudioStreamFromWave with AudioStreamFromMic . If you want to use MFCC instead of MSLS, replace MSLSExtraction with MFCCExtraction . If you want to change the dimension of the MSLS, change the property. See HARK document for details.

Table 14.13: Sample files of feature extraction
 

Feature

 

Network

MSLS  

$\Delta $ MSLS

Power

$\Delta $Power

Preprocessing

Corresponding

Generated file

file name

13 dim

13 dim

1 dim

1 dim

no dim

section

 

demo1.n

Yes

       

14.5.2

MFBANK13_0.spec

demo2.n

Yes

Yes

     

14.5.3

MFBANK26_0.spec

demo3.n

Yes

 

Yes

   

14.5.4

MFBANK14_0.spec

demo4.n

Yes

Yes

Yes

Yes

 

14.5.5

MFBANK28_0.spec

demo5.n

Yes

Yes

 

Yes

 

14.5.6

MFBANK27_0.spec

demo6.n

Yes

Yes

 

Yes

Yes

14.5.7

MFBANK27p_0.spec