Privacy-Aware Acoustic Assessments of Everyday Life Additional Material

Bitzer J, Kissner S, Holube I (2016). Privacy-Aware Acoustic Assessments of Everyday Life. Journal of the Audio Engineering Society, 64(6), 395-404. Download the OpenAccess paper from the Journal of the Audio Engineering Society.

Content

  1. Matlab Framework
  2. Smartphone feature extraction
  3. Additional Figures
  4. Erratum

Matlab Framework

Please clone the Git repository. In order to start with the framework, go to matlab/Paper/ and read README.md.

The code to produce Figures 3 to 7 (system evaluation) can be found in matlab/evaluation/.

Smartphone feature extraction

java/ contains the feature extraction algorithms used on the Android system. See Sec. 2 of the paper for details.

Additional Figures

As stated in the paper far more features have been computed to compare the smartphone based feature extraction to the conventional audio based extraction methods. Here are the final results:

The delta derivates of the MFCC coefficients

Delta values for the first MFCC
Delta values for the second MFCC
Delta values for the third MFCC
Delta values for the fourth MFCC

The standard deviation of the MFCC coefficients

Standard deviations of the first MFCC
Standard deviations of the second MFCC
Standard deviations of the third MFCC
Standard deviations of the fourth MFCC

The centroid results

Centroid of the spectrum
Standard deviations of the Centroid

The power spectrum entropy (PSE) results

PSE
Standard deviations of the PSE
Delta of PSE

The Broadband Envelope (Frequency Domain) (BEF) results

BEF
Delta of BEF
Standard deviations of the BEF

The Broadband Envelope Correlation and Lag (Frequency Domain) results

Correlation value
Lag of the maximum correlation

The Broadband Envelope Correlation and Lag (Time Domain, RMS-based) results

Correlation value
Lag of the maximum correlation

Erratum

Fig. 4 in the article shows the noise level exhibited by the microphones as a function of frequency. The ordinate is specified as Noise Level in dB SPL. This was calculated by appying the broadband calibration to every bin of the power spectral density of the measured noise. This is wrong for two reasons:

(Note that relative levels shown in the published figure are still valid.)

The figures below show the noise level in dB SPL, properly scaled, in (third-)octaves. The code for producing these figures can be found in matlab/evaluation/analyze_noise_eratum.m

The recording system's noise level in 1/3-octaves.
The recording system's noise level in octaves.