We are in the process of obtaining sound activity annotations for every sound file in the COSWARA database via human listening. Below is some analysis of the 286 annotated sound files. An example of annotations is shown below. The blue shade corresponds to the human annotation. The red trace corresponds to intensity obtained from PRAAT.

File duration vs. Activity duration

Intensity

We compute the intensity over short-time segments (40 msec, and hop by 1 sample) using PRAAT. A higher inensity value indicates louder sound. We show comparison between segments annotated as activity and no activity segments, with 60 sound files (30 healthy and 30 COVID).

  • For the below plot we pooled annotated segments from all sound categories.

  • For the below plot we show a categoy wise cmparison. Also, we show two groups, healthy and COVID, for each sound category. The below plot is for comparison of activity.

  • Similar to above but comparison of noise (or background) activity.

Harmonic-to-noise ration (HNR) of segments

We compute HNR over short-time segments (40 msec, and hop by 1 sample) using PRAAT. A larger HNR value indicates large harmonic content. We show comparison between segments annotated as activity and no activity segments, with 60 sound files (30 healthy and 30 COVID).

  • For the below plot we pooled annotated segments from all sound categories.

  • For the below plot we show a categoy wise cmparison. Also, we show two groups, healthy and COVID, for each sound category. The below plot is for comparison of activity.

  • Similar to above but comparison of noise (or background) activity.

Zero-crossing rate (ZCR) of segments

We compute the intensity over short-time segments (25 msec, and hop by 10 msec) using Librosa. This should lie between 0 and 1 (indicating the fraction of zero-crossings in the time window). We show comparison between segments annotated as activity and no activity segments, with 60 sound files (30 healthy and 30 COVID).

  • For the below plot we pooled annotated segments from all sound categories.

  • For the below plot we show a categoy wise cmparison. Also, we show two groups, healthy and COVID, for each sound category. The below plot is for comparison of activity.

  • Similar to above but comparison of noise (or background) activity.

Spectral Flatness of segments

We compute the intensity over short-time segments (25 msec, and hop by 10 msec) using Librosa. This should lie between 0 and 1 (closer to 1 indicating a flat spectrum). We show comparison between segments annotated as activity and no activity segments, with 60 sound files (30 healthy and 30 COVID).

  • For the below plot we pooled annotated segments from all sound categories.

  • For the below plot we show a categoy wise cmparison. Also, we show two groups, healthy and COVID, for each sound category. The below plot is for comparison of activity.

  • Similar to above but comparison of noise (or background) activity.

Intensity overlaid with human annotations

Sound Category Healthy COVID-19
cough shallow Click here Click here
cough heavy Click here Click here
vowel-o Click here Click here
vowel-e Click here Click here
vowel-a Click here Click here
breathing-shallow Click here Click here
breathing-deep Click here Click here
counting-normal Click here Click here
counting-fast Click here Click here

Classifying actvity and no activity regions

Using GMM

Using Random Forest

Using Logitic Regression

Spectrograms

  • Narrowband
Sound Category Healthy COVID-19
cough shallow Click here Click here
cough heavy Click here Click here
vowel-o Click here Click here
vowel-e Click here Click here
vowel-a Click here Click here
breathing-shallow Click here Click here
breathing-deep Click here Click here
counting-normal Click here Click here
counting-fast Click here Click here
  • Wideband
Sound Category Healthy COVID-19
cough shallow Click here Click here
cough heavy Click here Click here
vowel-o Click here Click here
vowel-e Click here Click here
vowel-a Click here Click here
breathing-shallow Click here Click here
breathing-deep Click here Click here
counting-normal Click here Click here
counting-fast Click here Click here

Harmonic to noise ratio, overlaid with human annotations

Sound Category Healthy COVID-19
cough shallow Click here Click here
cough heavy Click here Click here
vowel-o Click here Click here
vowel-e Click here Click here
vowel-a Click here Click here
breathing-shallow Click here Click here
breathing-deep Click here Click here
counting-normal Click here Click here
counting-fast Click here Click here