COVID 19 Detection From Speech Using Deep Learning
Overview:
You know how COVID 19 is dangerous and detection of covid is very critical if you have symptoms of COVID 19 it doesnāt mean that you have COVID 19. First, you have to do an RT PCR test. There are also some university and organization which is providing COVID 19 patient data to researchers. Many Data Scientists work on this problem. You can also find COVID 19 from Chest X-Ray but for final surety give priority to RT CPR or other valid tests of COVID 19.
IISC Banglore is working on a project called Coswara ( The word is an amalgamation of Co (from corona) and Swara (sound in Sanskrit) )
Do check out Full information about Coswara.
Under this Project, they are providing the Speech Data of COVID 19 Patients for research. Do check out Cosware Github hereā¦
About Dataset:
The dataset contains approximately 422 Patients data. Each patient must have a total of 9 Audio files and 1 metadata.json file which will give information about the COVID Status of that Particular Patient.
Audio Files Information:
(1) breathing-deep: Deep breathing voice file of the patient.
(2) breathing-shallow: Shallow breathing voice file of the patient.
(3) cough-heavy: Heavy cough voice file of the patient.
(4) cough-shallow: Shallow cough voice file of the patient.
(5) count-fast: Count 1 to 20 fast
(6) count-slow: Count 1 to 20 slow
(7) vowel-a: The voice file of the patient speaking vowel āaā.
(8) vowel-e: The voice file of the patient speaking vowel āeā.
(9) vowel-o: The voice file of the patient speaking vowel āoā.
JSON File Information:
Classes Information:
Here we have total of 7classes.
(1) healthy
(2) no_resp_illness_exposed:
(3) positive_asymp:
(4) positive_mild:
(5) positive_moderate:
(6) recovered_full:
(7) resp_illness_not_identified:
Coding Strategies:
I have tried 4 strategies for COVID 19 Status Classification.
(1) Basic Audio Features Extraction:
In this Strategy, I tried Basic Audio Features Extraction like RMSE, Chroma STFT, Spectral Centroid, Spectral Bandwidth, Spectral Rolloff, Zero Crossing Rate.
Information About Audio Features Extraction
Music Feature Extraction Medium
After making this data frame I reshape the data into 422 x 7 x 9. After this, I make one Deep Learning model which classifies into 7 classes.
Train Accuracy: 65.88 %
Test Accuracy: 70.59 %
(2) MFCCs Feature Extraction:
In this Strategy, I tried Basic MFCCs Feature Extraction. I extract 20 MFCCs Coefficient. In this Strategy
After that, I reshape the data into 422 x 9 x 21. Then I made another Deep learning model which classifies into 7 classes.
Train Accuracy: 66.77 %
Test Accuracy: 71.76 %
(3) Simple Features and MFCCs Feature Extraction:
In this Strategy, I combined both the above strategies. So now we have a total of 6 Simple Features and 20 MFCCs Coefficients.
After that, I reshape my data into 422 x 9 x 27. Then I made another Deep Learning model which classifies into 7 categories.
Train Accuracy: 66.77 %
Test Accuracy: 69.41 %
(4) MFCCs Features Extraction and CNN:
In this Strategy, I have tried to extract MFCCs features and then save them into JSON File and then I applied the Convolution Neural Network model.
Train Accuracy: 72.50 %
Test Accuracy: 73.77 %
Code:
For full code do check out My GitHub repositories.
Feel free to give ā to the repository.
Research Paper:
Do check out the Official Research Paper Launched by IISC.
The link is hereā¦
Conclusion:
I hope now you have an idea about how a thing works.
Try it by Yourselfā¦!!
If you have new strategies share them with us through comments/responses.
Github:
LinkedIn:
Thanks for reading! If you enjoyed this article, please hit the clap šbutton as many times as you can. It would mean a lot and encourage me to keep sharing my knowledge. If you like my content follow me on medium I will try to post as many blogs as I can.
