I built this to explore whether you can pull clinically meaningful signal out of a cough or voice recording using only acoustic features — no deep learning, no black box.
Upload a WAV or MP3 file and it extracts the features that show up in respiratory research:
- Jitter and Shimmer — cycle-to-cycle instability in pitch and amplitude
- Spectral Centroid — where the energy sits in the frequency spectrum
- Zero Crossing Rate — separates voiced sound from turbulent airflow
- Pitch tracking (F0) — fundamental frequency over time
It outputs a waveform, spectrogram, pitch contour, and an irregularity score showing how far the sample deviates from stable phonation.
pip install -r requirements.txt
streamlit run app.pyUpload your own file or use the included Coswara samples.
Built on real recordings from the Coswara dataset — clinical audio collected at IISc Bangalore with healthy and COVID-positive subjects across cough, breathing, and vowel tasks. 67 processed samples are included. Full dataset at the link above.
librosa · streamlit · matplotlib
Functional for feature extraction and visualization. Includes a simple rule-based audio-type suggestion (Cough / Voice / Breathing) as a first step toward a classification layer.
Not a diagnostic tool. Research and learning only.