Source-linked AI summary
Coswara -- A Database of Breathing, Cough, and Voice Sounds for COVID-19 Diagnosis
Neeraj Sharma, Prashant Krishnan, Rohit Kumar, Shreyas Ramoji, Srikanth Raj Chetupalli, Nirmala R., Prasanta Kumar Ghosh, Sriram Ganapathy
TL;DR
RT-PCR testing is costly, time-consuming, requires expertise, violates social distancing, and is difficult to deploy at scale, motivating alternate COVID-19 diagnostics. The paper presents Coswara, an openly released crowdsourced database of breath, cough, and voice sounds analyzed with signal processing and machine learning. The released dataset contained 941 participants, while a nine-class acoustic classifier achieved 66.74% test accuracy; collection, release, and analysis remained ongoing.
Problem
RT-PCR testing is expensive, time-consuming, requires expertise, violates social distancing, and is difficult to deploy at large scale.
Method
Coswara collects anonymized breath, cough, speech, and metadata through a web application, then uses signal processing and machine learning to identify acoustic biomarkers.
Results
66.74% test accuracy was obtained for a nine-class classifier distinguishing the sound categories, with 941 participants in the released dataset.
Takeaways & Limitations
Coswara provides an open-access respiratory-sound database intended to support sound-based point-of-care COVID-19 diagnostic tools.
Takeaways & Limitations
Data collection, dataset release, and analysis were still in progress, and the diagnostic tool required clinical validation and healthcare-authority approval.
Abstract
from arXiv · showhide
The COVID-19 pandemic presents global challenges transcending boundaries of country, race, religion, and economy. The current gold standard method for COVID-19 detection is the reverse transcription polymerase chain reaction (RT-PCR) testing. However, this method is expensive, time-consuming, and violates social distancing. Also, as the pandemic is expected to stay for a while, there is a need for an alternate diagnosis tool which overcomes these limitations, and is deployable at a large scale. The prominent symptoms of COVID-19 include cough and breathing difficulties. We foresee that respiratory sounds, when analyzed using machine learning techniques, can provide useful insights, enabling the design of a diagnostic tool. Towards this, the paper presents an early effort in creating (and analyzing) a database, called Coswara, of respiratory sounds, namely, cough, breath, and voice. The sound samples are collected via worldwide crowdsourcing using a website application. The curated dataset is released as open access. As the pandemic is evolving, the data collection and analysis is a work in progress. We believe that insights from analysis of Coswara can be effective in enabling sound based technology solutions for point-of-care diagnosis of respiratory infection, and in the near future this can help to diagnose COVID-19.
1. Introduction
COVID-19 created an urgent need for diagnostic tools that avoid RT-PCR’s cost, time, expertise, social-distancing, and scaling constraints. The paper motivates using respiratory acoustics and introduces Coswara as an open-access sound database project.
- RT-PCR results take 2–48 hours and require chemical reagents, devices, expertise, and close-contact sampling, limiting inexpensive large-scale deployment.
- COVID-19 symptoms include dry cough, breathing difficulty, chest pain or pressure, and loss of speech or movement, appearing 2–14 days after exposure.
- Speech breathing patterns are tied to respiratory-system anatomy and physiology, motivating acoustics research for point-of-care diagnosis.
- Coswara is named for coronavirus and swara, the Sanskrit term for sound, and its dataset is released openly as collection and analysis continue.
2. Scientific Rationale
The scientific rationale links respiratory sound production to airway and vocal physiology, while prior studies suggest that cough, breathing, and speech acoustics can contain disease-related information.
- 2.1. Cough Sounds: Cough sounds arise from airflow turbulence, tissue vibration, and fluid movement, producing broadband spectra across multiple temporal phases.Each cough in the illustrated sequence lasts close to 300 ms and shows broad spectral spread.
- 2.1. Cough Sounds: Prior work reported good accuracy for preliminary cough-based COVID-19 detection, although the investigation involved a small number of subjects.
- 2.2. Breath Sounds: Breathing difficulty is a common COVID-19 symptom, and smartphone breath spectrograms have shown distinct patterns for asthmatic versus healthy individuals.
- 2.3. Voice sounds: Studies associate lung diseases with speech-breathing biomarkers, while vocal-fold pressure and laryngeal dysfunction relate to vocal fatigue and altered breathing patterns.
3. Coswara Overview
Coswara builds a low-cost COVID-19 diagnostic project around crowdsourced breath, cough, and speech recordings. Its staged pipeline collects anonymized data, models acoustic biomarkers, and targets a validated web or mobile screening tool.
- 3. Coswara Overview: Coswara targets a simple, cost-effective COVID-19 diagnostic tool using breath, cough, and speech sounds, organized into data collection, modeling, and diagnostic development.
- 3.1. Data collection: The dataset includes nine sound categories spanning shallow and deep breathing, shallow and heavy coughs, three sustained vowels, and normal and fast digit counting.
- 3.1. Data collection: Participants provide age, gender, location, health status, and comorbidity metadata; personally identifying information is not collected and data are anonymized during storage.
- 3.2. Modeling: Collected recordings are analyzed with signal processing and machine learning to identify acoustic biomarkers, with curated releases provided openly through GitHub.
- 3.3. Diagnostic tool development: The planned application prompts voice recording and returns a COVID-19 infection-probability score, pending clinical validation and healthcare-authority approval.
4. Dataset Description
Coswara is an ongoing, openly released crowdsourced database whose recordings are manually curated and analyzed across nine respiratory-sound categories. An initial random-forest experiment evaluates whether acoustic features distinguish these categories.
- The project remains ongoing, with data collection, release, and analysis still in progress.
- Worldwide crowdsourcing used a web application for metadata entry and microphone recordings, with users spending approximately 5–7 minutes per interaction.
- As of 7 August 2020, the dataset contained 941 participants, each contributing nine audio files spanning the sound categories.
- All recordings were sampled at 48 kHz and manually curated through an annotation interface that verified category labels and audio quality.
- The nine-class experiment represented each 500-millisecond segment with 28-D temporal and spectral features and trained a 30-tree random forest using a 70–30% train-test split.
- 66.74% test accuracy was obtained on balanced nine-class data, with less confusion among vowels and greater confusion among digit-counting, cough, and breathing samples.
5. Conclusion
The paper presents Coswara as a sound-based diagnostic-tool effort built on respiratory-sound research and a curated database. It reports ongoing work toward machine-learning classification of health conditions and identification of COVID-19 sound biomarkers.
- Coswara is described as a sound-based diagnostic tool for COVID-19 that builds on prior studies of respiratory disorders.
- The project uses different stimuli to capture complementary respiratory-sound information.
- The next development phase will apply machine learning to classify health conditions and identify sound-based biomarkers for COVID-19.