Covid.wav
Senior Project · 2026 Download brief →

COVID-19
Acoustic Detector

A cough carries a signature. We read it with a trained ear and return a screening verdict.

Run a screening
Team · CS Senior Design Spectral Classifier CNN / MFCC & Mel Spectrogram No hardware required

THE PROBLEM

The sound of a cough holds a measurement medicine still ignores.

Testing today is invasive, slow, and centralized. We translate a recorded cough into a mel-spectrogram, run it through a compact convolutional classifier, and return a real-time risk assessment with next-step guidance in seconds — on any smartphone or computer, with longitudinal health tracking for patients and medical professionals alike.

COVID-19
AUDITOR.

waveform · cough_017.wav

Raw samples, 16 km mono. Silence-trimmed and normalized.

spectrogram · mel-30ms

Frequency image fed to the CNN backbone.

Sound → Verdict

STFT · MEL · CNN · REPORT

(01)

Audio Upload

Record a cough in the browser with the Web Audio API — no specialized hardware — or drop a short .wav clip. It arrives, silence is trimmed, samples resampled to one fixed rate.

(02)

Feature Extraction

Slide a 30 ms frame, compute the mel spectrum and MFCC deltas — a shape doctor already reads.

(03)

Classifier + Verdict

The CNN returns a COVID / healthy distribution with confidence, mapped to a Low / Medium / High risk level with clinical guidance on next steps.

MODEL & DATA LAYER

Signal Preprocessing

MFCCs and mel-spectrograms, normalized and augmented with SpecAugment, pitch/time shifts, and additive noise — so the model hears the cough, not the room you recorded it in.

MFCC SPECTROGRAM SPECAUGMENT PITCH/TIME SHIFT NOISE

Feature + Model

A 2D CNN baseline fine-tuned with ResNet transfer learning in PyTorch, trained on the CoughVid dataset (~2,800 recordings).

CNN RESNET PYTORCH LIBROSA

Evaluation & Report

Accuracy, precision, recall, F1, and ROC-AUC on a held-out split with k-fold cross-validation — then a patient-readable report: verdict, confidence, and a referral note.

ACCURACY PRECISION RECALL F1 AUC K-FOLD CV

DELIVERABLE

The project plan, milestone by milestone.

Four phases across Fall 2026: data preprocessing and augmentation, deep-learning model development, evaluation, and full-stack web integration — a detection tool for medical professionals and potential COVID patients. One printable page, ready to save as a PDF.

Download Project Plan → Contact the team
M1 · SEP 28 · DATA M2 · OCT 26 · MODEL M3 · NOV 23 · SHIP

THE TEAM

Four hands, one breathprint.

Data pipeline, requirements, models, and frontend — we meet at the wave, not just the slide.

  • 01 · Data & Audio Pipeline
    Drew Quashie (dquashie2024@my.fit.edu)
  • 02 · Requirements & Docs
    Amanda Ogbonna (kogbonna2025@my.fit.edu)
  • 03 · Model Development
    Loleyi Oluwatomisin (ooluwatomisi2023@my.fit.edu)
  • 04 · Frontend · React
    Richard Alonso Garcia (ralonsogarci2023@my.fit.edu)
Team · CS 2026 @COVID.wav All rights, 2026
Faculty advisor · Dr. Zahra Nematzadeh — znematzadeh@fit.edu

Senior design · Fall 2026

Semester milestones

Milestone 1 · September 28

Planning · Data · Baseline
  • Requirements documentation and system architecture design
  • CoughVid dataset pipeline: filtering, MFCC / spectrogram features, normalization
  • Data augmentation: additive noise, pitch / time shifts, SpecAugment
  • Baseline PyTorch CNN and React audio-recording component

Milestone 2 · October 26

Training · Evaluation · Backend
  • Train and fine-tune CNN and ResNet transfer-learning models
  • Initial evaluation reports: F1 score, ROC-AUC
  • FastAPI backend endpoints for audio uploads and inference
  • React UI for prediction results and risk feedback

Milestone 3 · November 23

Integration · Tracking · Delivery
  • End-to-end real-time predictions across React, FastAPI, and PyTorch
  • Longitudinal cough-tracking visualization dashboard
  • Full system testing, performance optimization, and bug resolution
  • Final deliverables: demonstration video, poster, final presentation, and paper

Milestone 1 · Task Matrix

Full plan →

Amanda Ogbonna — Requirements & plan docs · Completed

Drew Quashie — Preprocessing & augmentation · In progress

Loleyi Oluwatomisin & Amanda Ogbonna — PyTorch CNN · Pending

Richard Alonso Garcia — React audio interface · Pending