TinyML Cough Detection Wearable

Cough frequency is a useful signal in respiratory studies, but counting coughs today means either manual tallying or shipping audio to the cloud — expensive, power-hungry and a privacy problem. TinyML changes the trade: a microcontroller can run a small neural network on-device, so the wearable decides locally and only the event (timestamp + confidence) leaves the device. This project builds that wearable. An ESP32-S3 board with an INMP441 MEMS microphone samples audio in short windows, computes mel-scale spectral features, and runs a TensorFlow Lite Micro CNN trained on cough-vs-non-cough audio (built from the public COUGHVID corpus plus custom-collected samples). A majority-vote smoother turns window labels into cough events, which are logged to flash and synced over BLE to a phone dashboard with the day's count, an hourly chart and a confidence-filtered event log. The report is explicit about the model's limits: it counts cough-like sounds, it does not diagnose anything, and accuracy is measured on the student's own test set during the build.

1. The INMP441 microphone streams audio to the ESP32-S3 over I2S in short overlapping windows.
2. Firmware computes mel-scale spectral features for each window — the same feature recipe used in training.
3. The quantized TFLite Micro CNN scores each window as cough or non-cough; scores below the confidence threshold are discarded.
4. A majority-vote smoother across consecutive windows merges sustained detections into single cough events with timestamps.
5. Events are written to on-flash storage and advertised over BLE; the phone dashboard pulls new events on connect.
6. The dashboard renders today's count, the hourly distribution chart and the confidence-filtered event log.
7. Training happens offline: cough and non-cough clips are curated, augmented, and used to train and quantize the model that is flashed to the device.

Tech stack:

• ESP32-S3 (TinyML-capable, BLE on board) with Arduino/ESP-IDF firmware
• INMP441 MEMS I2S microphone
• TensorFlow Lite Micro (quantized CNN, on-device inference)
• Python training pipeline (feature extraction, augmentation, training, quantization)
• COUGHVID public corpus + custom-collected samples for training data
• Phone web dashboard over BLE (count, hourly chart, event log)

• [On-device cough classifier] TFLite Micro CNN on the ESP32-S3 labels short audio windows as cough or non-cough locally — raw audio never leaves the device.
• [MEMS microphone front end] INMP441 I2S microphone with configurable gain; firmware documents the sampling setup and wind-noise behaviour.
• [Spectral feature pipeline] Mel-scale spectrogram features computed on-device from each audio window before inference.
• [Event smoothing] Majority-vote window over consecutive labels converts raw detections into discrete cough events with timestamps.
• [BLE phone dashboard] Cough count today, hourly bar chart, event log with confidence scores, and device battery level over Bluetooth Low Energy.
• [On-flash event log] Events persist in flash across power cycles and sync to the phone when it reconnects.
• [Privacy by design] No audio recording, no cloud upload — only timestamped event labels leave the wearable, stated as an explicit design guarantee.
• [Complete training pipeline] Dataset curation, augmentation, training and on-device evaluation notebooks; confusion matrix computed on the student's held-out test set.

• Complete ESP32-S3 firmware (audio capture, feature extraction, TFLite Micro inference, event logging, BLE)
• Training pipeline (dataset curation, augmentation, training, quantization, evaluation notebooks)
• Phone dashboard web app (count, hourly chart, event log, battery) over BLE
• Wearable assembly guide (board, microphone placement, battery, enclosure)
• Data-collection procedure for custom cough samples (consent and hygiene notes included)
• Project report PDF (background, TinyML design, training, evaluation, limitations)
• PPT presentation for final review
• Viva Q&A preparation document (TinyML, quantization, spectral features, BLE, evaluation)
• Setup guide (environment, training, flashing, dashboard pairing)

• This is an academic prototype, not a medical device: cough counts are indicative and must never be used for diagnosis, treatment decisions or health monitoring. The report and dashboard carry this warning.
• The model detects cough-like sounds — throat clearing, loud laughter and similar transients can trigger false positives; the evaluation documents the confusion patterns openly.
• Wind noise and a loose microphone mount degrade detection; the assembly guide documents placement and the firmware's gain staging.
• No clinical accuracy is claimed; performance is measured on the student's own held-out test set, which cannot represent all voices, ages and environments.
• Battery life is estimated from the power budget — continuous inference plus BLE advertising drains faster than the headline figure in noisy environments.
• Training a robust model needs a diverse custom dataset; the COUGHVID corpus helps, but the report is honest that dataset bias is the main accuracy risk.