TinyML Baby Cry Detector

New parents buy Wi-Fi baby monitors that stream audio to the cloud around the clock — expensive, subscription-tied, and a privacy worry in the nursery. This project takes the opposite approach: a small bedside ESP32 unit listens locally and recognizes the acoustic signature of an infant cry with a TinyML model, alerting only when crying is actually detected. Because inference happens on the microcontroller, there is no account, no app pairing and no audio upload; the unit works on a shelf with just a USB charger. The build covers the full TinyML pipeline honestly — dataset curation from a public infant-cry corpus, mel-spectrogram features, model quantization, and false-alarm testing against everyday household sounds.

1. The INMP441 microphone streams 16 kHz audio to the ESP32 over I2S.
2. Firmware computes log-mel spectrograms over sliding 1-second windows.
3. The quantized CNN outputs a cry probability per window; probabilities are averaged over 3 seconds for a stable decision.
4. If the smoothed score exceeds the sensitivity threshold, the chime sounds, the LED lights, and the episode timer starts on the OLED.
5. When crying stops for 10 continuous seconds, the episode closes and its start time and duration are written to the local log.
6. Sensitivity levels adjust the decision threshold; all settings persist across power cycles.

Tech stack:

• ESP32 (dual-core Xtensa, 520 KB SRAM)
• INMP441 I2S MEMS microphone
• TensorFlow Lite Micro, INT8 quantized CNN
• 0.96" SSD1306 OLED, chime buzzer, status LED
• Donate-a-cry corpus — training reference
• Python training pipeline (librosa, TensorFlow)

Dataset & model details

• Dataset: The Donate-a-cry corpus — a public research dataset of infant cry recordings collected for cry-classification studies — used as the training/evaluation reference for the cry class; household background sounds (speech, TV, appliances) form the negative class. The report documents the corpus's size, recording conditions and its limits.
• Task: 1-second audio-window binary classification (infant cry vs background); input = 40×49 log-mel spectrogram, output = cry probability per window, smoothed over 3 s.
• Model: Small CNN (two conv blocks + dense head, ~30 KB after INT8 quantization), TensorFlow Lite Micro on ESP32 (design target).
• Metrics: Window-level precision/recall and false alarms per day on held-out clips — reported as design targets in the build notes, with the actual measured figures recorded during the buyer-run nursery test. No accuracy is claimed as measured until that run. Educational/assistive prototype; no medical or safety-critical claims.

• [On-device cry classification] A quantized CNN scores 1-second audio windows on the ESP32 — no cloud, no streaming, no audio leaves the nursery.
• [Cry vs noise discrimination] Trained to separate infant crying from TV, talking, fans and traffic noise, with multi-second confirmation to reject brief sounds.
• [Gentle local alerts] A soft chime plus LED indicator on confirmed crying — no harsh siren near the baby.
• [Cry-duration timer] The OLED shows how long the current episode has lasted and a count of today's episodes.
• [Episode log] Timestamps of cry episodes are stored locally so parents can review patterns over days.
• [Adjustable sensitivity] A button cycles low/medium/high sensitivity for different room sizes and noise floors.
• [Training pipeline included] Python scripts for dataset prep, training and TFLite Micro conversion ship with the project.

• Assembled detector unit: ESP32 + microphone + OLED + chime + LED
• Complete firmware (audio pipeline, inference, alerts, log)
• Quantized TFLite Micro model + Python training/conversion pipeline
• Wiring diagram and nursery placement guide
• Project report PDF (TinyML background, dataset documentation, methodology, nursery test procedure, results, safety disclaimer)
• PPT presentation for final review
• Viva Q&A preparation document (CNN, quantization, mel features, precision/recall, false alarms)

• Detection performance is a design target until the buyer's nursery test — loud siblings, barking dogs and TV at high volume will challenge it, documented in the report.
• This is an assistive educational prototype, NOT a safety-critical baby monitor — it must never be the sole supervision for an infant, stated plainly.
• Designed for one room (~3–5 m); it cannot monitor through closed doors or across a house.
• The model recognizes crying sounds, not distress causes — hunger, discomfort and pain sound similar to the classifier.
• Continuous mains power is assumed; there is no battery backup in the base build.
• Like all ML classifiers, it can miss real cries (false negatives) — the report discusses this failure mode honestly.