Autism Sensory Meltdown Risk Detector (Acoustic AI)

Assistive Acoustic Deep Learning Model for Real-Time Auditory Trigger Detection & Sensory Overload Prevention
Developed for Datathon 2025 by Ristek UI • Engineered by Zeiniah (@zeyniaa)

---

📌 Project Overview

Sensory sensory overload is a frequent and distressing experience for autistic and neurodivergent individuals, often provoked by acute, loud, or repetitive environmental noises. Without proactive warning mechanisms, rapid escalation into sensory meltdowns can occur.

The Autism Sensory Meltdown Risk Detector is an assistive acoustic intelligence system designed to continuously monitor acoustic environments, classify trigger noise events in real-time, and compute an immediate sensory distress risk index.

Built and benchmarked during Datathon 2025 by Ristek UI, the system employs Librosa for spectral decomposition (Mel-Frequency Cepstral Coefficients / MFCC) and trains an optimized 1D-Convolutional Neural Network (1D-CNN) to achieve 90% classification accuracy across multi-class environmental triggers.

---

🎯 Benchmark Performance

The model undergoes strict multi-metric evaluation across unseen test distributions:

Performance Metric	Score	Description
Classification Accuracy	90.00%	Multi-class overall prediction accuracy
Average F1-Score	0.89	Balanced harmonic mean of precision and recall
Average Precision	0.91	Low false positive rate across ambient soundscapes
Inference Latency	< 45 ms	Real-time classification suitable for wearable edge devices

---

🔬 System Architecture & Audio Pipeline

flowchart TD
  subgraph Ingestion ["🎙️ Acoustic Ingestion"]
    Wave["Raw Audio Waveform (.wav / .mp3)<br/>Mono • 22,050 Hz"]
    Segment["Temporal Windowing<br/>Fixed 3.0s Duration • Zero-Padding"]
  end

  subgraph FeatureEng ["🎛️ Spectral Feature Extraction (Librosa)"]
    STFT["Short-Time Fourier Transform (STFT)"]
    Mel["Mel-Frequency Scaling (Triangular Filter Banks)"]
    MFCC["40-Dimensional MFCC Feature Vector<br/>Temporal Mean Pooling"]
    Tensor[("Input Tensor: (Batch, 40, 1)")]
  end

  subgraph ModelArchitecture ["🧠 1D-CNN Deep Learning Architecture"]
    Conv1["Conv1D Block 1 (64 filters, k=5, ReLU)<br/>BatchNorm • MaxPool1D(2) • Dropout(0.25)"]
    Conv2["Conv1D Block 2 (128 filters, k=5, ReLU)<br/>BatchNorm • MaxPool1D(2) • Dropout(0.30)"]
    Conv3["Conv1D Block 3 (256 filters, k=3, ReLU)<br/>BatchNorm • GlobalAveragePooling1D"]
    Dense1["Dense Layer (128 units, ReLU)<br/>Dropout (0.40)"]
    Softmax["Softmax Output Layer (6 Classes)"]
  end

  subgraph DiagnosticOutput ["🚨 Risk Assessment & Advisory"]
    Cat["Trigger Classification<br/>(Crying, Siren, Appliance, Barking, Bang, Neutral)"]
    Risk["Meltdown Risk Level<br/>[CRITICAL / HIGH / MEDIUM / SAFE]"]
  end

  Wave --> Segment --> STFT --> Mel --> MFCC --> Tensor
  Tensor --> Conv1 --> Conv2 --> Conv3 --> Dense1 --> Softmax
  Softmax --> Cat --> Risk

Loading

---

🎧 Auditory Trigger Taxonomy & Clinical Advisories

Trigger Category	Acoustic Characteristics	Risk Severity	Real-World Advisory
siren_alarm	High-intensity repetitive modulation (80–110 dB)	CRITICAL	Emergency sirens, fire alarms. Relocate to quiet decompression room immediately.
crying_screaming	Intense high-frequency human distress signals	HIGH	Sharp vocal bursts. Ear protection or noise-canceling headphones recommended.
percussive_bang	Sudden transient onset (< 20ms attack time)	HIGH	Fireworks, slamming doors, thunder. High risk of startle-induced sensory shock.
appliance_vacuum	Continuous broadband motor noise + hum	MEDIUM	Blenders, vacuum cleaners, lawnmowers. Acoustic fatigue trigger over time.
dog_barking	High dynamic impulse barks	MEDIUM	Sudden domestic barking. Prepare calming sensory tools.
ambient_neutral	Low-decibel diffuse envelope (< 55 dB)	SAFE	Calm conversation, soft rain, library environment.

---

📁 Repository Structure

autism-meltdown-detector/
├── notebooks/
│   └── autism_meltdown_colab.ipynb  # Interactive Google Colab notebook
│
├── src/
│   ├── __init__.py
│   ├── audio_processing.py          # Librosa MFCC & spectral feature extractor
│   ├── models/
│   │   ├── __init__.py
│   │   └── cnn1d.py                 # 1D-CNN model architecture
│   ├── train.py                     # CLI training script with Callbacks
│   └── predict_audio.py             # CLI audio file inference & risk advisory
│
├── data/
│   ├── sample_audio/                # Directory for user test audio samples
│   └── README.md                    # Trigger taxonomy & dataset documentation
│
├── requirements.txt                 # Project dependencies
├── .gitignore                       # Git filters for audio data and checkpoints
└── README.md                        # Project documentation

---

⚡ Quickstart & Usage

1. Installation

Clone the repository and install required packages:

git clone https://github.com/zeyniaa/autism-meltdown-detector.git
cd autism-meltdown-detector
pip install -r requirements.txt

2. Model Training

Train the 1D-CNN acoustic classifier:

python src/train.py --epochs 100 --batch-size 32 --lr 0.0005

3. Audio File Prediction & Risk Advisory

Analyze any audio file for sensory meltdown risks:

python src/predict_audio.py --audio sample.wav

Output preview:

=======================================================
ACOUSTIC MELTDOWN RISK ANALYSIS REPORT:
Detected Trigger : SIREN_ALARM (95.42% confidence)
Meltdown Risk    : [CRITICAL]
Clinical Advisory: High decibel repetitive frequency. Relocate to quiet room.
=======================================================

4. Run on Google Colab

Click the badge below to run the complete pipeline on Google Colab:

---

👩‍💻 Author & Acknowledgments

• Zeiniah (@zeyniaa): Lead AI & Audio Signal Processing Engineer.
• Datathon 2025 by Ristek Universitas Indonesia (UI).
• Dedicated to advancing accessible and human-centric assistive technologies for neurodivergent individuals.