👶 Infant Cry Classification using Deep Learning

A production-ready deep learning system for automatically classifying infant cries into five different categories using PyTorch, Torchaudio, and ResNet34.

The model converts raw infant cry recordings into Mel Spectrograms and leverages transfer learning to recognize the underlying reason behind the cry.

📖 Project Overview

Understanding infant cries is a challenging task for new parents and healthcare professionals.

This project uses Deep Learning and Transfer Learning to automatically classify infant crying sounds into one of five categories:

• Hungry
• Tired
• Discomfort
• Belly Pain
• Burping

The pipeline converts raw audio into Mel Spectrograms before using a pretrained ResNet34 model for classification.

The project follows a modular production-ready architecture separating data preparation, dataset loading, model definition, training, evaluation, prediction, and deployment.

✨ Features

✔ Dataset balancing using random oversampling

✔ Automatic metadata generation

✔ PyTorch Dataset pipeline

✔ Audio preprocessing

✔ Mel Spectrogram generation

✔ Transfer Learning using ResNet34

✔ Automatic evaluation

✔ Confusion Matrix generation

✔ Classification Report generation

✔ Gradio Web Interface

✔ Production-ready folder structure

✔ Modular codebase

⚙ Installation

git clone https://github.com/Omshinde01/InfantCryClassification.git

cd InfantCryClassification

pip install -r requirements.txt

python src/app.py

📂 Project Structure

InfantCryClassification/
│
├── data/
│   ├── raw/
│   ├── balanced/
│   └── metadata.csv
│
├── models/
│   └── best_infant_cry_resnet.pth
│
├── src/
│   ├── config.py
│   ├── utils.py
│   ├── dataset.py
│   ├── model.py
│   ├── train.py
│   ├── evaluate.py
│   ├── predict.py
│   └── app.py
│
├── requirements.txt
├── README.md
└── LICENSE

🎵 Dataset

The dataset contains infant crying recordings categorized into five classes.

Class	Description
Hungry	Cry due to hunger
Tired	Cry caused by sleepiness
Discomfort	Wet diaper, uncomfortable position etc.
Belly Pain	Gastrointestinal discomfort
Burping	Need to burp after feeding

Dataset balancing is performed using random oversampling to ensure equal class distribution.

🧠 Model Architecture

The project uses a pretrained ResNet34 model as the feature extractor.

The first convolution layer is modified to accept a single-channel Mel Spectrogram instead of a three-channel RGB image.

The final classification layer is replaced with a custom classifier for five infant cry categories.

🚀 Training

The model is trained using

• CrossEntropyLoss
• AdamW Optimizer
• ReduceLROnPlateau Scheduler

Audio Processing

• Sample Rate: 16000 Hz
• Duration: 5 seconds
• Mel Bands: 64
• FFT Size: 1024
• Hop Length: 512

🌐 Web Interface

The project includes a Gradio-based interactive web application that enables users to upload an infant cry recording and instantly receive the predicted cry category along with the model's confidence score.

🏠 Home Interface

The home screen allows users to upload a .wav audio recording of an infant cry for analysis.

---

📊 Prediction Result

After processing the uploaded audio, the application displays:

• ✅ Predicted Cry Category
• 📈 Confidence Score
• ⚡ Real-time inference using the trained ResNet34 model

🚀 Production REST API

The project includes a production-ready FastAPI service designed for real-time infant cry diagnosis. The API follows a two-stage AI inference pipeline that first verifies whether the uploaded audio actually contains an infant cry before performing detailed cry classification.

This design minimizes false predictions on environmental sounds and provides a more reliable inference pipeline for real-world deployment.

---

🏗 System Architecture

---

🔄 Two-Stage AI Inference Pipeline

Stage 1 – Infant Cry Verification (YAMNet)

The uploaded audio is first analyzed using Google's YAMNet model from TensorFlow Hub.

YAMNet extracts high-level acoustic embeddings and predicts hundreds of environmental sound classes.

The API combines the confidence scores of cry-related categories such as:

• Baby Cry
• Infant Cry
• Crying
• Sob
• Whimper
• Wail
• Scream

If the combined cry confidence is below the predefined threshold, the audio is considered environmental noise.

Instead of attempting cry classification, the API immediately returns the Top-3 detected ambient sounds.

This significantly reduces false-positive predictions.

---

Stage 2 – Infant Cry Diagnosis (ResNet34)

If YAMNet confirms that an infant cry is present, the audio is processed further.

The preprocessing pipeline performs:

• Audio decoding
• Mono conversion
• Resampling to 16 kHz
• Fixed-length padding/trimming (5 seconds)
• Mel Spectrogram generation
• Log-amplitude conversion
• Feature normalization

The generated Mel Spectrogram is then passed to a custom ResNet34 classifier trained using transfer learning.

The classifier predicts one of five infant cry categories:

• Hungry
• Tired
• Discomfort
• Belly Pain
• Burping

---

☁ AWS Cloud Storage

Every uploaded audio recording is automatically archived to Amazon S3 using FastAPI Background Tasks.

This asynchronous approach ensures that cloud uploads never block the inference pipeline, resulting in lower response latency.

Archived recordings can later be used for:

• Dataset expansion
• Model retraining
• Performance monitoring
• Audit history

---

⚡ API Features

• Two-stage AI inference pipeline
• FastAPI-based REST API
• Google YAMNet verification model
• Custom ResNet34 classifier
• Mel Spectrogram feature extraction
• Background AWS S3 archival
• Dockerized deployment
• JSON responses
• Real-time inference
• Production-ready architecture

---

📡 API Endpoint

Predict Infant Cry

POST /predict

Request

Upload an audio file using multipart/form-data.

Parameter	Type	Description
file	Audio (.wav)	Infant cry recording

---

✅ Example Response (Cry Detected)

{
  "cry_detected": 1,
  "combined_cry_score": 0.91,
  "prediction": "Hungry",
  "archive_name": "20260710_123045_ab12cd34.wav",
  "breakdown": {
    "Hungry": 91.42,
    "Tired": 2.64,
    "Discomfort": 1.88,
    "Belly Pain": 2.05,
    "Burping": 2.01
  }
}

---

🚫 Example Response (No Cry Detected)

{
  "cry_detected": 0,
  "combined_cry_score": 0.04,
  "prediction": "Noise Detected",
  "archive_name": "20260710_124050_f34ab671.wav",
  "top_noises": {
    "Vacuum Cleaner": 46.81,
    "Speech": 28.34,
    "Music": 12.56
  }
}

---

🐳 Docker Deployment

Build the Docker image

docker build -t infant-cry-api .

Run the container

docker run -p 7860:7860 infant-cry-api

📈 Why a Two-Stage Pipeline?

Instead of directly classifying every uploaded sound, the system first determines whether the audio actually contains an infant cry.

This design offers several advantages:

• Reduces false-positive predictions
• Handles environmental noise gracefully
• Improves inference reliability
• Makes the API more suitable for real-world deployment
• Separates cry detection from cry diagnosis

🚀 Future Work

• EfficientNet backbone
• ConvNeXt backbone
• Vision Transformer
• ONNX Export
• TensorRT Optimization
• Mobile Deployment
• REST API
• Real-time Streaming Prediction

👨‍💻 Author

Om Shinde

AI/ML Engineer

LinkedIn: https://linkedin.com/in/omshinde01

GitHub: https://github.com/Omshinde01