Apppublic

arshadykbk2/Child-Speech-Emotion-Recognition

sourceHugging Facemitupdated 2mo agoView on Hugging Face

0likes

App README

🎙️ Child Speech Emotion Recognition (SER) System

A SOTA Deep Learning & Healthcare Platform for Automated Child Vocal Emotion Classification Built with Microsoft WavLM Transformer, 1D Residual CNNs, Attention Pooling, and FastAPI.

🌟 Executive Summary

This project implements a state-of-the-art Child Speech Emotion Recognition (SER) System designed for clinical healthcare assessment, developmental psychology research, and academic demonstration.

The system classifies child vocalizations into 6 distinct emotion classes:

—😃 Happy
—😢 Sad
—😡 Angry
—😨 Fear
—😐 Neutral
—🤢 Disgust

📊 Key Performance Metrics

The model was evaluated on the independent test dataset, achieving state-of-the-art classification performance:

Metric	Score
Validation Accuracy	97.18%
F1-Score (Macro)	97.14%
Recall (Macro)	97.24%
Precision (Macro)	97.06%
Average ROC-AUC	0.99

🏗️ Neural Network Architecture

text

Input Audio (16kHz PCM Signal)
          │
          ▼
┌──────────────────────────────────────────────┐
│  Microsoft WavLM Base (768-dim Acoustic)    │  <-- Pre-trained Transformer Encoder
└──────────────────────────────────────────────┘
          │
          ▼
┌──────────────────────────────────────────────┐
│  1D Residual CNN (3 Residual Blocks)         │  <-- Local Temporal Feature Extraction
└──────────────────────────────────────────────┘
          │
          ▼
┌──────────────────────────────────────────────┐
│  Attention Pooling Mechanism                 │  <-- Dynamic Time-Frame Weighting
└──────────────────────────────────────────────┘
          │
          ▼
┌──────────────────────────────────────────────┐
│  Linear Classification Head (6 Classes)      │  <-- Softmax Probability Distribution
└──────────────────────────────────────────────┘

🚀 Quick Start Guide (Offline Desktop Software)

🍏 On Mac:

1.Double-click `RUN_OFFLINE_MAC.command`.
2.Safari/Chrome will open automatically to http://localhost:8000.

🪟 On Windows:

1.Double-click `RUN_OFFLINE_WINDOWS.bat`.
2.Chrome/Edge will open automatically to http://localhost:8000.

☁️ Hugging Face Cloud Deployment

This project is pre-configured for 1-click hosting on Hugging Face Spaces:

1.Create a new Space on Hugging Face (SDK: Docker $\rightarrow$ Blank).
2.Upload the project files or push via Git CLI:

bash

   git clone https://huggingface.co/spaces/YOUR_USERNAME/YOUR_SPACE_NAME
   cd YOUR_SPACE_NAME
   cp -r /path/to/project/* .
   git add .
   git commit -m "Deploy Child Emotion AI Web App"
   git push

📁 Repository Directory Structure

text

├── RUN_OFFLINE_MAC.command       # 1-Click Mac Executable Launcher
├── RUN_OFFLINE_WINDOWS.bat       # 1-Click Windows Executable Launcher
├── Dockerfile                    # Containerization manifest for Cloud Deployment
├── server.py                     # FastAPI Backend Inference Server
├── index.html                    # Commercial React + Tailwind Healthcare UI
├── config.py                     # System Parameters & Hyperparameters
├── requirements.txt              # Python Dependencies List
├── OFFLINE_APP_INSTRUCTIONS.md   # User Guide for Offline Executables
│
├── models/                       # Neural Network Architectures
│   ├── wavlm_classifier.py       # WavLM Base Foundation Model
│   ├── cnn.py                    # 1D Residual Convolutional Layers
│   └── attention.py              # Attention Pooling Mechanism
│
├── dataset/                      # Data Processing & Loader Pipeline
│   ├── parser.py                 # Audio Parser & Label Extractor
│   ├── loader.py                 # PyTorch DataLoader & Collate Logic
│   └── preprocessing.py          # 16kHz Resampling & Normalization
│
├── training/                     # Model Training Engine
│   ├── trainer.py                # Trainer Class with AMP & Early Stopping
│   ├── losses.py                 # CrossEntropy & Focal Loss Functions
│   └── metrics.py                # F1, Precision, Recall, ROC-AUC Computations
│
├── utils/                        # Utility Helpers
│   ├── checkpoints.py            # Model Checkpoint Saver/Loader
│   └── seed.py                   # Reproducibility Seed Manager
│
└── outputs/                      # Trained Artifacts & Weights
    ├── label_mapping.json        # Class Index Mapping
    └── checkpoints/
        └── emotion_model.pth     # 385 MB Trained Checkpoint (97.18% Accuracy)

📄 License & Attribution

—Developed for: B.Tech Final-Year Capstone Project & Healthcare Deliverable.
—Dataset: Child Speech Emotion Dataset (C-BESD).
—Frameworks: PyTorch, Transformers, FastAPI, Tailwind CSS.