CodeAlpha_EmotionRecognition

CodeAlpha — Machine Learning Internship | Task 2

Objective

Recognize human emotions (neutral, calm, happy, sad, angry, fearful, disgust, surprised) from speech audio files using deep learning.

Dataset

• Name: RAVDESS — Ryerson Audio-Visual Database of Emotional Speech and Song
• Source: Downloaded automatically via Zenodo (official source)
• Size: 1,440 audio files — 24 actors (12 male, 12 female), 8 emotions
• Format: .wav files, 24kHz stereo

Emotion Classes

Code	Emotion	Code	Emotion
01	Neutral	05	Angry
02	Calm	06	Fearful
03	Happy	07	Disgust
04	Sad	08	Surprised

Tools & Libraries

• Python 3.10
• librosa — audio processing & MFCC extraction
• TensorFlow / Keras — deep learning model
• NumPy, pandas — data handling
• matplotlib, seaborn — visualization
• scikit-learn — preprocessing & evaluation

Approach

Feature Extraction

• Loaded each .wav file using librosa
• Extracted 40 MFCC (Mel-Frequency Cepstral Coefficients) per audio clip
• Audio clipped to 3 seconds, padded/truncated to fixed length (174 frames)
• Final feature shape per file: (174 timesteps × 40 MFCC coefficients)

Model Architecture — CNN + LSTM Hybrid

Input (174, 40)
    ↓
Conv1D (64 filters) → BatchNorm → MaxPool → Dropout
    ↓
Conv1D (128 filters) → BatchNorm → MaxPool → Dropout
    ↓
Conv1D (256 filters) → BatchNorm → MaxPool → Dropout
    ↓
LSTM (128 units)
    ↓
Dense (128) → BatchNorm → Dropout
    ↓
Dense (64) → Dropout
    ↓
Dense (8, softmax) → Emotion Prediction

Training Strategy

• Optimizer: Adam (lr=0.001)
• Loss: Categorical Cross-Entropy
• EarlyStopping (patience=10)
• ReduceLROnPlateau (patience=5)
• Train / Val / Test split: 70% / 15% / 15%

What's Inside the Notebook

Step	Description
1	Install libraries
2	Import libraries
3	Download RAVDESS dataset automatically
4	Understand file naming convention
5	EDA — emotion distribution, waveform, spectrogram, MFCC visualization
6	MFCC feature extraction for all files
7	Label encoding & train/val/test split
8	Build CNN + LSTM model
9	Train model with callbacks
10	Plot training history (accuracy & loss)
11	Evaluate — accuracy, classification report, confusion matrix, F1 per class
12	Save model (.h5) and label encoder (.pkl)
13	Predict emotion from a new audio file
14	Key insights & findings

Key Results & Findings

• CNN layers extract local temporal patterns from MFCCs
• LSTM captures long-range sequential dependencies in speech
• Angry and happy emotions tend to achieve higher F1-scores due to distinct acoustic features
• Neutral and calm are sometimes confused due to acoustic similarity
• Model saved as emotion_recognition_model.h5 for reuse and deployment

How to Run

pip install librosa tensorflow numpy pandas matplotlib seaborn scikit-learn jupyter
jupyter notebook Task2_Emotion_Recognition.ipynb

Run all cells top to bottom. The dataset downloads automatically in Step 3 (~215 MB).

Repository Structure

CodeAlpha_EmotionRecognition/
├── Task2_Emotion_Recognition.ipynb   ← Main notebook
├── README.md                          ← This file
├── emotion_recognition_model.h5       ← Saved model (generated after training)
├── label_encoder.pkl                  ← Saved label encoder (generated after training)
├── emotion_distribution.png           ← EDA plot
├── waveform_spectrogram.png           ← Waveform & spectrogram
├── mfcc_visualization.png             ← MFCC heatmap
├── training_history.png               ← Accuracy & loss curves
├── confusion_matrix.png               ← Confusion matrix
└── f1_per_emotion.png                 ← F1-score per emotion