Dazzlin-32/SER-Amharic

speech-emotion-recognition-Amharic # 🎤 Speech Emotion Recognition (SER) using Amharic Dataset *Author:* SEKINA HUSSEIN SHERIF *Date:* July 2025 *Thesis Base Project* | Detecting emotions from human speech using machine learning This project serves as a key part of my AI thesis --- ## 📌Project Overview This notebook classifies human emotions from voice using MFCC features and a MLP model trained on the ASED(Amharic Speech Emotion Detection) dataset. It builds on an existing Kaggle-baased appproach, improved by fine-tuning features and architecture - resulting in **26.19% accuracy improvement** (final accuracy **89.15%**) --- ## 📂Dataset - **Source** - ASED on Github - **Access** - Cloned into data/ - **Format** - `.wav` audio files with emotion labels embedded in the filename. - **Classes Used** - - 🎭 Neutral - 😊 Happy - 😢 Sad - 😡 Angry - 😨 Fearful   >Note: The raw dataset is not included in this repository due to size and licesnsing. Your `.wav` files to the `/data` folder with the first block of code. --- ## 🧠 Methodology - *Audio Feature Extraction*: librosa to compute *40 MFCCs* per sample - *Preprocessing*: StandardScaler for normalization, LabelEncoder for labels - *Model Architecture*: - Built a 1D CNN with: 4 × Conv1D layers (ReLU activation) MaxPooling after each conv layer Dropout for regularization Flattened and passed through dense layers Output layer: Softmax (5 emotion classes) Optimizer: Adam Loss function: Categorical Cross-Entropy Output: Softmax layer (5 classes) - *Training*: 50 epochs, batch size 32 - *Evaluation*: Accuracy, precision/recall, and confusion matrix --- ## 📊 Results | Metric | Value | |----------------|---------| | Baseline Accuracy | ~62.96% | | Final Accuracy | *~89.15%* | | Accuracy Gain | +26.19% | ### 🔍 Observations - Highest confusion between fearful and neutral - MFCC dimensionality boost and scaling improved generalization - Model is lightweight and Colab-friendly --- ## 🧪 How to Run ### 🔧 Setup (Locally or on Col …