Vocal Emotion Recognition

This repository contains a speech emotion recognition using emodb (German) and Hamid (Persian) datasets.

Link to datasets:

• Emo-db
• Hamidi

Experimental results

Five different models applied for emotion recognition purpose (SVM, Random Forest, NN, CNN and LSTM). Each model trained and evaluated using datasets of four classes (Sad, Happy, Neutral and Angry). The datasets mixed together to examine the result of each network on a mixed language dataset. Furthermore, datasets samplerate downgraded to 8000 Hz. Following are experimental results from each model (test_size=0.2, random_state=42):

1. SVM accuracy: 76%
2. RF accuracy: 64%
3. NN accuracy: 80%
4. CNN accuracy: 85%
5. LSTM accuracy: 86%

Installing dependencies

Dependencies are listed below and in the requirements.txt file.

• h5py
• Keras
• scipy
• sklearn
• speechpy
• tensorflow
• tqdm

Install one of python package managers in your distro. If you install pip, then you can install the dependencies by running pip2 install -r requirements.txt

If you prefer to accelerate keras training on GPU's you can install tensorflow-gpu by pip2 install tensorflow-gpu

Directory Structure

• speechemotionrecognition - Package folder which contains all the code files corresponding to package
• dataset - Contains the speech files in wav formatted seperated into 7 folders which are the corresponding labels of those files
• models - Contains the saved models which obtained best accuracy on test data.
• example.py - Contains examples on how to use the package

Details of the package

• utilities.py - Contains code to read the files, extract the features and create test and train data
• mlmodel.py - Code to train non DL models. We have three models
  • 1 - SVM
  • 2 - Random Forest
  • 3 - Neural Network
• dnn.py - Code to train Deep learning Models. Supports two models given below
  • 1 - CNN
  • 2 - LSTM

Using the package

Look at example.py for sample usage.