audio_emotion_analysis

The objective of this project is to predict the emotion present in any audio file/signal.

Clone the project and move to the directory in shell/command line.

##Requirements:

1. Anaconda - This installs python along with most popular python libraries including sklearn. If not already installed, install it from https://www.continuum.io/downloads .
2. python_speech_features - Python library for feature extraction.
3. pyaudio - Python library for recording and playing of audio samples.

The requirements 2 and 3 can be installed by executing the following using shell/cmd, in the cloned repository directory:

pip install -r requirements.txt

##Installation:

Once the requirements are installed, just type the following in shell/cmd python setup.py install

##Preparing dataset:

• Put all the unlabelled audio files in a folder named calls, or any other folder and update the name of folder in label_dataset.py.
• Run the script label_dataset.py.
  • It will scan all the audio files, create a set of 30 sec audio chunks.
  • It will play each chunk and then ask for a label (positive/neutral/negative).
  • Enter 1 for negative, 2 for neutral and 3 for positive.
  • Continue until all the chunks are labelled.
  • A new dataset will be prepared in a new folder named data, with each audio file of the name in the format <label>_<counter>.wav. For example:

   > positive_1.wav
   > negative_2.wav
  

Scripts description:

feature_extractor.py

Extracts the features from any audio file or signal and returns a feature vector.

emotion_analysis.py

This is the backbone of the project. It contains modules that

• extracts features using the feature_extractor.py
• trains the model
• tests the model
• evaluates the model on new dataset

##Evaluating a new dataset:

• Put all the new audio files in a folder named test_calls, or any other folder and update the name of the folder in main_script.py.
• Execute the script main_script.py by typing the following in shell/cmd, in the containing directory:

python main_script.py

• The script will load an existing model or train a new model on the dataset prepared, extract features for each of the new audio files and feed it to the trained model.

###Results

• For each file, following three things are evaluated:
  • Overall emotion
  • Emotion transition from first half to second half
  • Emotion present in each 20 sec chunk of the file
• These results for each audio file are written to a .csv file, on which further analysis can be done.