Maschinelle Eigenschaftsanalyse

This is a project dedicated to the classification of emotional speech and was created in class with Prof. Dr. Burkhardt at Technische Universität Berlin. After holding a short speech in class, the aim of the project was to analyze the voice recording by taking a closer look at target acoustic features such as the HNR, the mean F0Hz or jitter.

The project uses Jupyter, an open-source web application that allows you to create and share documents that contain live code that can be used for statistical modeling and data visualization.

Learning Objectives

• Introduction to Python and Juypter Notebook
• Classification of Emotional Speech
• Acoustic Analysis through Extraction and Analysis of Expert and Brute-Force Features
• Visualization of Statistical Findings

Preparation

In order to run this project, some preliminary steps are necessary:

• Create a virtual environment for the project and collect the necessary imports. If a certain plug-in is missing, it is recommended to install it via Pip in the terminal, so that the import into Jupyter Notebook can run smoothly. You will definitively need: Pandas, os, Matplotlib, seaborn, glob, NumPy, Parselmouth and SciPy.
• Prepare the audio file for the analysis. The format should be: wav, 16kHz, 16bit, PCM (possible tool: SoX/Audacity).
• Segment and annotate your data according to the dimension (or a dimension/category of your choice) on a 10-level Likert scale (possible tool: Audacity, Speechalyzer).

Documentation

After configuring a python environment with Jupyter notebook the audio and its annotations are first imported into a pandas table and then analyzed step by step according to the Feinberg PraatScripts. In intermediate steps, the obtained data is combined with the annotations priorly given to the different audio segments. Necessary information on the single steps is included in the notebook.

Extracting Acoustic Features

The PraatScripts and this notebook offer a comprehensible code to perform your analyses. Nevertheless, you might need to adapt the scripts to fit your needs.

The overall analysis includes:

• Formants
• HNR
• Pitch
• Intensity
• Jitter
• Shimmer
• Speech Rate
• Pauses
• Vocal-tract Length Estimates

Visualization

The project includes different approaches to visualize the data using the tools seaborn and Matplotlib for e.g. scatter plots, box/violin plots or cluster plots.

Credits

• Speechalyzer
• Feinberg PraatScripts

Author: @lou-ukw