The speech signal encodes a variety of speaker and language characteristics, such as voice quality, speaking rate, emotion, or accent in addition to the words spoken. Many of these properties, however, are difficult to quantify and are intertwined. We are therefore working on the development of generative mathematical models for the discovery and targeted manipulation of voice characteristics. This allows us to manipulate, for example, gender or more nuanced characteristics such as nasality or breathiness of the voice without significantly changing other characteristics of the speaker. Such systems can then help in the training of clinical linguistics or actors. If you enjoy mathematics, programming, and machine learning, we would be happy to invite you to our office to discuss potential thesis topics.
