"The human voice is a complex construct made up of superimpositions of various influencing factors. As a result, it has different characteristics that are difficult to identify," says Professor Dr. Reinhold Häb-Umbach, Professor of Communications Engineering at Paderborn University and one of the leaders of sub-project C06. "By breaking down speech signals into different components, we can learn more about what makes our voices unique."
The components are divided into linguistic content properties - what someone says - and tonal properties - how the voice sounds. In their publication, the researchers show how the individual components are connected at the tonal level. To do this, they created a model from neural networks that separates the different tonal aspects from one another. This can be used to generate a new synthetic language with specifically modified properties, for example a desired average pitch.
The researchers presented the results in their article "Speech Disentanglement for Analysis and Modification of Acoustic and Perceptual Speaker Characteristics". "With this publication, we are helping to understand how we can use computers to understand and modify different aspects of speech," summarizes Frederik Rautenberg, co-author of the article and also a researcher in subproject C06. "This allows us to develop speech modification programs that can help people with speech difficulties, for example."
The article was presented at the 49th Annual Conference on Acoustics (DAGA). The DAGA is the largest conference on acoustics in the German-speaking world and took place from March 6-9 in Hamburg.
