Stuttering is a speech disorder which is characterized by blocks, prolongations, and repetitions of words, and syllables. It has a prevalence of about 1% of the population. The condition is treatable but not curable. Good results in therapy can be achieved by teaching a 1technique called “Fluency Shaping”. One institute in Germany that oers this kind of therapy is our project partner Kasseler Stottertherapie (KST).
The research project is based in the area of machine learning and pattern recognition. Recognizing stuttering events is a hard medical problem. To recognize such events, methods from pattern recognition and machine learning are being used, colloquially labeled as artificial intelligence, to enable automatic detection. To achieve this medical expert knowledge as well as deep learning methods are combined to dierentiate between fluent speech, repetitions, blocks and unnatural prolongations as well as detecting if a client is using the speech technique learned during therapy.
As speech data is highly sensible personal data it is of utmost importance to guarantee peoples privacy. To ensure this all algorithms developed must be adapted to run on smartphones
or smart-health devices. This is supposed to guarantee data privacy by not sending speech data into the cloud, as it is usually done for speech analysis. To enable this algorithms must
be adapted to be performance and energy eicient on devices with limited resources.
A main goal of this research is to be able to successfully classify stuttering events on devices with limited resources. My research is supposed to lay the foundation for technical aids that
can be used in stuttering therapy, but can in part be transferable to other speech disorders and pathologies. This shall enable a lasting control of therapy success.