Panel CLARIN2021: The Role of Corpora for the Study of Language Use and Mental Health Conditions

Stefan Goetze is Senior Lecturer at the University of Sheffield, UK. Before, he was deputy head of the department “Hearing, Speech, and Audio Technology”, Oldenburg's institute branch of the Fraunhofer-Institute for Digital Media Technology IDMT, head of the group “Automatic Speech Recognition" and head of "Audio System Technology for Audiology and Assistive Systems". He received his Dr.-Ing. in 2013 at the University of Bremen, Germany. His research interests include sound processing, enhancement, and assessment for hands-free systems and assistive listening devices as well as automatic speech and sound recognition systems for assistive technologies, human-machine-interaction and computer dialogue systems for persons with normal or pathological hearing and speech production. Stefan Goetze was lecturer at the University of Bremen from 2007 to 2020 and project leader and coordinator of national and international research projects in the fields speech and text analysis and audio signal processing for human-machine interaction and analysis of normal and dysarthric speech. He and his team developed screening and training systems for persons suffering from speech and speaking problems due to Parkinson's or late provision with cochlear implants.

Availability of data is particularly important for speech and language technology. Although the problem of limited data for specific (research) questions, such as analysis of mental health conditions from the speech signal, is identified and large efforts are already undertaken to collect and annotate such data, the amount and variability is yet far away from what we know from e.g. the field of automatic speech recognition, where researchers and companies made enormous efforts to collect audio and text data in the past.

In addition to collecting more and more data for specific research questions, focus should be put on how this data could be used to answer different research questions in addition. Partly researchers e.g. already use speech data collected from persons suffering from a particular illness or condition and aim at answering research questions for different conditions. This should be kept in mind already while collecting, annotating, and archiving the data.

 

Efficient annotation and storage of audio and text data are essential to develop speech recognition and analysis systems as well for ‘normal’ as for pathological speech and language data. Stefan Goetze and his team regularly recorded audio data and used annotation tools like Audacity and self-developed tools to enrich publicly available datasets e.g. like Open SLR (https://openslr.org/).

Stefan Goetze and his team contributed research and development for speech and speaking analysis and training systems for persons with Morbus Parkinson and for at least severely hearing-impaired patients who suffer from speech and voice disorders due to insufficient auditory input (audiogenic speech disorders). To develop such system in collaborative research projects with partners from industry, large efforts for data collection of speech material was essential. The material had to be phonetically balanced and reflect typical distortions on phoneme, word and sentence level.

 

Creating awareness of the larger context in which research data could be used is important in the field as well as connecting researchers from different domains. Furthermore, information on available data and resources is still spread over the internet and sometimes difficult to find. Here CLARIN makes a clear contribution.