Ludwig-Uhland-Institut für Empirische Kulturwissenschaft

Automatic Swabian Recognition (ASR): Processing and Evaluating Dialect Speech

Gefördert vonInnovation Fund Program of the Cluster of Excellence „Machine Learning: New Perspectives for Science”
ProjektleitungProf. Dr. Thomas Thiemeyer, Prof. Dr. Lea Frermann (Seminar für Sprachwissenschaft),  Prof. Dr. Gerhard Jäger (Seminar für Sprachwissenschaft), Prof. Dr. Kerstin Ritter (Hertie Institute for AI in Brain Health) 
Projektmitarbeiter*innenDr.in Valeska Flor, NN, NN
Projektlaufzeit:2027 - 2029 (+1 Jahr)

The goal of Automatic Swabian Recognition (ASR) is to make the Arno Ruoff archive fully usable for both the scientific community and the wider public by creating a curated, open dataset with parallel tiers of (i) sound recordings, (ii) existing transcriptions in German orthography, and (iii) phonetic transcriptions in IPA produced by a dialect-adapted automatic speech recognition pipeline.

Developing robust automatic speech recognition models on archival dialect data enables transfer to contemporary applications facing similar challenges: regional variation, spontaneous speech, and limited training data. The project addresses a critical gap: current automatic speech recognition systems fail on regional dialects, limiting cognitive assessment for neurodegenerative diseases. By validating dialect-aware models on the TREND cohort (1,200 participants, Tübingen region, longitudinal recordings), we establish speech as a non-invasive digital biomarker while ensuring robustness across patient populations.

Building on this dual foundation of historical dialect archives and contemporary clinical speech data, the project integrates machine-learning methods, quantitative models of linguistic change, and qualitative methods of cultural-anthropological analyses of archival meaning and access.