Audio data archived in radio broadcast stations represent a rich source for various research purposes from phonetic questions up to training and test data for speech modelling. We present an efficient semi-automatic workflow for pre-processing, transcribing and analysing large linguistic-phonetic audio corpora. As a pilot study, we process radio broadcast news from a German public radio station containing recordings from 1956 until 2017. The workflow consists of basic preprocessing, automatic speech recognition, manual word correction, automatic generation of pairs of audio chunks and transcripts, plus an automatic word-, syllable- and phoneme-level segmentation of these chunks. The workflow is organised using the Octra Backend management tool, manual validation and correction of transcripts and chunking are performed using the Octra editor, and the BAS web services perform the segmentation. In an example analysis we show with our specific radio corpus how to use it for comparative longitudinal structure analyses of broadcast news, and for text- and signal-based studies on changes of speech and articulation rate.

This paper introduces an approach to automatically predict the speech fluency of preschool children as part of Language Proficiency Assessments. We use spontaneous speech data from children with German as native and second language aged 4–6 years, collected via a game–based elicitation method. The recordings were mainly annotated manually on various fluency-related phenomena. The resulting feature values were compared to human fluency ratings of the same data. The human ratings and the fluency-related acoustic features were used to build Cumulative Link Mixed Models (CLMMs) with and without splines to test their ability to predict the human ratings with multiple metrics (Spearman’s ρ, MAE, quadratic weighted κ). Results show that a parsimonious linear model already reaches near-human agreement (quadratic weighted kappa κ = 0.65) and that incorporating non-linear spline effects does not improve predictive accuracy. These findings suggest that relatively simple CLMMs can substitute additional human raters in fine-grained fluency assessment of preschool children, which is a task that is already challenging for trained listeners.

Natural language processing (NLP) and speech technologies have made significant progress in recent years; however, they remain largely focused on standardized language varieties. Dialects, despite their cultural significance and widespread use, are underrepresented in linguistic resources and computational models, resulting in performance disparities. To address this gap, we introduce Saar-Voice, a six-hour speech corpus for the Saarbrücken dialect of German. The dataset was created by first collecting text through digitized books and locally sourced materials. A subset of this text was recorded by nine speakers, and we conducted analyses on both the textual and speech components to assess the dataset’s characteristics and quality. We discuss methodological challenges related to orthographic and speaker variation, and explore grapheme-to-phoneme (G2P) conversion. The resulting corpus provides aligned textual and audio representations. This serves as a foundation for future research on dialect-aware TTS, which is proposed in the form of zero-shot and fine-tuned model adaptation in low-resource scenarios.

In this study we investigate the role of inhalation noises at the end of laughter events in two conversational corpora that provide relevant annotations. A re-annotation of the categories for laughter, silence and inbreath noises enabled us to see that inhalation noises terminate laughter events in the majority of all inspected laughs with a duration comparable to inbreath noises initiating speech phases. This type of corpus analysis helps to understand the mechanisms of audible respiratory activities in speaking vs. laughing in conversations.