Speaker
Description
Parliamentary recordings are a well-known resource for obtaining large-scale data for automatic speech recognition. In this contribution, we argue for a much broader use of parliamentary recordings and their accompanying official transcripts, rich speaker metadata, and generally non-problematic privacy and copyright status, especially in the context of less-resourced languages. Beyond speech recognition, such resources provide unique opportunities for large-scale research on acoustic, linguistic, prosodic, and discourse-level phenomena in semi-spontaneous speech. To illustrate these opportunities, we showcase ParlaSpeech, a multilingual collection of more than 6,000 hours of parliamentary recordings aligned with official transcripts in four Slavic languages: Croatian, Czech, Polish, and Serbian. The resource integrates several layers of linguistic and speech-oriented annotation. In addition to extensive parliamentary and speaker metadata inherited from the ParlaMint project, the corpus includes Universal Dependencies syntactic annotation, sentiment and topic annotations derived from the textual modality, and filled pause and primary stress annotations derived from the speech modality. The resulting infrastructure supports the study of interactions between prosody, acoustics, syntax, discourse structure, and speaker behavior across languages. We present several completed and ongoing research directions enabled by this infrastructure. Previous work has explored relationships between acoustic variables and sentiment expression, as well as cross-linguistic patterns in the frequency and duration of filled pause production. Ongoing analyses investigate patterns of primary stress across speakers with varying sociodemographic background, and the syntactic environments associated with occurrence of filled pauses. Future work will extend the collection with additional languages and annotation layers, opening up additional opportunities for under-researched languages.