Spoken cultural heritage can present considerably heterogeneous content as tales, stories, recitals, poems, theatrical representations and other form of folk literature. This work investigates the automatic detection and classification of those data type in large spoken audio archives. The corpus used for this study consists of 90 radio broadcast shows collected for preserving a large variety of Swiss French dialects. Given the variability of the language spoken in the recordings, the paper proposes a language-independent system based on structural features obtained using a speaker diarization system and various acoustic/prosodic features. Results reveal that such a system can achieve an F-measure equal to 0.85 (Precision 0.88/Recall 0.84) in retrieving folk literature in those archives. Prosodic features appear more effective and complementary to structural features. Furthermore, the paper investigates whether the same approach can be used to label speech segments into five large classes (Storytelling, Poetry, Theatre, Interviews, Functionals) showing F-measures ranging from 0.52 to 0.88. As last contribution, prosodic features for disambiguating between spoken prose and spoken poetry are investigated. In summary the study shows that simple structural and acoustic/prosodic features can be used to effectively retrieve and label folk literature in broadcast archives.
展开▼
