International audience We present the first large-scale corpus of spoken Haitian Creole (Kreyòl), namely Radio Haiti-Inter. The corpus was constructed using automatic speech recognition (ASR) with a state-of-the-art model specifically dedicated to Kreyòl. In addition to transcriptions, we provide part-of-speech (POS) tags, as well as time-aligned transcripts and confidence scores, enabling users to select the most reliable segments for their research. We conduct a manual evaluation of both the transcription quality and POS tagging accuracy to assess the reliability of the resource we present. To enable high-quality research with the resource we introduce, we are releasing 50 hours, comprising both the audios and attached annotations, drawn from the highest-quality segments. This corpus represents an invaluable resource for advancing the study of Kreyòl, with potential applications in phonetics, phonology, morphology, syntax, as well as the study of code-switching and code-mixing. As the recordings cover a large span of years, the corpus we introduce is also suited to micro-diachronic studies of Kreyòl.

1. Introduction

2. Related Work

3. Data & Annotation
-- 3.1. Source
-- 3.2. Automatic Speech Recognition
-- 3.3. Alignments
-- 3.4. Confidence Measure
-- 3.5. POS Tagging

4. Human Evaluation

5. Radio Haïti-Inter Corpus

6. Discussion

7. Conclusion

8. Acknowledgements

9. Bibliographical References