MELANZ: A Trilingual Kreol Morisien–English–French Code-Switching SpeechCorpus for Automatic Speech Recognition

MELANZ is the first code-switching (CS) speech corpus for Kreol Morisien (KM), a French-lexifier Creole language spoken natively by over 90% of the Mauritian population. In practice, KM speakers routinely alternate between KM, English, and French within a single utterance, yet no dedicated speech resource exists to reflect this multilingual reality. The MELANZ corpus comprises approximately 22 hours of broadcast news speech from 76 episodes of the Zournal Kreol television programme. The recordings were segmented into 4,037 utterances totalling 197,028 words, with all transcriptions manually verified. A 2-hour evaluation subset (271 utterances, 16,098 words) was annotated with word-level language identification tags, achieving an inter-annotator agreement of Cohen's kappa = 0.966. Corpus analysis confirms the trilingual nature of KM speech: 91.1% of utterances contain at least one language switch, 46.1% contain words from all three languages simultaneously, and the average Code-Mixing Index is 0.189. Baseline ASR experiments with three architecturally distinct models (Whisper large-v3, MMS, wav2vec2-XLSR-1B) reveal that French words are consistently the most difficult to recognise (WER 31.3-47.9%), error rates at switch points are 130-157% higher than elsewhere, and recognition degrades monotonically with CS intensity. MELANZ fills a critical gap as the first CS speech resource for any Creole language.