CAFE: Algerian Arabic, French, and English Code-Switched Conversational Speech Dataset

The CAFE dataset (Code-switching Algerian French English) addresses the lack of publicly available resources capturing spontaneous code-switching between Algerian Arabic (Darja), French, and English. Spanning over 37 hours of real-world, in vivo human–human dialogue from more than 100 speakers, it reflects Algeria’s rich sociolinguistic landscape through naturally occurring conversational features such as code-switching points, overlapping speech, non-lexical events (e.g., laughter, fillers, ambient noise), and dialectal variation. CAFE supports research in automatic speech recognition (ASR), dialect identification, and multilingual NLP for low-resource settings. Annotations were based on Latin script for French and English, while Arabic script was applied for Algerian Dialect and MSA. The dataset consists of three main subsets: cafe-small/Contains annotated audio clips featuring natural conversations. audio/: Mono-channel WAV files (16 kHz, 16-bit PCM). transcripts_raw/: A CSV file aggregating raw transcriptions for all files. transcripts_ZAEBUC/:One JSON file with enriched annotations using the ZAEBUC-Spoken format for each file. (including dialect intensity and language switching details). cafe-large/Includes a larger volume of audio clips with pseudo-labels generated using Whisper-based models. audio/: Mono-channel WAV files (16 kHz, 16-bit PCM). pseudo_labels/: CSV files containing pseudo-transcriptions per audio file. CAFE-small-no-overlap/A curated subset of cafe-small, filtered to remove speaker or content overlap for benchmarking or experimentation. audio/: Mono-channel WAV files (16 kHz, 16-bit PCM). transcripts_raw/: CSV files where each row includes the filename and its pseudo-transcription.