abnajlae/darija-asr-benchmark-6speaker

Darija ASR 6-Speaker Benchmark

A fixed, paired 20-utterance benchmark read identically by 6 held-out speakers (3 female: F1, F2, F3; 3 male: M1, M2, M3 -- none present in any training corpus), used to evaluate cross-speaker generalization for a Moroccan Darija (Arabizi) Whisper fine-tuning pipeline (paper not yet published -- citation forthcoming).

Consent and anonymization

Written informed consent was obtained from all six speakers for the recording and publication of their voice data for research purposes. Speaker identifiers have been anonymized (F1-F3, M1-M3) in place of the original names throughout this dataset and the accompanying paper.

Contents

• benchmark.jsonl -- 120 rows (6 speakers x 20 utterances). Fields: speaker_id, gender, utterance_id, attempt_used, file_name, audioPath (relative to chunks/), ground_truth (Arabizi transcription), ground_truth_arabic, duration_seconds.
• chunks/ -- 120 WAV files (16kHz mono), matching audioPath in the manifest.
• results.csv -- transcription hypotheses and CER/PER scores for all 120 recordings across three model configurations (zero-shot baseline, Phase 1 fine-tuned, Phase 2 fine-tuned), as reported in the paper's Multi-Speaker Generalization section.

Utterance categories

The 20 shared utterances span 4 categories, matched to the same stratification used in the paper's main evaluation sets: 9 Pure-Darija-short, 5 code-switched-short, 4 Pure-Darija-long, 2 code-switched-long.

Citation

Citation forthcoming -- paper not yet published, check back here for an update.

License / usage terms

This dataset is released for research use under the terms of the informed consent obtained from each participant (see above), not a standard open license (license: other in the metadata reflects this). Contact the corresponding author for questions about permitted use beyond research and academic citation.