{
    "Name": "Kirill",
    "Dialect Subsets": [
        {
            "Name": "Egyptian",
            "Volume": 510.0,
            "Unit": "hours",
            "Dialect": "Egypt"
        },
        {
            "Name": "Saudi",
            "Volume": 356.0,
            "Unit": "hours",
            "Dialect": "Saudi Arabia"
        },
        {
            "Name": "General Arabic",
            "Volume": 639.0,
            "Unit": "hours",
            "Dialect": "Classical Arabic"
        },
        {
            "Name": "Unspecified",
            "Volume": 440.0,
            "Unit": "hours",
            "Dialect": "mixed"
        }
    ],
    "HF Link": "https://huggingface.co/datasets/bitmanagerai-speech-lab/darbuka_arabic_tts_dataset",
    "Link": "https://huggingface.co/datasets/bitmanagerai-speech-lab/darbuka_arabic_tts_dataset",
    "License": "custom",
    "Year": 2026,
    "Language": "ar",
    "Dialect": "mixed",
    "Source": [
        "other"
    ],
    "Domain": [
        "general"
    ],
    "Form": "audio",
    "Annotation Style": [
        "machine annotation"
    ],
    "Description": "Darbuka Arabic brings together Arabic speech, automatic transcripts, and audio-quality measurements for speech recognition and text-to-speech research, with a focus on Egyptian and Saudi speech.",
    "Volume": 0.0,
    "Unit": "hours",
    "Provider": [],
    "Derived From": [],
    "Paper Title": "Darbuka Arabic: A Large-Scale Dataset for Arabic Speech Research",
    "Paper Link": "https://huggingface.co/bitmanagerai-speech-lab",
    "Script": "Arab",
    "Tokenized": false,
    "Host": "HuggingFace",
    "Access": "Free",
    "Cost": "",
    "Has Splits": false,
    "Partial": false,
    "Tasks": [
        "speech recognition",
        "text to speech"
    ],
    "Venue Title": "other",
    "Venue Type": "preprint",
    "Venue Name": "",
    "Authors": [],
    "Affiliations": [],
    "Abstract": "Darbuka Arabic is an Arabic speech dataset comprising 943,731 clips and approximately 1,944 hours of audio, with a focus on Egyptian and Saudi speech. Each 2\u201320-second clip includes transcriptions from three ASR models, a PriorityROVER consensus transcript, quality measurements, dialect labels, and source provenance. Of the total audio duration, 98.01% achieves a predicted DistillMOS score of at least 3.0, and 83.59% exhibits inter-ASR consistency of at least 80%. Distributed as streamable WebDataset shards with accompanying analytical metadata, Darbuka Arabic supports ASR research and TTS data preparation through flexible inspection and subset selection.",
    "Added By": "Kirill Borodin"
}