Mohaddz/sada-diarization-preview

SADA 2022 Arabic Diarization

Training-ready speaker-attributed ASR windows derived from SADA 2022. The source recordings are mirrored at khaledalganem/sada2022.

Splits

• train: 36,004 windows, 202.064 hours, 4,062 recordings
• validation: 853 windows, 4.774 hours, 88 recordings
• test: 901 windows, 5.006 hours, 111 recordings

Total: 37,758 windows and 211.844 hours.

The official SADA train, validation, and test partitions are preserved. Windows are 8–28 seconds, contain 1–4 locally remapped speakers, and have at least 45% annotated speech. Windows from the same source recording do not materially overlap. Ambiguous speaker labels and overlapping speech annotations are excluded from v1.

Audio is resampled to 16 kHz mono FLAC. Background noise and music are retained; no denoising is applied.

Columns

• audio: prepared FLAC window
• segments: speaker, timestamps, transcript, dialect, gender, age, environment
• target: SyvAI-style speaker/timestamp/text sequence
• source_file, source_start, source_end: source provenance
• speaker_count, turn_count, speech_coverage, environment
• audio_sha256, qa_flags, review_status, license

Target format

<|spltoken0|><|t:0.0|>السلام عليكم<|t:2.1|>

Speaker IDs are local to each window and ordered by first appearance. Timestamp tokens use 100 ms resolution.

License and attribution

SADA was created by the Saudi Data and Artificial Intelligence Authority (SDAIA) and the Saudi Broadcasting Authority (SBA). This derivative is released under CC BY-NC-SA 4.0 and is restricted to non-commercial use.

Changes made: selected timestamped regions, excluded unsupported annotations, remapped speaker IDs locally, resampled audio, encoded FLAC, and generated speaker-attributed sequence targets.

train-v1: balanced real + conversational mixtures

train-v1 is a non-commercial Arabic speaker-attributed ASR training configuration derived entirely from SADA2022. It contains 30,000 windows / 164.317 hours:

• 17,727 selected real SADA TRAIN windows (102.061 h)
• 12,273 deterministic conversational mixtures (62.256 h)
• speaker counts 1/2/3/4: 6,000 / 9,000 / 9,000 / 6,000
• synthetic active-speech overlap: 13.95%
• 60 Parquet shards with embedded 16 kHz mono FLAC

The official SADA validation and test material is not included in this configuration. Source recordings are disjoint from held-out recordings and the final audit found zero held-out ID overlap.

Fields

• audio: embedded 16 kHz mono FLAC (bytes, path)
• text: Arabic transcript without control tokens
• target: <|spltokenK|><|t:start|>text<|t:end|> sequence; no count token
• segments: structured local speaker, start/end seconds, and text
• rttm_text: exact RTTM activity labels
• speaker_count, overlap statistics/bucket, and pause durations
• real_vs_synthetic, recipe version, recipe JSON, and source provenance
• audio, row, shard, and source checksums
• split, attribution, and license metadata

Synthetic method and limitations

Mixtures use full annotated Arabic utterances from canonical SADA TRAIN recordings, deterministic seed 30072026, conversational pause sampling, and none/light/medium/heavy overlap buckets. At most two speakers overlap and no utterance is clipped to create an interruption.

SADA speaker labels are file-local. Different source recordings are treated as distinct speaker proxies, but they are not proven to contain different people. The material is Saudi broadcast speech and is not a complete model of all Arabic dialects, channels, or spontaneous conversation.

License and attribution

SADA was created by the Saudi Data and Artificial Intelligence Authority (SDAIA) and the Saudi Broadcasting Authority (SBA). This transformation is released under CC BY-NC-SA 4.0: non-commercial use only, with attribution and ShareAlike. Cite the original SADA dataset and paper.

Integrity: shard manifest 1f7854d8333e2aa139f0ded6aa56f41d3ecaa3ef4988b7dff40b3e75cb94b3ab; validation report 28e6c96a4c0bf1d55b946451bb9f70be1f206901d45c77ba7c31f6dd8001a8eb.