Rabe3/youtube-001-part01

YouTube-300 — Saudi Arabic Speech

Single-speaker-per-clip Saudi Arabic speech mined from YouTube, cut into short utterances with transcripts, and filtered for TTS / voice-cloning use. Built with youtube-to-dataset.

Snapshot. This is a partial release: the extraction run is still in progress over a larger playlist. The numbers below describe what is currently uploaded.

At a glance

Clips	77,304
Duration	128.3 hours
Source videos	323
Channel	سعوديه بدون تكلف (1 channel)
Speakers	323 (speaker is per-video; cross-video linking not applied)
Language	Arabic (Saudi dialect), 100% of clips
Audio	24 kHz mono WAV, EBU R128 normalised per video
Clip length	2.0–12.4 s (mean 5.97 s, median 5.72 s)
Mean SQUIM MOS	4.60
Median CER (YouTube subs vs. Cohere ASR)	0.024

Usage

from datasets import load_dataset

ds = load_dataset("Rabe3/youtube-300", split="train")
ex = ds[0]
ex["audio"]   # decoded waveform at 24 kHz
ex["text"]    # transcript

Streaming works too, which is the easier option at ~21 GB:

ds = load_dataset("Rabe3/youtube-300", split="train", streaming=True)

Fields

Core

Field	Description
audio	24 kHz mono clip
audio_file	Clip filename, {video_id}_{chunk_index:04d}.wav
text	Transcript used for training (see text_source)
duration, start, end	Clip length and its span in the source video
speaker	Diarised speaker id, scoped to the video ({video_id}_SPK{n})
global_speaker	Cross-video speaker id; null — linking has not been run
language, asr_language	Decoded language of the clip
video_id, video_url, title, channel, upload_date	Provenance

Transcript provenance

text_yt is the YouTube subtitle text for the clip's span, text_cohere is an independent transcription by CohereLabs/cohere-transcribe-arabic-07-2026, and cer_yt_vs_cohere is the character error rate between them. text_source records which one text came from: yt_auto (99.8% of clips) or cohere (used when a video had no subtitles). Agreement between the two independent sources is the dataset's main fidelity guarantee — clips where they disagree were dropped.

Quality measurements (each clip also passed a threshold on these)

squim_mos, squim_stoi, squim_pesq, squim_si_sdr, snr_db, peak_dbfs, rms_dbfs, clipping_ratio, lufs, music_score, music_label, speech_score, vocal_ratio, accompaniment_ratio, align_score, aligned_words, boundary_clipped, speaker_conf, asr_confidence, asr_confidence_alt, arabic_ratio, script_ratio, chars_per_sec, sample_rate.

How clips were selected

Every clip in this dataset cleared all of the following. 16,469 candidate clips were rejected.

• Speech only. Silero VAD for speech regions; DiariZen (BUT-FIT/diarizen-wavlm-large-s80-md-v2) for speaker turns. Regions with overlapping speakers are dropped outright — two voices in one clip are poison for voice cloning.
• No music. Demucs vocals-vs-accompaniment energy ratio ≤ 0.10, plus an AST AudioSet music score ≤ 0.20.
• Speaker consistency. ECAPA embedding within cosine distance 0.45 of the speaker's own centroid.
• Transcript fidelity. CER between YouTube subtitles and Cohere ASR ≤ 0.35; 4–18 characters/second; ≥ 2 words; ≥ 60% of characters in the expected script.
• Alignment. Forced alignment with MahmoudAshraf/mms-300m-1130-forced-aligner; clips whose boundary cuts a word in half are dropped.
• Audio quality. SQUIM MOS ≥ 3.0, STOI ≥ 0.75, PESQ ≥ 1.8, SNR ≥ 12 dB, clipping ratio ≤ 0.001, peak between −35 and −0.5 dBFS.

Limitations

• One channel, one dominant voice. All clips come from a single YouTube channel, so accent, recording chain, and topics are narrow. speaker ids are per-video, meaning the same person carries a different id in every video; do not treat the 323 ids as 323 distinct people.
• Transcripts are machine-generated. Mostly YouTube auto-captions, validated against a second ASR model rather than by humans. Diacritics are absent and dialectal spelling is inconsistent.
• Partial run. Shard count and totals will change as the extraction continues.

Provenance and license

The audio is derived from publicly available YouTube videos and the underlying content remains the property of its original creators. No license is granted to the source material here; it is redistributed for research use. If you are the rights holder and want material removed, open a discussion on this repo.

metadata.jsonl at the repo root is the pipeline's original per-clip metadata for the same 77,304 clips, in the same order as the parquet rows.