Rabe3/saudi-tts-higgs-synthetic-200k

Saudi Arabic Synthetic TTS Corpus (Higgs TTS 3)

199,955 single-speaker Saudi Arabic utterances, 24 kHz mono, ~306 hours.

from datasets import load_dataset

ds = load_dataset("Rabe3/saudi-tts-higgs-synthetic-200k", split="train")
print(ds[0]["text"])
print(ds[0]["audio"]["sampling_rate"], ds[0]["audio"]["array"].shape)

How it was made

Teacher	bosonai/higgs-tts-3-4b via vLLM-Omni
Voice	zero-shot clone from a single 8.6 s Saudi reference clip
Text	199,955 unique Saudi dialect lines, undiacritized
Code-switching	~30% of lines mix Arabic and English
Sample rate	24 kHz mono

Text preparation

Source lines were cleaned before synthesis: Unicode NFKC normalisation (Arabic presentation forms → base letters), Farsi/Urdu confusable codepoints mapped to Arabic, bidi and zero-width marks stripped, markdown artefacts removed, tanween ordering unified, and lines containing foreign scripts dropped. 99.26% of the source corpus survived unchanged.

Diacritics were deliberately removed. The source carried a hand-built pronunciation lexicon (وِشْ / بَسْ / الحِينْ), but this model's tokenizer has no diacritised-Arabic entries and falls back to character splitting — بَسْ becomes 4 tokens against 1 for بس, a 3.12x inflation on exactly those words. Measured on identical sentences, diacritised input produced 1.90x the audio duration and markedly worse fluency. The Saudi accent comes from the reference clip instead.

Quality control

Every clip was screened for: unreadable/empty files, silent audio, non-finite samples, clipping, low voiced fraction, excessive trailing silence, and duration-vs-text-length outliers (runaway or truncated generation). 199,955 of 200,021 passed (99.97%); 66 were dropped.

Limitations

• Audio is synthetic, not human speech.
• QC was audio-only. It cannot detect word-level repetition or text/audio mismatch that stays within a plausible duration. A sampled ASR check suggested roughly 1% of clips may carry such defects.
• Single speaker — no speaker diversity.
• The teacher has limited Saudi dialect coverage in its tokenizer, so some dialect-specific words may be pronounced with more standard Arabic phonology.