A deduplicated collection of dialectal Arabic sentences assembled from openly-downloadable sources, every line tagged with a BCP-47 lang code. Saudi Arabic is the focus, but all labelled dialects are retained. Built as the text side of a Saudi TTS / phonemizer pipeline.

Files

all.tsv — the corpus: id<TAB>lang<TAB>source<TAB>text.
all.enriched.tsv — adds two phonetic columns: id<TAB>lang<TAB>source<TAB>text<TAB>diac<TAB>ipa, where
- diac is dialect-aware diacritization (arbtok rawi-lattice fusion: the bundled rawi ensemble scores dialect-licensed hypotheses under a joint beam; written diacritics are hard constraints), and
- ipa is arbtok's pausal dialect-aware transcription of the sentence, routed by the line's lang code — e.g. قهوة → Najdi ɡahawa (ar-SA), Egyptian ʔahwa (ar-EG), MSA qahwa. Pipeline accuracy on the blind-verified Arabic gold: PER 0.015 on vocalized input, 0.177 on bare input (vs 0.357 for espeak-ng ar).
summary.txt, provenance.json — per-lang / per-source counts, licenses, and status.

Split by dialect by filtering the lang column (e.g. ar-SA, ar-EG, ar-x-lav).

Intended use

Text source for downstream TTS/phonemizer pipelines: filter by lang to get a dialect's sentence pool, then either synthesize directly (text) or consume the pre-computed diac/ipa columns from all.enriched.tsv as phonemizer input or regression fixtures. Not intended as a standalone NLP/LM corpus without checking per-source licensing in provenance.json first.

Noise / quality profile

Text-only — no audio in this repository, so no bandwidth/DNSMOS/UTMOS profile applies. The ipa/diac columns' reliability is the closest analogue to a quality metric: PER 0.015 on vocalized input vs. 0.177 on bare input (see below); treat that gap as the confidence signal per row, not a per-row correctness guarantee.

Provenance and licensing

Sources vary in license (several research-use or CC-BY-NC; some unspecified) — see provenance.json. This repo is private and intended for internal research use; verify each source's terms before any redistribution. The diac and ipa columns are provisional (the diacritizer is MSA-trained; dialect phonology is approximate — Saudi is routed to Najdi) and are meant as input to human phonetic QA, not as final labels.

Sources and reproduce

Built by Salesteq-Agent/salesteq-dataprocessing-scripts: scripts/donor_tts/gather_dialect_text.py pulls raw Arabic sentences from a registry of sources (local dialect corpora, UBC-NLP multi-dialect HuggingFace sets, and Saudi-specific sets), cleans and globally deduplicates them, and tags every line with a BCP-47 lang code derived from the source's own dialect label. provenance.json in this repo records exactly which sources ran, their licenses, and per-source counts — it is the authoritative source list. The diac/ipa enrichment columns are produced with arbtok (dialect-aware diacritization + pausal IPA, routed by each line's lang code).

Related datasets

The text side of the Saudi TTS / phonemizer pipeline: filter by lang to get a dialect's sentence pool, then synthesize or feed the pre-computed ipa to a phonemizer. The synthetic audio counterpart (text rendered to speech across dialects) is Salesteq/arabic-multidialectal-omnivoice-synthetic.

Downloads last month: 17

Total file size:

71.2 MB