masterofaudio2077/egy-g2p-dialects

Egyptian Arabic dialect G2P (Cairo, Sharqia, Sohag, Aswan)

Arabic text in, phonemic IPA out, for a chosen Egyptian sub-dialect. Text only, no audio.

Read this first: the pronunciations are synthetic and unverified. They were written by a large language model (ChatGPT) from short, draft descriptions of each dialect, then filtered by a program that checks structure only. No native speaker has reviewed them (verified = false, review_status = unreviewed). Use them as a starting point for training or for a review pass, not as ground truth.

What a row is

One row = one (sentence, dialect) pair. Every sentence is pronounced in all four dialects, so rows with the same group_id are the same sentence in Cairo, Sharqia, Sohag and Aswan. Target size: 10,000 sentences x 4 dialects = 40,000 rows.

column	meaning
id, group_id	row id and the id shared by the 4 versions of a sentence
item_n	number of the sentence inside its category
text	the Arabic sentence, diacritics removed, otherwise unchanged
normalized_text	conservative normalization (alif forms and ى unified; ة and ه kept apart)
dialect, region, governorate, locality	arz_cairo, arz_sharqia, arz_sohag, arz_aswan; locality is empty
category	common, conversational, dialect_sensitive, minimal_pairs, special
text_source	Hub dataset the sentence came from, or generated
ipa	phonemic IPA of the whole sentence, words separated by spaces
words	JSON list of {w, ipa}, one entry per orthographic word (aligned)
variants	JSON, alternate pronunciations by word index (written a~b by the model)
rule_ids	dialect rules (C1, C2, V1) that explain differences from the Cairo form
annotation_source, verified, review_status	llm_generated, false, unreviewed
agreement	empty; reserved for a later second-pass agreement score
review_sample	about 1% of sentences flagged for native-speaker review
split	dev (10%) or test (90%) by hash of the sentence; no train split on purpose
dedup_key, seed	normalized-text hash; the generation spec

Dialects and how well they are documented

id	region	documentation	notes
arz_cairo	Cairo	high	the reference; other dialects are written as differences from it
arz_sharqia	Sharqia (Delta, east)	medium	the region is heterogeneous (ق as ʔ or ɡ varies by town)
arz_sohag	Sohag (Upper Egypt)	medium	ق as ɡ, ج as ɟ or dʒ
arz_aswan	Aswan (far south)	low	thinly documented; modelled the same as Sohag, Nubian contact is not modelled

The differences allowed against Cairo are a short fixed list (ق: ʔ to ɡ; ج: ɡ to ɟ or dʒ; ē/ō to aj/aw). A word may differ from the Cairene form only through one of those rules, which is why most words are identical across dialects. This keeps the model from inventing differences, but it also means real differences outside the list are missing, and Sohag and Aswan are nearly indistinguishable in this version.

How the rows were cleaned

A program checks every item: allowed IPA symbols only, one token per word, no three consonants in a row, each consonant must have a matching Arabic letter, and a dialect may differ from Cairo only through a listed rule that acts on its own letter. Failures are dropped. A few mechanical rules were repaired instead of dropped (emphatic letters, the Upper Egyptian ق and ج rules, the Sharqia ق variant). In Sharqia every word with ق lists both ʔ and ɡ forms; Sohag and Aswan are identical in this version.

Where the sentences come from

About 60% existing Egyptian sentences (categories common, conversational), taken from the author's own Hub datasets (chats, persona dialogues, decision-data prompts, call scripts). Those datasets were themselves generated by ChatGPT, so "existing" does not mean "human-written". About 40% were written by ChatGPT for this dataset to cover dialect-sensitive words, minimal pairs and special cases.

Limits

• Synthetic and unverified; errors in the IPA are expected, especially for Sohag, Sharqia and Aswan.
• The filter checks structure (allowed IPA symbols, word alignment, differences explained by a rule). It cannot tell whether a pronunciation is how a region really speaks.
• Phonemic only; no phonetic detail, no audio, no village-level variation.
• Arabizi input is out of scope.
• Check the licences of the source datasets and of ChatGPT output before reuse; cc-by-4.0 above is a placeholder.