mohammed-bahumaish/arabic-internal-dialect-annotated

Amphion Arabic Dialect-Annotated Utterances

83,266 Whisper-transcribed Arabic utterances exported from the Amphion corpus, each labeled with a country-level dialect code. Annotations come from manual labeling of speakers and channels in the Amphion editor; this is a small bootstrapped seed for the dialect-identification pipeline (see newtts/docs/training/dialect-identification-pipeline.md).

Schema

Each line in data/{dialect}.jsonl.gz matches the Arabic Reddit dialect corpus schema so the two can be concatenated for DID training:

{
  "text": "...",
  "source": "internal:saudi:speaker:789",
  "dialect": "saudi",
  "subreddit": null,
  "kind": "utterance",
  "score": null,
  "char_len": 312
}

• text — Whisper transcription, normalized (NFC, zero-width and tatweel stripped, shadda-first canonicalization, Arabic-Indic digits → Latin).
• source — internal:{dialect}:speaker:{speaker_id} when the dialect came from a manual speaker annotation, internal:{dialect}:channel:{channel_id} when it was inherited from the channel. Speaker dialect takes precedence.
• subreddit, score — always null; kept for column-compatibility with the Reddit shards.
• kind — always "utterance" for this corpus.

Per-dialect counts

dialect	utterances	hours
ye	83,266	455.8

Source split

• speaker (manual): 4,002
• channel (inherited): 79,264

Filters applied

• words ≥ 4
• confidence ≥ 0.5
• dnsmos ≥ 0.0
• only_completed: False
• excluded labels: mixed, unsure

Notes

• This is a domain-matched DID seed (Whisper-transcribed Arabic); pairs with the larger out-of-domain Reddit/QADI sets used for tokenizer/MLM pretraining.
• Country-level macro-F1 is hard for short utterances (intra-Gulf near-indistinguishable). Aggregate to region for production gates.