DOLMA-speech

Speech for Central Kurdish and five neighbouring Iranian languages — Gilaki, Mazanderani, Zazaki, Talysh and Laki — each utterance paired with a sentence transcription in its own language and an English translation. 41,700 utterances in total, split into one config per language.

Configs (rows per split)

Config	train	test	Script
hawrami	10,166	1,263	Arabic
southern_kurdish	5,912	757	Arabic
english	17,567	—	Latin
gilaki	2,961	625	Arabic
mazanderani	875	249	Arabic
laki_kurdish	755	313	Arabic
zazaki	201	50	Latin
talysh	5	1	Arabic
Total	38,442	3,258	—

Columns

Column	Notes
audio	MP3 (voice_<id>.mp3), decoded automatically by datasets
sentence	the utterance in the config's language
english	English translation of the same sentence (not in the english config)
gender	male / female
language	the config name, e.g. southern_kurdish
speaker_id	SPEAKER_<n>
duration	seconds (short utterances, ~0.8 – 11 s)
original_full_path	recording path, e.g. dataset/talysh/voice_16525.mp3

Because english is present alongside sentence, the dataset doubles as a parallel translation corpus — the same English sentence appears across configs, e.g. "he left the hotel and turned up the street in the direction for the restaurant." with southern_kurdish, mazanderani and gilaki renderings of it.

Samples

southern_kurdish : لە هتلەگە هاتەو دەیشت ئەڵگردرامەو وەرە ڕستورانەگە کەفتە ڕێیا.
mazanderani      : همه چی روشن بی یه
zazaki           : No yew kitab o derheqê yew mêrdekî de yo ke yew girawa venge de tena mendo.
talysh           : اَ نی شَه و بَرْش دَوَسَه.
hawrami          : ئەی دنیاء
english          : he left the hotel and turned up the street in the direction for the restaurant.

Usage

from datasets import load_dataset

ds = load_dataset("razhan/DOLMA-speech", "southern_kurdish")
print(ds["train"][0]["sentence"], "->", ds["train"][0]["english"])

# all configs at once
full = load_dataset("razhan/DOLMA-speech")

Known limitations

• Very uneven coverage. english (17,567 rows) dominates; talysh has 5 rows and zazaki 201 — those two configs are demonstration scale, not training scale.
• Script varies by language: Zazaki is Latin-script, the rest Arabic-script. Do not tokenise all configs with one Arabic-script vocabulary.
• The english column is a translation, so it inherits translationese; it is not a human reference parallel corpus.
• Content is news/political in register and shares source sentences across configs, so splits can leak the same underlying sentence into train and test for different languages (fine within a config, risky across configs).
• speaker_id/gender were derived automatically; expect occasional mislabels.
• Durations are short (mostly a few seconds) and no long-form recordings exist.

Related

• razhan/hawrami_speech — dedicated studio Hawrami ASR data.
• razhan/asosoft-speech — Sorani read speech.
• razhan/hac_to_en — Hawrami → English text pairs.