Asosoft speech (asosoft-speech)

3,840 short Central Kurdish (Sorani) utterances with sentence-level transcriptions — studio-quality read speech, with an audited test split.

At a glance

Rows	3,840 — train 3,240 / test 600
Columns	audio, transcription (string), duration (float64)
Parquet on disk	702.4 MB
Audio format	WAV, 16-bit (files named F01146001.wav, …)
Utterance length	3.16 – 13.49 s in the test split
Language	Central Kurdish / Sorani (ckb)
Task	automatic speech recognition

Field profile

Measured on the test split (600 rows):

Column	Value
transcription	mean 58 chars, median 56, max 114; 100% Arabic script
duration	min 3.159 s, max 13.489 s (mean ≈ 5.7 s on the first rows)

Sample transcriptions:

پاش موسڵمان بوونی بڕیاری دا بەرگی باڵاپۆشیش لە بەر بکات.
ئافرەتانی کورد بۆ لە فرەژنی دەترسن.

Usage

from datasets import load_dataset

ds = load_dataset("razhan/asosoft-speech")
print(ds["test"][0]["transcription"])
audio = ds["test"][0]["audio"]      # decoded array + sampling rate

Test-split corrections

The test split was checked against the recordings and the transcriptions listed below were corrected. The audio itself is unchanged — only the reference text was fixed, so if you have cached an older copy of the split, re-download it.

Row	Original reference	Corrected reference
49	ئەو کلیپە بریتییە لە چیرۆکی خۆشەویستییەکی قووڵ کە بە خیانەت کۆتایی پێدێت	ئەو کلیپە بریتییە لە چیرۆکێکی خۆشەویستی قوڵ کە بە خەیانەت کۆتایی پێدێت
432	یاسای بودجەی عێراق لە بەرژەوەندیی هاوڵاتییانی هەرێمی کوردستانە	removed — the first two words are cut off in the recording
433	کۆمپانیای دەرهێنانی نەوت گەشبینە بە کارکردنی لە هەرێمی کوردستان.	removed — the first word is cut off in the recording
434	حکوومەتی ئێران بە فەرمی ناوی دراوی وڵات دەکات بە تمەن	removed — the first word is cut off in the recording
485	ئێمە پێشنیاری گۆڕانکاری لە بڕگەی چوارەمی پڕۆژەکەدا دەکەین.	پێشنیاری گۆڕانکاری لە بڕگەی چوارەمی پڕۆژەکەدا دەکەین.

Because three clips were dropped for truncation, the test split holds 600 usable utterances; keep that in mind when comparing against published baselines computed on the uncorrected file.

Known limitations

• Read speech only: no spontaneous conversation, no background noise, no telephone channel — models trained here will not transfer directly to broadcast or phone audio.
• Transcriptions are not punctuation-normalised (، and . are both used).
• Audio is stored as WAV inside parquet (~180 KB per utterance), so the whole set is ~700 MB.
• No speaker IDs or gender labels are provided, so you cannot build a speaker-disjoint split from this dataset alone.
• Small: ~3,840 utterances is fine-tuning scale, not pretraining scale.

Related

• razhan/asosoft-speech-v2 (private) — the same recordings with the 2026-08-04 orthography corrections and a transcription_original column plus CORRECTIONS.jsonl.
• razhan/DOLMA-speech — multi-language (Kurdish, Gilaki, Mazanderani, Zazaki, Talysh, Laki) speech with English translations.