Diyako Hashemi — Kurdish speech from YouTube (diyako_hashemi_yt)

24,207 short Central Kurdish speech clips with transcriptions, cut from a Kurdish-language video series and paired with the spoken text.

At a glance

Rows	24,207 (single train split)
Columns	audio, transcription
Parquet on disk	6.77 GB across 17 shards (7.84 GB uncompressed)
Row groups	15 per shard (1,000 rows per group)
Script	Arabic script, 100% of rows
Task	automatic speech recognition

Field profile

Measured over a 1,000-row sample (shard 11):

Column	Value
transcription	mean 49 chars, median 46, max 124
audio	WAV, path like zimani kurdi 201305_16 120).زمانی کوردی_167.wav

Sample transcriptions:

لەم ژیانە نوێدا زۆر گۆڕان بەسەر ئەو وشانە نەهاتوو ئەوانەیە کە ئێستە باسم
بۆتان لەبەرەوەی کە قۆناغەکە خۆی وەک خۆی ماوەتەوە بەڵام

The audio filenames keep the source video's title (زمانی کوردی, "the Kurdish language") plus a sequential clip index, so clips can be grouped back into their source recording.

Usage

from datasets import load_dataset

ds = load_dataset("razhan/diyako_hashemi_yt", split="train")
print(ds[0]["transcription"])
print(ds[0]["audio"]["path"])

Filter or group by the filename prefix if you need a speaker-disjoint or video-disjoint evaluation split — the dataset itself does not carry those fields.

Known limitations

• No speaker or video id column. Grouping must be done by parsing the audio filename, which contains spaces, parentheses and non-ASCII characters — quote paths carefully in any shell script.
• Transcriptions look automatic and are unpunctuated; there are no sentence boundaries and no alignment information.
• Continuous speech was cut into fixed-length clips, so clips can begin or end mid-word.
• One source series: register, accent and recording setup are all similar, which limits how well models trained here generalise.
• Rights: derived from third-party YouTube video. Reuse is bound by YouTube's terms and the original creator's rights; this repository is not a licence grant.
• No duration column — read lengths from the WAV headers.
• 6.77 GB of audio: stream or download shards individually.

Related

• razhan/yt-ckb — a much larger (116 GB) YouTube-derived Kurdish speech set.
• razhan/asosoft-speech, razhan/hawrami_speech — studio read speech.