Riste Mirwari — audiobook speech (ristey_mirwari)

31,535 transcribed speech segments cut from a reading of the book ڕشتەی مرواری (Rishtey Mirwari), with a speaker and gender column alongside each segment.

At a glance

Rows	31,535 (single train split)
Columns	audio, file, transcription, speaker, gender, book, is_gold_transcript
Parquet on disk	4.99 GB across 11 shards (5.50 GB uncompressed)
Row groups	29 per shard (1,000 rows per group)
Audio	WAV, voice-isolated tracks: iUUYlq9RROA_01_vocals_58.wav
Language	Central Kurdish / Sorani (ckb), Arabic script
Task	automatic speech recognition

Field profile

Measured over all 31,535 rows (not a sample):

Column	Value	Distinct values
transcription	mean 55 chars, median 57, max 107	—
file	29 chars, <youtube_id>_<track>_vocals_<segment>.wav	—
speaker	SPEAKER_00 in every row	1
gender	Male in every row	1
book	ڕشتەی مرواری in every row	1
is_gold_transcript	False in every row	1

Sample transcription:

دیسان هەر تێریشم نەئەخوارد ئەمە ویست زیاتر ببێ لەپاشا کە ئەمبینی گەلێکی تر وەکوو من
بینایی چاوان فەرمووی ماشەڵا ماشەڵا لە خێڵی ڕۆغزایی لە قەیتوولی سوورکەوە سوار

The filename structure is informative: the audio was separated from a source video (iUUYlq9RROA), the vocal stem was kept (_vocals_), and the result was cut into numbered segments (_58, _580, …). file and audio.path carry the same value.

The speaker and gender columns carry no information

speaker is SPEAKER_00 for all 31,535 rows, gender is Male for all of them, and book is constant too. Treat all three as dead columns:

• you cannot build a speaker-disjoint or gender-balanced split from this dataset — the metadata cannot distinguish speakers at all;
• if you need speaker diversity, group by the source video id embedded in file (iUUYlq9RROA, …) instead, or use a dataset with real diarisation labels;
• an apparently perfect "single speaker" result from a model trained here is a property of the data, not evidence of speaker generalisation.

Usage

from datasets import load_dataset

ds = load_dataset("razhan/ristey_mirwari", split="train")
row = ds[0]
print(row["file"], row["book"])          # speaker/gender are constants
print(row["transcription"])

Known limitations

• is_gold_transcript is False in all 31,535 rows — treat every transcription as automatic (ASR/alignment output) rather than a human-verified reference.
• No usable speaker or gender labels (see above).
• One book, read aloud: narrow register (literary prose), a single reading style, and one source video id per chapter — no cross-speaker or cross-domain coverage.
• Continuous prose was cut into short segments, so segments can start or end mid-sentence; there is no duration column (read it from the WAV headers).
• Rights: derived from a third-party video/audiobook. Reuse is constrained by the platform's terms and the rights holders' permissions; the underlying book is in copyright.
• 4.99 GB of audio: stream or download shards individually.

Related

• razhan/kteb, razhan/kteb-dataset — Kurdish book text.
• razhan/asosoft-speech, razhan/hawrami_speech — studio read speech with real speaker ids.