yt-ckb — Kurdish speech from YouTube

~775k short WAV clips of Central Kurdish (Sorani) speech with their transcriptions, sharded into 56 tar archives (116.4 GB total) plus a metadata.csv index.

At a glance

Files	56 shards + metadata.csv
Total size	116.35 GB
Shard size	2,147,328,000 bytes (~2.0 GiB) each, last one 273,244,160 bytes
Clips	775,172 (with header line: 775,173 lines in metadata.csv)
Index	metadata.csv, 124.4 MB, columns file_name,path,transcription
Audio	WAV, named data/<youtube_id>_<n>.wav
Language	Central Kurdish / Sorani (ckb)

⚠️ The shards are not gzip despite the .gz name

train.tar.gz, train.tar-2.gz, … train.tar-55.gz are uncompressed tar archives with a misleading extension. Verified by reading the file headers:

• no gzip magic (1f 8b) at offset 0 — gunzip fails with Error -3 while decompressing data: incorrect header check
• a valid ustar magic at offset 257 in the first shard, with POSIX octal mode (0000755) and size fields, i.e. a plain tar

Extract them directly:

tar -xf train.tar.gz            # works
tar -xf train.tar-2.gz          # ... shards are independent, extract any subset
# gunzip train.tar.gz           # FAILS — not gzip

Every shard is a self-contained tar (each begins with its own data/ directory entry), so you can download and unpack only the shards you need.

Usage

import csv, tarfile

# 1) read the index
with open("metadata.csv", encoding="utf-8") as f:
    rows = list(csv.DictReader(f))
print(len(rows), rows[0])
# {'file_name': 'data/4msEW3nk6o8_0.wav', 'path': 'data/4msEW3nk6o8_0.wav',
#  'transcription': 'کەس نەبێتەوە'}

# 2) pull audio out of one shard
with tarfile.open("train.tar-2.gz") as tf:
    m = tf.getmember("data/HlzeYMJ4kxE_410.wav")
    audio = tf.extractfile(m).read()

file_name and path are identical in every sampled row; use file_name as the key to join against the tar members.

Sample transcriptions:

کەس نەبێتەوە
تاوەکو سەرنجی ملیۆنەها کەسی لە چەند ساڵێکدا بۆلای بەرنامە و نووسینەکانی خۆی ڕاکێشا
مامۆستا شێخ محەممەد بەرزنجی

How it was built

Audio from Kurdish-language YouTube videos, segmented into short clips (a few seconds each) with one transcription per clip; clips are grouped by source video id and packed into fixed-size tar shards.

Known limitations

• Rights: the recordings come from third-party YouTube videos. Reuse is constrained by YouTube's terms and the original creators' rights; this dataset is not a licence grant for the underlying audio.
• Transcriptions are automatic, not human-verified: expect missing punctuation, occasional mis-segmentation at clip boundaries, and no speaker labels.
• Clips are cut mechanically, so a clip can start or end mid-word (see the _0, _1, … suffixes, which are just clip indices within a video).
• No split is defined and no duration field is present — audio length must be read from the WAV headers.
• Downloading the whole set costs 116 GB; always stream or fetch single shards.
• metadata.csv mixes videos, speakers and audio quality; there is no per-video metadata beyond the id embedded in the filename.

Related

• razhan/diyako_hashemi_yt — a different YouTube-derived Kurdish speech set.
• razhan/DOLMA-speech — studio-style Kurdish and Iranian-language speech with English translations.