Blgn94/mongolian-stt-dataset

Mongolian Speech Dataset (v24 corpus)

Mongolian (Cyrillic Khalkha) read speech for ASR fine-tuning: 146.9 hours across Common Voice v24, FLEURS, and MBSpeech.

2026-07-30 — two changes, read this if you pulled before that date.

1. YouTube-sourced audio removed. 598 clips (559 train / 39 validation, ~1.1 h) are gone. Every remaining row is read speech from a redistributable public corpus.
2. This repo now hosts the v24 corpus. It previously held the v20 blend (57,320 train / 3,017 validation). The Common Voice portion moved from warmestman/common-voice-20-mn-normalized to onlysainaa/common-voice-mn-24, and row counts changed as shown below. The former -v24 repo is retired.

Re-download if you have an older local copy — it still contains the removed rows and, if it predates this, a different corpus.

Splits

Split	Total	commonvoice (v24)	fleurs	mbspeech
train	95,428	87,988	3,770	3,670
validation	2,260	1,869	215	176

Duration by source (measured):

source	train	validation	avg clip
commonvoice	123.44 h	2.88 h	~5.1 s
fleurs	13.20 h	0.75 h	~12.6 s
mbspeech	6.34 h	0.29 h	~6.2 s
total	142.99 h	3.91 h

Known limitations — read before benchmarking

• Heavy text reuse. 95,428 train rows cover only 11,669 distinct sentences (8.2x overall). Common Voice is the worst offender: 87,988 rows / 6,083 distinct sentences = 14.46x.
• Validation is not text-disjoint. 2,056 of 2,260 validation sentences (91.0%) appear verbatim in train — 99.9% of the Common Voice rows. In-domain WER measured on this split largely reflects text memorisation, not acoustic modelling. Build a text-disjoint dev split if you need an honest number.
• Not speaker-disjoint. The split is random at row level, so the same speakers appear in train and validation.
• All read speech. Since the YouTube removal there is no spontaneous or conversational audio in this corpus. Expect models trained on it to degrade on natural speech.
• ⚠️ other is unvalidated. The Common Voice pool includes v24 other clips, which are not human-confirmed and may contain mistranscriptions/misreads. This was a deliberate volume-over-purity choice.

What changed vs the v20 corpus

• Common Voice source: warmestman/common-voice-20-mn-normalized (v20) → onlysainaa/common-voice-mn-24 (v24, raw corpus).
• Pool: v24 validated + other (91,791 clips).
  • train = (validated ∪ other) − dev − test
  • validation = dev (verified clips only)
  • test is held out (not included).

Previous corpus (v20 blend): train 57,320 / validation 3,017.

Features

• audio — 16 kHz mono. Mixed on-disk encoding: fleurs/mbspeech are stored as 16 kHz WAV bytes; the v24 commonvoice clips are stored as their original mp3 bytes with the Audio feature pinned to 16 kHz (resampled on read). This keeps the folder ~5 GB instead of ~14 GB. The distinction is invisible to consumers — every row decodes to a 16 kHz array.
• sentence — transcript (Cyrillic, capitalized + punctuated, whitespace-cleaned).
• source — commonvoice | fleurs | mbspeech.

Usage

from datasets import load_dataset
ds = load_dataset("Blgn94/mongolian-stt-dataset")

Licensing

Combined corpus is CC-BY-4.0. Common Voice is CC0-1.0; onlysainaa/common-voice-mn-24 is declared Apache-2.0 on the Hub. FLEURS is CC-BY-4.0. With the YouTube portion removed, every component now carries an explicit redistribution license.