btsee/fleurs-mn

FLEURS Mongolian (cleaned)

A quality-filtered, normalised subset of google/fleurs, mn_mn, prepared for training Mongolian (Khalkha Cyrillic) text-to-speech with oron-tts.

Built by oron-cleaner. Every threshold was calibrated on this corpus, and every number and column on this page is read from the shipped data rather than asserted.

from datasets import load_dataset
ds = load_dataset("btsee/fleurs-mn", split="train")
print(ds[0]["text"], ds[0]["audio"]["sampling_rate"])

What is in it

clips	1,908
audio	4.30 h
distinct sentences	1,285
identified speakers	0
format	parquet shards, audio inline as 24 kHz mono wav
upstream audio	16 kHz native, so bandwidth is capped near 7.7 kHz
median clip	7.76 s
median measured bandwidth	7,734 Hz
median DNSMOS OVR	3.90
median alignment score	0.899
median ASR CER	0.112

Splits

split	clips	hours	speakers
train	1708	3.84	0
withheld	200	0.46	0

All rows ship in one file set; the split column says which split a clip belongs to, so you can filter without re-downloading.

withheld holds clips whose sentences were removed from train, so text used as an evaluation target is genuinely unseen. It is a text holdout, not a speaker holdout — the two are separate requirements, and intersecting them leaves too little data to be useful.

Gender

gender	clips	hours
female	1012	2.33
male	896	1.97

gender_resolved is declared where the contributor gave it, propagated across a speaker's other clips where one of them declared it, and otherwise inferred from median F0. gender_source records which of the three applied, so inferred labels can be dropped if you do not trust them.

Columns

column	type	example
audio	audio	24 kHz mono wav, stored inline
align_score	float64	0.9331557547294662
asr_transcript	string	дээшарсан ахиу ихэвчлэн илүү зөөлөн дуу гарг
audio_path	string	wavs/fleurs_10009537996232518777.wav
bandwidth_hz	float64	7578.125
cer	float64	0.06481481481481481
clip_id	string	fleurs_10009537996232518777
dnsmos_bak	float64	3.425452424609938
dnsmos_ovr	float64	3.2594630522193726
dnsmos_p808	float64	2.771478898233047
dnsmos_sig	float64	3.2443036437034607
duration_s	float64	6.748
gender	string	female
gender_resolved	string	female
gender_source	string	declared
id	int64	1286
lang_group_id	int64	2
lang_id	int64	60
language	string	Mongolian
len_ratio	float64	0.9629629629629629
mean_f0_hz	float64	213.8156923536814
num_samples	int64	147840
path	string	/root/.cache/huggingface/datasets/downloads/
pitch_confidence	float64	0.7014218009478673
raw_transcription	string	Дээш харсан нахиу ихэвчлэн илүү зөөлөн дууг
snr_db	float64	19.123080681097697
speaker_known	bool	False
split	string	train
text	string	Дээш харсан нахиу ихэвчлэн илүү зөөлөн дууг
transcription	string	дээш харсан нахиу ихэвчлэн илүү зөөлөн дууг

What the derived columns mean

column	meaning
text	normalised transcript — the training target
clip_id	stable identifier, unique in this corpus
audio_path	path in the source corpus layout, kept for traceability
gender_resolved / gender_source	see above
snr_db	speech-region RMS against true non-speech RMS
bandwidth_hz	measured lowpass shelf, not the container sample rate
dnsmos_sig / dnsmos_bak / dnsmos_ovr / dnsmos_p808	learned perceptual quality
align_score	MMS-FA forced-alignment confidence against text
cer / len_ratio / asr_transcript	ASR agreement with text
mean_f0_hz / pitch_confidence	pitch, used for gender inference
duration_s	duration of the shipped audio, after edge trimming
split	train, validation, test or withheld

How it was cleaned

Each clip passes every gate below or is dropped. Gates run in order, and a clip stops at its first failure.

1. Duration — outside 1–20 s is dropped.
2. Clipping and DC offset — consecutive full-scale samples, or a DC bias.
3. Silero VAD — a clip that is mostly silence is a failed recording.
4. SNR — speech-region RMS against true non-speech RMS on the untrimmed signal. When a clip is spoken end to end there is no noise floor to measure; that is recorded as unmeasurable rather than treated as a failure, and the background judgement is left to DNSMOS-BAK.
5. Bandwidth — the real lowpass shelf, which is what an MP3 encoder or a 16 kHz source leaves behind.
6. DNSMOS (SIG / BAK / OVR) — learned perceptual quality.
7. MMS-FA forced alignment against the given transcript. This is the primary transcript gate: a low score means the audio does not contain those words.
8. ASR CER as a secondary check.

Every measurement is kept as a column, so you can re-filter to your own thresholds instead of trusting ours.

Text is normalised with oron_tts.text.MongolianNormalizer: numeral expansion that refuses rather than guesses case suffixes, small-caps folding (ЭЗЭНий becomes Эзэний), and abbreviation and unit expansion. The normalised string is simultaneously the published text, the CER reference and the training target, so a wrong expansion would be published, scored against itself, and learned.

provenance.json records the pinned model and dataset revisions, package versions, a fingerprint of the normaliser source, and a content hash.

=== Corpus summary ===

split          clips    hours  speakers   male h  female h
train          1,708      3.8         1      1.8       2.0
validation         0      0.0         0      0.0       0.0
test               0      0.0         0      0.0       0.0
withheld         200      0.5         1      0.1       0.3

Text diversity:
  clips                 1,908
  distinct sentences    1,285  (1.48x repetition)
  letters covered       34/35
  letter bigrams        714 distinct
  NEVER APPEARS         щ
  under 100 occurrences ъ

CER by source (the gate's recogniser trained on Common Voice):
  source         clips   median    mean
  fleurs          1908    0.112   0.113
  A markedly lower median for cv is contamination, not quality.

total: 4.3 h   male 2.0 h   female 2.3 h

Acceptance criteria:
  total >= 25 h            FAIL  (4.3 h)
  male  >= 5 h             FAIL  (2.0 h)
  male speakers >= 3       FAIL  (1)

Using it for oron-tts finetuning

python scripts/build_f5_dataset.py --corpus <corpus-dir> --name oron_mn --splits train
python scripts/preflight.py --data data/oron_mn_pinyin --target-updates 30000

manifest.jsonl ships alongside the parquet in the original corpus layout, which is what build_f5_dataset.py expects.

Considerations

• Licence: CC-BY-4.0. Read-speech recordings collected for a public benchmark and released under CC-BY-4.0. Attribution is required.
• Voice cloning. This corpus is suitable for training zero-shot TTS, which can clone a voice from seconds of audio. Contributors consented to speech research, not to having their individual voice reproduced. Do not synthesise a named or identifiable person's voice without that person's consent.
• Coverage. One language variety, read speech. Accent, dialect and spontaneous-speech coverage is untested.
• Inferred gender labels. Where gender_source is f0, the label is a guess from pitch, not a self-report. Filter on gender_source if that matters.
• The gates encode judgements. Thresholds were calibrated on this data; a different application may want different ones.

Citation

@misc{orontts_fleurs_mn,
  title  = {FLEURS Mongolian (cleaned)},
  author = {Badral, Battseren},
  year   = {2026},
  url    = {https://huggingface.co/datasets/btsee/fleurs-mn}
}

Please also cite the upstream corpus: google/fleurs, mn_mn.

Used by

btsee/oron-tts — Mongolian text-to-speech, trained on this corpus.