btsee/WorldSpeech-mn

WorldSpeech Mongolian (cleaned)

A quality-filtered, normalised subset of disco-eth/WorldSpeech, config 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/WorldSpeech-mn", split="train")
print(ds[0]["text"], ds[0]["audio"]["sampling_rate"])

What is in it

clips	4,228
audio	5.15 h
distinct sentences	4,218
identified speakers	0
format	parquet shards, audio inline as 24 kHz mono wav
upstream audio	24 kHz native -- the only Mongolian source with real full-band content
median clip	3.48 s
median measured bandwidth	7,664 Hz
median DNSMOS OVR	3.60
median alignment score	0.875
median ASR CER	0.120

Splits

split	clips	hours	speakers
train	3828	4.67	0
withheld	400	0.48	0

Each split is a separate file set, so load_dataset(..., split="withheld") downloads only that split. The split column carries the same value.

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	638	0.79
male	3005	3.67
unknown	585	0.69

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	0.8447552051705619
asr_transcript	string	төр засгийн догтвортой ажиллах нөхцлийг ханг
audio_path	string	wavs/ws_train_0.wav
bandwidth_hz	float64	6757.8125
cer	float64	0.0625
clip_id	string	ws_train_0
country	string	Mongolia
dnsmos_bak	float64	2.9860486290495984
dnsmos_ovr	float64	3.3701887197499913
dnsmos_p808	float64	2.4493732452786605
dnsmos_sig	float64	3.206337054570516
duration_s	float64	3.226
gender_resolved	string	female
gender_source	string	f0
human_transcript	string	төр засгийн тогтвортой ажиллах нөхцөлийг хан
language	string	mn
len_ratio	float64	1.0
mean_f0_hz	float64	193.82103338635343
original_asr_transcript	string	төр засгийн тогтвортой ажиллах нөхцөлийг хан
original_cer	float64	0.03999999910593033
original_dnsmos_bak	float64	3.1745293140411377
original_dnsmos_ovr	float64	2.4409024715423584
original_dnsmos_p808	float64	3.455521821975708
original_dnsmos_sig	float64	3.0558812618255615
original_duration	float64	3.259999990463257
original_snr	float64	16.926071166992188
pitch_confidence	float64	0.7821782178217822
segment_id	string	mn_mn_2016-07-05_0440_00020077
session_date	string	2016-07-05
snr_db	float64	0.0
source	string	parliament_mn
source_end_s	float64	204.02999877929688
source_start_s	float64	200.77000427246094
source_url	string
split	string	train
text	string	төр засгийн тогтвортой ажиллах нөхцөлийг хан

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          3,828      4.7         1      3.3       0.7
validation         0      0.0         0      0.0       0.0
test               0      0.0         0      0.0       0.0
withheld         400      0.5         1      0.4       0.1

Text diversity:
  clips                 4,228
  distinct sentences    4,218  (1.00x repetition)
  letters covered       34/35
  letter bigrams        650 distinct
  NEVER APPEARS         щ
  under 100 occurrences ф ъ

CER by source (the gate's recogniser trained on Common Voice):
  source         clips   median    mean
  ws              4228    0.120   0.115
  A markedly lower median for cv is contamination, not quality.

total: 5.2 h   male 3.7 h   female 0.8 h

Acceptance criteria:
  total >= 25 h            FAIL  (5.2 h)
  male  >= 5 h             FAIL  (3.7 h)
  male speakers >= 3       FAIL  (1)

Using it for oron-tts finetuning

build_f5_dataset.py reads a corpus directory of wav files plus a manifest. This repo ships the audio inside the parquet, not as loose wavs, so materialise it first:

from datasets import load_dataset
import soundfile as sf, json, pathlib

out = pathlib.Path("oron_mn_corpus"); (out / "wavs").mkdir(parents=True, exist_ok=True)
ds = load_dataset("btsee/WorldSpeech-mn", split="train")
with open(out / "manifest.jsonl", "w", encoding="utf-8") as fh:
    for r in ds:
        name = r["clip_id"] + ".wav"
        sf.write(out / "wavs" / name, r["audio"]["array"], r["audio"]["sampling_rate"])
        row = {k: v for k, v in r.items() if k != "audio"}
        row["audio_path"] = "wavs/" + name
        fh.write(json.dumps(row, ensure_ascii=False) + chr(10))

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

manifest.jsonl is also shipped at the repo root for inspection without downloading the audio.

Considerations

• Licence: CC-BY-NC-4.0. Broadcast and parliamentary recordings. NON-COMMERCIAL ONLY -- CC-BY-NC-4.0 forbids commercial use, and any model trained on this corpus inherits that restriction. Speakers are public figures recorded in public proceedings and broadcasts; none consented to voice cloning.
• 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_ws_mn,
  title  = {WorldSpeech Mongolian (cleaned)},
  author = {Badral, Battseren},
  year   = {2026},
  url    = {https://huggingface.co/datasets/btsee/WorldSpeech-mn}
}

Please also cite the upstream corpus: disco-eth/WorldSpeech, config mn_mn.

Not used by the model. This corpus failed the pipeline's 15% pass-rate gate, and btsee/oron-tts never trained on it — so the model carries no non-commercial restriction from this dataset.