th cleaned speech corpus

This refresh provides a converged full train manifest, a frozen eval manifest, and a token-coverage subset of approximately 5000 hours. All three manifests reuse the existing validated FLAC-in-Parquet audio pool; no audio was copied, repacked, or transcoded for this refresh.

Corpus overview

Variant	Rows	Hours
train	11,756,895	13412.724
coverage_5000h	4,338,927	4999.983
eval	1,772	2.001

All audio locators are portable and have the form audio/<subset>/<group>/part-N.parquet#row=N.

The release audit records exact input/output hashes, rejected audio mappings, language-tag validation, and zero train/eval audio-identity overlap.

from datasets import load_dataset
ds = load_dataset("WTForbes/th", "th-5000h", split="train")

Update 2026-09-10 — incremental additions and contamination removal

This revision does not rewrite any existing audio part. It (1) appends ten new subsets, (2) removes 3,903 rows from corpus/th.train.clean.jsonl whose audio or text collides with public ASR benchmarks, and (3) adds provenance columns. Everything else is byte-identical to the previous revision.

New subsets (audio/<subset>/<hash>/part-*.parquet, rows appended to corpus/th.train.clean.jsonl)

subset	source	licence	rows	hours	text cleaning
central	Porjai Thai Dialect Corpus, central split (HF)	CC BY-SA	335,674	697.7	word-spaced transcripts joined; R1–R10 normaliser
korat / khummuang / pattani	Porjai Thai Dialect Corpus, dialect splits	CC BY-SA	40,704 / 24,543 / 39,361	48.2 / 33.9 / 41.2	same; dialect column set; khummuang/pattani dev were used for training, do not evaluate on them
isan	Typhoon Isan dataset	CC BY-SA	9,971	27.7	official thai_spelling column is unreliable → Thai spelling rebuilt from the [isan]<thai> raw pairs; common loanwords → Thai script (LLM-audited)
lotusdis	LOTUSDIS far-field meeting corpus	research	133,320	100.9	loanwords → Thai script (94 % audit pass); mic/distance kept in mic
thai_ser	THAI-SER (scripted turns only)	CC BY-SA	21,562	31.4	improvised turns have no transcript and are excluded
elderly	Thai Elderly Speech (GitHub, 3 parts)	CC BY-SA	19,200	26.6	—
gowajee	Gowajee corpus	research	18,796	15.9	phrase spaces restored from the turbo model's segmentation
yodas_clean	YODAS2 th000 (YouTube subtitles)	CC BY 3.0	25,513	155.9	windows ≤ 28 s; spam-subtitle filter; kept only when large-v3 phoneme error ≤ 0.10 and a Qwen3.8 judge accepted the subtitle (typos minimally fixed, hallucinations dropped)

All new rows carry: text (normalised: no punctuation, ๆ expanded, digits spelled out, fillers standardised), text_raw (the source transcript untouched), text_v1 (before the semantic loanword pass, where it differs), quality_version = th_add_2026-09-10, id_orig, speaker, split, dialect, flags, plus the standard audio_id / audio_parquet / audio_row locators. Audio is 16 kHz mono FLAC, PCM_16, verified by sha256 and pcm_sha256 exactly like the existing parts.

Rows whose raw transcript contained Arabic digits are kept (they were excluded from our training because the spoken form is ambiguous, but that is a training choice, not a data defect).

Rows removed from the train manifest (3,903)

subset	removed	reason
gigaspeech2	3,874	audio or exact/near-exact text matches GigaSpeech2 test / dev or the 1,000-utterance typhoon-ai/gigaspeech2-typhoon benchmark
common_voice16	19	text matches TVSpeech / GS2 test sentences
aadonis	10	text matches GS2 dev sentences

The full list (audio_parquet, audio_row, subset, match type, benchmark id) is in /workspace/report/wtf_th_delete_rows.json (mirrored as corpus/removed.contamination.2026-09-10.jsonl). The audio parts themselves are untouched, so the rows can be restored by re-adding them to the manifest.

Not obtained (so not in this revision)

• GigaSpeech2 (gated part): the official HF release is gated; only the mirror shards already inside wtf_gigaspeech2 are used. No further GS2 audio was downloaded.
• Common Voice 26 (th): requires a Mozilla Data Collective login; HF mirrors stop at CV 11/13, so nothing beyond the existing common_voice16 subset was added.
• There is no other Thai source waiting for download; the fine-tuning line is paused (analysis phase), so the next data change will be the worldspeech ำ/า repair, not new collection.

Known issues left as-is

• worldspeech: 19.4 % of rows write ำ as า (a PDF-extraction defect in the source). Not repaired in this revision; a substitution-only LLM repair is piloted but not yet validated.
• gigaspeech2 labels are the EThai refined labels (they already were); no change.