voidful/tw-speech

TW Speech

Taiwan speech annotations generated with OpenMOSS-Team/MOSS-Audio-8B-Thinking in BF16 without quantization, using voidful/tw-ice-tts-moss-annotation.

Access requires a request and manual approval by the repository owner. The remote release is complete only when publication.json reports remote_upload_complete: true; files may be arriving before that marker exists. Source-specific rights and allowed-use metadata are preserved in canonical_record_json. Some upstream rows carry license_pending_review / private_training_pending_rights_review. This repository does not grant a new open license for the source recordings.

Configuration	Split	Audio windows
tw_tts	train	14,049,932
tw_tts	validation	14,165
earica_speech	train	1,009,291

All 13,687 pinned source files are covered. Every provided audio item, channel and complete nonoverlapping 20-second window is retained, including short final windows. Official source splits are unchanged. The 25 original Earica audio=[] rows have no audio to annotate and are preserved separately in source-issues/earica_speech/train.jsonl.

Parquet shards embed the exact original-rate PCM16 WAV bytes. Their SHA256 is stored in audio_sha256. audio can be loaded directly with Hugging Face Datasets; no external audio paths are required. labels contains the eleven structured annotation fields. canonical_record_json preserves the complete canonical record, source metadata, raw response and model provenance; its audio path is a historical local locator. Use the embedded audio column.

Transcript calibration (Breeze-ASR-25 × Qwen3-ASR cross-validation)

The tw_tts_calibration and earica_speech_calibration configurations add a calibrated transcript for every audio window (15,073,388 rows; no audio — join on record_id with the audio configurations). The original MOSS transcript in labels.transcript is unchanged and is repeated here as transcript_moss.

Both ASR systems transcribed every window independently, without any transcript prompt:

System	Revision	Decoding
MediaTek-Research/Breeze-ASR-25	cffe7ccb404d025296a00758d0a33468bec3a9d0	vLLM 0.32, FP16, greedy, forced `<
Qwen/Qwen3-ASR-1.7B	7278e1e70fe206f11671096ffdd38061171dd6e5	vLLM 0.32, BF16, greedy, automatic language

Normalization (tw-speech-asr-normalize-v3, comparison only): per character NFKC → OpenCC t2s → casefold; punctuation, spaces and symbols removed; Chinese digit characters (零〇一…九兩) mapped to 0–9 so 1971 equals 一九七一 (units such as 十百千萬 are kept, so 三百 is never merged with 300); Latin word boundaries ignored.

Correction rule (two-ASR character vote, MOSS text as the base): a span of the MOSS transcript is changed only when Breeze-ASR-25 and Qwen3-ASR independently propose exactly the same replacement for exactly the same span, both neighbouring characters are matched by both systems, the original source_text does not support the MOSS reading at that position, and the edit is not (a) adding/removing fillers or immediate repetitions (Breeze removes disfluencies by design), (b) a number-format change, (c) a Latin/Chinese script switch, or (d) a 他/她/它 or 的/得/地 choice. Replacement characters are taken from Breeze-ASR-25 (Traditional script). MOSS 00:00 Speaker A: timestamp/speaker tags are removed (moss_tags_stripped). Every applied edit is listed in edits (JSON; moss_span indexes the tag-stripped MOSS text when moss_tags_stripped is true). Text edits are applied only to rows whose MOSS languages include 臺灣華語 (Breeze-ASR-25 is decoded as zh).

Status (calibration_status, train_use): after correction, cer_qwen and cer_breeze (Breeze side with fillers/repetitions removed on both texts) measure agreement. verified: both ≤ 0.10; plausible: min ≤ 0.15 and max ≤ 0.30; low_agreement: otherwise; language_mismatch: labelled Taiwan Mandarin but Qwen3-ASR detects another language (e.g. Cantonese); not_mandarin_label: not labelled Taiwan Mandarin (not calibrated); no_transcript. train_use is false for low_agreement, language_mismatch and no_transcript. ASR outputs are machine-generated evidence, not human references.

from datasets import load_dataset
cal = load_dataset("voidful/tw-speech", "tw_tts_calibration", split="train", streaming=True, token=True)
print(next(iter(cal)))

See calibration-provenance.json for code checksums, counts and the exact policy.

Status	Rows
verified	9,295,979
plausible	3,574,637
low_agreement	1,752,982
not_mandarin_label	392,867
language_mismatch	51,410
no_transcript	5,513

3,827,344 rows (25.4%) have a changed transcript (4,908,419 edits; 106,641 rows had MOSS speaker/timestamp tags removed); train_use is true for 13,263,483 rows (43,485 h). On Taiwan-Mandarin rows the mean CER of the label against Qwen3-ASR drops from 9.14% to 8.31% and against Breeze-ASR-25 from 10.43% to 9.57%. Independent check: in 40 random shards, 87.8% of edited rows moved closer to the original source_text (2.2% moved away); source_text was only used to veto edits, never to propose them.

from datasets import load_dataset, Audio
ds = load_dataset("voidful/tw-speech", "tw_tts", split="train", streaming=True, token=True)
ds = ds.cast_column("audio", Audio(decode=False))
example = next(iter(ds))
print(example["labels"])

Annotations are model-generated and have not been manually reviewed. An unavailable transcript remains null with its reason in the canonical record. Original text is preserved separately and was not used as an ASR prompt. Official train/validation source groups can overlap; annotation completion does not certify a leakage-free training split. See provenance.json. manifest.json lists every shard, its row count and exact file checksum.