sejongwang/ipapack_plus_clean

IPAPACK++ Cleaned-Label Overlay (v1)

A labels-only, strictly additive cleaned overlay on IPAPACK++ (Zhu et al., ZIPA, ACL 2025). It adds a regenerated phoneme label (phones_v1_clean) alongside the paper's original phones, a drop_flag for the rows we could not repair, a per-utterance sigs list of which defect (if any) was fixed, and a per-row license tag. The original paper label is preserved on every row, including the dropped ones — so you can compare, ablate, or refuse the cleanup entirely.

This is not a re-publication of the audio and not a replacement for the paper. It is an additive second pass on the labels, built on and crediting IPAPACK++/ZIPA, offered back to the community that uses it. The IPAPACK++ authors were themselves explicit that the G2P-generated transcriptions are noisy, especially for low-resource languages, and named label quality among the paper's limitations; this overlay is extra coverage on top of theirs.

Status: v1, labels-only. No retraining, no audio–IPA forced alignment, no phonetician sign-off yet — see Limitations. Every check verifies a defect is absent from the regenerated label, not that the label matches the audio.

---

TL;DR

• What this is. A cleaned-label overlay over the full IPAPACK++ corpus: 8,290,179 cuts / 8,281,384 active / 17,109.4 h across 415 shards, as Lhotse JSONL cut manifests. We add phones_v1_clean, drop_flag, sigs, and license. Audio is not bundled — re-pair it from the public anyspeech/ipapack_plus_* shards by cut.id.
• What we fixed. Nine mechanically-identifiable label-defect classes ("signatures"): six fixed directly in the labels, one dropped because it is audio-dependent, two deferred to an eval-time canonicalize layer.
• How much changed. ~1.9 M rows regenerated; 6,383,316 rows (77 %) ship unchanged at the paper baseline. drop_flag covers 8,795 utterances (0.11 %).
• License. Mixed, per-source (recorded per row in the license field) — most rows are CC0/CC-BY, but four crowd-sourced OpenSLR slices (~778 k rows) are CC-BY-SA. See Source corpora & licenses.
• How to use it. Train on cut["supervisions"][0]["custom"]["phones_v1_clean"]; filter out …["custom"]["drop_flag"]; re-pair audio by cut.id. The paper baseline is preserved on every row.

---

How to use it

This is a label overlay: you get cleaned label manifests; you bring your own audio.

from huggingface_hub import snapshot_download
from lhotse import CutSet

# 1) pull the cleaned cut manifests (labels only, no audio)
local = snapshot_download("sejongwang/ipapack_plus_clean", repo_type="dataset",
                          allow_patterns="cuts.*.jsonl.gz")

NO_ROUTE = {"mn", "ia", "ceb", "jv", "ff", "tg", "skr"}  # ship at ipa_v0; filter by ISO

cuts = CutSet.from_jsonl(f"{local}/cuts.000000.jsonl.gz")   # one shard; loop over all 415
for cut in cuts:
    sup    = cut.supervisions[0]
    custom = sup.custom
    if custom.get("drop_flag", False):
        continue                                  # skip Sig-3 / untonable / residue
    if sup.language in NO_ROUTE:
        continue                                  # carries v3_backend="unchanged"
    label   = custom["phones_v1_clean"]           # train on this
    baseline = custom.get("original") or custom.get("phones")   # paper baseline (heterogeneous)
    row_license = cut.custom["license"]           # per-row source license (e.g. cc-by-sa-4.0)
    # re-pair audio: member {cut.id}.flac inside the original recording.NNNNNN.tar (same shard index)

Fields (all on the supervision's custom block unless noted):

• phones_v1_clean — the canonical v1 cleaned label; train on this.
• drop_flag — True for Sig-3 (digit), Sig-4-untonable, and residue drops. Filter these out.
• original (a.k.a. ipa_v0) — paper baseline; heterogeneous: absent on ~1.445 M MLS rows, which keep phones. Read as custom.get("original") or custom.get("phones").
• sigs — list of signatures fixed on this row.
• cut.custom.corpus — source provenance tag (cut level), e.g. cv:rw, mls:german, openslr:bengali.
• cut.custom.license — per-row source license (cut level). Filter on this if you must avoid ShareAlike rows.

Two warnings worth tattooing on your dataloader. supervisions[0]["text"] is a stale phone string, not orthography. PII (gender, speaker, age, accents, variant) has been stripped on every row.

Re-pairing audio. Join by cut.id. Each utterance's audio is the member {cut.id}.flac inside the original recording.NNNNNN.tar at the same shard index as the cut shard (positionally paired). The original audio is the public anyspeech/ipapack_plus_* dataset (16 kHz). (The lhotse from_shar re-pairing path was not executed end-to-end here — sanity-check it against your lhotse version.)

---

What the audit found: the nine signatures

A signature is a single, mechanically-identifiable defect class in IPAPACK++'s phoneme labels. ipa_v0 is the paper-baseline label; ipa_v3 (= phones_v1_clean) is the cleaned label. The audit, the detector, and the measurements are our work; the dataset, its Table-6 hour accounting, and the G2P tools (CharsiuG2P, Epitran) are the paper's.

#	Signature	What it is	Policy	Scope
1	Apostrophe letter-name (U+0027)	Orthography split on the apostrophe; orphaned 's rendered as the letter-name /ɛs/ ("ess") instead of the genitive sibilant	CLEAN	849,482 utts / ~916 h + +109.9 h additional
2	Typographic apostrophe (U+2019)	Same split path on the typographic '; folded with Sig-1	CLEAN	(folded with Sig-1)
3	Digit silent drop	ASCII-only \d never matches native-script digits (Bengali ১৯৪৭, Devanagari, Burmese, Tamil, Arabic-Indic); silently discarded	DROP	537 active drops (~14.2 h)
4	Non-Mandarin tone strip	byT5 emits Chao tones but a post-G2P step strips them; 7 of 8 tonal cells 100 % stripped, Yoruba partial	CLEAN	8 cells, 6 langs / 79.6 h
5	Dutch/Swedish Epitran rule artifact	Epitran's nld/swe-Latn rule table is phonologically wrong on 9 patterns, deterministically	CLEAN	35,164 utts (CV+FLEURS)
6	Length-marker ː over-insertion	Epitran's word-final length rule over-fires; affects nl, sv	CLEAN	shared with Sig-5
C1	Mandarin Chao-tone zero	byT5 under <cmn-s>: emits zero Chao tone letters across 6,955,717 chars; routed around with pypinyin	CLEAN	122,220 utts / ~199 h
7	French ø/œ notation drift	Train vs eval write different but both PHOIBLE-attested vowels — a notation drift, not an error	CANONICALIZE (eval-time; no data change)	~10 h, eval-only
8	Spanish r/R notation drift	Tap ɾ / trill r are a genuine phonemic contrast (pero/perro); a blanket merge would over-collapse it	CANONICALIZE (eval-time; no data change)	~10 h, eval-only

So: six fixed in the data (1, 2, 4, 5, 6, C1), one dropped (3), two eval-time only (7, 8). The largest single re-route is Kinyarwanda (rw, 977,882 utts), which has no citation-form gold and is validated only by inter-tool agreement — weight it accordingly. Per-utterance regeneration uses a measurement-driven per-ISO matrix (BEST_G2P_PER_LANG) drawn from thirteen G2P backends (ten exercised in v1).

The full forensic detail — every measurement, the false-positive byte-match audit (~1,200 h of MLS/LibriSpeech cells over-flagged and excluded before the signatures were finalized), the additional-damage sweep, and the zero-tone proof — is available on request.

---

Limitations

1. No retraining, ablation, or forced alignment at v1 — this is the big one. Every PASS verifies self-consistency with the generating backend (the defect is absent), not label-to-audio correctness. PERs are cited against citation-form lexicon gold, not IPAPACK++ audio. Tone-restored zh/yue/th apply no sandhi (你好 ships as nǐ hǎo, unsandhied) — treat as "tone present, value unverified." Because there is no reliable token-level alignment in seq2seq G2P, each fix re-transcribes the whole utterance, so non-target tokens adopt that backend's conventions (reduced forms, stress/length marks) — likewise not audio-validated. A full audit of all 1.9 M regenerated rows confirms this re-transcription introduces a small rate of new errors — ~38 malformed labels, ~1,100 illegal genitive clusters (en/ca), ~1,000 gold-regressions in two re-routed cells — but the fixes outweigh them by **100:1** for the main signatures. The exception is the Malay (ms) re-route, which is net-negative (the routed backend emitted English-style pronunciations) — fall back to v0 there.
2. Most of the package is unchanged paper baseline. ~1.9 M rows touched; 6,383,316 (77 %) ship unchanged, including a parse-cell blind spot of ~2.49 M cuts (MLS, OpenSLR, hyphenated zh-CN/sv-SE cells) never routed through the detectors. Treat any v3_backend="unchanged" row as "paper-baseline quality, not audited by this release."
3. The 7 no_route ISO codes are unmeasured. mn, ia, ceb, jv, ff, tg, skr (~19 k utts) sit at ipa_v0 with drop_flag=False.
4. The shipped vocab predates the cleanup. Encoding phones_v1_clean with ipa_simplified/unigram_127.model sends ≈1.36 M utts to <unk> (≈680 k after ɡ→g). These are not label defects — normalize ɡ→g or extend the vocab before training.
5. Sig-7/8 are eval-time only (apply ø→œ for fr, r→ɾ for es to both hyp and ref before scoring; the es merge hides genuine /ɾ/–/r/ contrasts). Also ~800,764 cuts share a duplicate phones_v1_clean — dedup/group-by-label when splitting to avoid train/test leakage.

---

Source corpora & licenses

This dataset is an additive, labels-only overlay. It redistributes regenerated IPA phoneme labels (phones_v1_clean), orthographic transcript text, utterance IDs, and a Lhotse cut manifest — NO AUDIO. Re-pair audio yourself from the sources below. Original IPAPACK++ phones are preserved on every row.

This is a mixed-license release. The applicable license is recorded per row in the license field. Rows from ShareAlike sources are offered under CC-BY-SA and may not be treated as permissively licensed.

Primary citation: Zhu, J., Samir, F., Chodroff, E., Mortensen, D. R. ZIPA: A Family of Efficient Models for Multilingual Phone Recognition. ACL 2025. https://aclanthology.org/2025.acl-long.961/ · arXiv:2505.23170

Source	Rows	License	Attribution / citation
Common Voice (cv:*)	5,285,019	CC0-1.0	Mozilla Common Voice — https://commonvoice.mozilla.org/
Multilingual LibriSpeech (mls:*, SLR94)	1,445,339	CC-BY-4.0	Pratap et al. (2020) — https://www.openslr.org/94/ · labels regenerated
FLEURS (fleurs:*)	155,927	CC-BY-4.0	Conneau et al. (2022), Google — https://huggingface.co/datasets/google/fleurs · labels regenerated
LibriSpeech (openslr:librispeech/*, SLR12)	281,209	CC-BY-4.0	Panayotov et al. (2015) — https://www.openslr.org/12/ · labels regenerated
AISHELL-1 (aishell, SLR33)	120,078	Apache-2.0	Bu et al. (2017), arXiv:1709.05522 — https://www.openslr.org/33/
Kazakh KSC (openslr:kazakh, SLR102)	147,165	CC-BY-4.0	Khassanov et al. (EACL 2021) — https://www.openslr.org/102/ · labels regenerated
IISc-MILE Tamil (openslr:tamil, SLR127)	77,136	CC-BY-2.0	Madhavaraj et al. (2022), arXiv:2207.13331 — https://www.openslr.org/127/ · labels regenerated
Bengali ASR (openslr:bengali, SLR53)	218,377	CC-BY-SA-4.0	© 2016–2018 Google, Inc.; Kjartansson et al. (SLTU 2018) — https://www.openslr.org/53/ · ShareAlike
Javanese ASR (openslr:javanese, SLR35)	184,984	CC-BY-SA-4.0	© 2016–2017 Google, Inc. (w/ Reykjavik Univ., Univ. Gadjah Mada) — https://www.openslr.org/35/ · ShareAlike
Sinhala ASR (openslr:shinhala, SLR52)	178,001	CC-BY-SA-4.0	© 2016–2018 Google, Inc.; Kjartansson et al. (SLTU 2018) — https://www.openslr.org/52/ · ShareAlike
Kazakh KSD (openslr:kazakh2/*, SLR140)	196,944	CC-BY-SA-4.0	Mansurova & Kadyrbek (2023), Al-Farabi Kazakh National Univ. — https://www.openslr.org/140/ · source CC-BY-SA-3.0, adapted labels offered under CC-BY-SA-4.0 (permitted ShareAlike upgrade)

Excluded: Magicdata — non-redistributable per IPAPACK++; confirmed absent from this release.

ShareAlike notice. The Bengali, Javanese, Sinhala, and Kazakh-KSD slices are offered under CC-BY-SA-4.0 — their regenerated IPA labels are Adapted Material. (Bengali/Javanese/Sinhala sources are CC-BY-SA-4.0; the Kazakh-KSD source is CC-BY-SA-3.0, upgraded to 4.0 under the ShareAlike "this version or later" clause.) Changes were made (phoneme labels regenerated via grapheme-to-phoneme).

G2P toolchain (credit, not a license obligation on the labels). Labels were generated with OLaPh (Wirth, 2025; en/de/fr/cs), Epitran, phonemizer + espeak-ng (nl/sv), CharsiuG2P, pypinyin + pinyin-to-ipa, ToJyutping, viphoneme, PyThaiNLP, indic_nlp_library, and commonvoice-utils. espeak-ng is GPL-3.0 and commonvoice-utils is AGPL-3.0, used in-process during generation; per the FSF GPL FAQ, program output is not covered by the program's copyright, so no GPL/AGPL obligation attaches to these IPA label strings.

---

Citation

Please cite both this overlay and the original IPAPACK++ paper.

@misc{kim2026ipapack_cleanup_v1,
  title  = {Cleaning IPAPACK++: A Surgical Audit of Multilingual Phoneme Labels},
  author = {Kim, Junehwi},
  year   = {2026},
  note   = {IPAPACK++ Cleaned-Label Overlay (v1), Hugging Face Datasets},
}

Zhu, Jian and Samir, Farhan and Chodroff, Eleanor and Mortensen, David R. ZIPA: A Family of Efficient Models for Multilingual Phone Recognition. Proc. 63rd ACL 2025 (Vol. 1: Long Papers). https://aclanthology.org/2025.acl-long.961/

The label-cleanup work and this release are by Junehwi Kim; compute was a local 2080 Ti × 3.