BanglaBox's Bangla voice-cloning corpus is tagged CC BY 4.0, and its own ethics statement says the consent behind it is research-only

A corpus licensed for commercial reuse, built on consent that forbids it

BanglaBox is a real engineering result. Table 3 of the paper reports 24 native Bangla raters scoring 1,824 ratings per system across three five-point scales: BanglaBox reaches Naturalness 4.56, speaker-similarity MOS 4.51 and Clarity 4.69, ahead of every synthesized baseline and within 0.20 to 0.28 points of human recordings (4.76 / 4.79 / 4.84). Every pairwise difference against the five baselines clears p<0.01 under Wilcoxon signed-rank testing with Bonferroni correction. On zero-shot voice cloning from a 10-second reference, BanglaBox reaches a speaker-encoder cosine similarity (SECS) of 0.627, against IndicF5's 0.37 and a same-corpus XTTS-v2 fine-tune's 0.47, trained on 520.3 hours of audio against the 3,850 hours reported for the prior state of the art, BnTTS.

None of that is in dispute. What the paper's tables do not resolve is who may use the corpus for what. The Hugging Face dataset repository Banglabox/bangla-corpus, confirmed live via the Hugging Face API on 13 October 2026 (799 downloads), carries the metadata tag license:cc-by-4.0 and a dataset card whose entire licence section reads: "CC BY 4.0. Identifying information is withheld for double-blind review." CC BY 4.0 is one of the most permissive Creative Commons licences: it allows commercial use, modification and redistribution, conditioned only on attribution. Section 6 of the same paper, the Ethics Statement, describes a narrower grant for the same recordings: "Those consent terms cover research use only and exclude any commercial use of the recordings."

What a team adapting Chatterbox TTS to Bangla built, and what it measured

arXiv:2610.09211v1, BanglaBox: A Phonetically-Balanced Corpus and Data-Efficient Foundation-Model Adaptation for Bangla Text-to-Speech with Zero-Shot Voice Cloning, by Emtiaz Uddin Ahmed, Araf Mahmud, Sajib Hossain, Tarikul Islam Tamiti, Sajid Fardin Dipto and Anomadarshi Barua, was submitted 6 October 2026. No institutional affiliation is listed in the paper or in its arXiv metadata; the one listed contact is a Gmail address. The paper adapts Resemble AI's MIT-licensed Chatterbox TTS (a 520M-parameter Llama-style autoregressive transformer with built-in zero-shot cloning) to Bangladeshi Bangla through three changes: a merge-consistent tokenizer extension from 704 to 4,240 tokens, Bangla text normalization with honorific expansion, and a prompt-masked dual-loss training objective.

The corpus itself has two tiers: Tier A is 367.3 hours from 13 speakers recorded in a Dhaka studio (48kHz/24-bit, SNR ≥44dB) under a 14-week coverage-feedback loop that selects scripts to minimize a tiered Jensen-Shannon divergence over phone, diphone, triphone and juktakkhor (conjunct-cluster) distributions. Tier B is 153.0 hours curated from seven public Bangla speech datasets, after dropping 16 of 39 initial speakers (94 of 247 hours) that failed an SNR or ablation-driven quality filter. The corpus, model weights, LoRA adapter and complete evaluation materials are hosted at huggingface.co/Banglabox.

What the licence says against what the consent says

Rows checked live by BLOMEGA on 13 October 2026 via the Hugging Face API (huggingface.co/api/datasets, /api/models) and a direct fetch of the dataset README. Paper quotations are verbatim from arXiv:2610.09211v1.

The footnote 1 distinction matters: "unconditional" in the abstract modifies when the artifacts appear (not gated behind acceptance), not what someone may do with them. The Responsible NLP Checklist embedded in the paper (Appendix, item 2.b) separately confirms that "Tier A recordings use professional voice talent who provided written informed consent" for research use, and that Tier B audio is used "only when their licences permit downstream research use" (our emphasis). Nowhere in the paper's own text does a speaker's consent extend to commercial use, yet the artifact a developer actually downloads carries a licence tag that grants it.

How a research-only consent form becomes a commercially-tagged download

The mismatch is not a hypothetical about Creative Commons law; it is a specific, checkable gap between two documents produced by the same team about the same 367.3 hours of studio audio.

This pattern is not unique to BanglaBox. Research teams routinely tag an artifact CC BY or CC BY-NC without the dataset card's licence field being reviewed against the actual consent language buried in the paper's ethics section, because the two documents are written by different co-authors at different times and rarely cross-checked before submission. For a data buyer, the Hugging Face page is usually the only document read.

What this means for a team evaluating Bangla or other low-resource voice data

• Read the ethics section, not just the licence badge. A dataset card's licence field is set by whoever uploads it and is not legally binding on the underlying rights if the source consent was narrower. BanglaBox is a documented, checkable case of exactly that gap.
• The watermark is not a licence. Resemble AI's Perth watermark, inherited by BanglaBox's fine-tune and applied to every generated waveform, lets a third party detect that audio came from this system. It does nothing to resolve whether using the training corpus commercially is permitted.
• Tier B inherits a second, unverified licensing layer. 153.0 hours are drawn from seven public Hugging Face/Bengali.AI sources "retained only when their original licenses permit" (Appendix C.7); we did not independently re-verify each of those seven upstream licences, so a commercial user inherits that same unchecked chain twice over.
• Data efficiency is real and worth separating from the licence question. BanglaBox's 520.3-hour recipe reaches SECS 0.627 against IndicF5's 0.37 and a same-corpus XTTS-v2 fine-tune's 0.47, and Table 6's ablation shows the prompt-masked dual-loss objective is the single biggest driver of cloning fidelity (removing it drops SECS to 0.51). The method is reusable even where the specific corpus is not.
• Consent scope should travel with the data, not just the licence string. This is the gap BLOMEGA is built to close: a chain-of-title record naming exactly what each contributor agreed to, attached to the data itself rather than inferred from a one-line badge.

How to check the licence-versus-consent gap yourself

# Dataset card and licence tag, live
curl -s "https://huggingface.co/api/datasets?author=Banglabox" | python3 -m json.tool
curl -s "https://huggingface.co/datasets/Banglabox/bangla-corpus/raw/main/README.md"
# -> license: cc-by-4.0 (YAML front matter) and "## Licence / CC BY 4.0" in the card body

# Model weights licence
curl -s "https://huggingface.co/api/models/Banglabox/chatterbox-bangla-tts" | python3 -c "import json,sys;print(json.load(sys.stdin)['cardData'])"
# -> {'language': ['bn'], 'license': 'mit', ...}

# The paper's own consent language
curl -sL https://arxiv.org/pdf/2610.09211 -o banglabox.pdf
python3 -c "import pypdf; print('\n'.join(p.extract_text() for p in pypdf.PdfReader('banglabox.pdf').pages))" | grep -A3 "Research-only use terms"
# -> "Research-only use terms. Those consent terms cover research use only and exclude any commercial use of the recordings."

The subjective and objective score tables (Tables 3-5) can be recomputed from the public evaluation dataset Banglabox/banglaeval, which the dataset card describes as holding "every evaluation input, output and score behind the paper's tables."

What would prove this reading wrong

Our claim is narrow: as published and as hosted on 13 October 2026, the corpus's stated licence (CC BY 4.0) is broader than the consent scope the same paper describes for the people recorded in it (research use only, no commercial use). We would be wrong if the authors have a separate, broader commercial-use agreement with Tier A speakers that simply is not described in the Ethics Statement, or if "CC BY 4.0" on the dataset card is a placeholder pending a correction that had already been made by the time a reader checks. We predict that by 31 January 2027 one of two things will have happened: either the dataset card's licence is changed to a non-commercial term (e.g. CC BY-NC) to match the stated consent, or the paper's camera-ready version (this is a double-blind submission as of October 2026) adds language explicitly reconciling the two. If neither has happened and the CC BY 4.0 tag still stands unchanged next to the same Ethics Statement text, the gap should be treated as an unresolved error in the artifact, not a one-time oversight caught here.

Sources

1. Emtiaz Uddin Ahmed, Araf Mahmud, Sajib Hossain, Tarikul Islam Tamiti, Sajid Fardin Dipto, Anomadarshi Barua. BanglaBox: A Phonetically-Balanced Corpus and Data-Efficient Foundation-Model Adaptation for Bangla Text-to-Speech with Zero-Shot Voice Cloning. arXiv:2610.09211v1, submitted 6 October 2026. Abstract, §1-6, Tables 1-7, Ethics Statement, Responsible NLP Checklist.
2. huggingface.co/datasets/Banglabox/bangla-corpus, licence tag cc-by-4.0, README checked via direct fetch and the Hugging Face API, 13 October 2026.
3. huggingface.co/Banglabox/chatterbox-bangla-tts, licence tag mit, Hugging Face API, 13 October 2026.
4. huggingface.co/datasets/Banglabox/banglaeval, evaluation materials dataset, Hugging Face API, 13 October 2026.
5. Resemble AI. Chatterbox TTS, MIT licence, cited in arXiv:2610.09211v1 §3.
6. BLOMEGA computations and live checks, 13 October 2026: Hugging Face API queries, dataset card fetch, and paper-text extraction cross-check (pypdf).

FAQ

Is BanglaBox's Bangla voice corpus free to use commercially?

The Hugging Face dataset card for Banglabox/bangla-corpus says "Licence: CC BY 4.0", a licence that by itself permits commercial reuse with attribution. But the paper's own Ethics Statement says every Tier A speaker's written informed consent "cover[s] research use only and exclude[s] any commercial use of the recordings." A licence grant and the underlying consent it was built on disagree, confirmed live on 13 October 2026.

How good is BanglaBox compared with commercial Bangla TTS?

In the paper's own 24-rater evaluation (1,824 ratings per system), BanglaBox scores highest on Naturalness (4.56), SMOS (4.51) and Clarity (4.69) against Google TTS, Azure TTS, Narakeet, IndicF5 and fine-tuned YourTTS/XTTS-v2, with every pairwise difference significant at p<0.01. It reaches speaker-cloning SECS of 0.627 against IndicF5's 0.37, trained on 520.3 hours versus a prior system's reported 3,850.

Can I reproduce BanglaBox's numbers myself?

Yes. The paper (arXiv:2610.09211) ships a public Hugging Face evaluation dataset, Banglabox/banglaeval, described as holding every evaluation input, output and per-rating score behind its tables, plus the model weights and inference code in Banglabox/chatterbox-bangla-tts (MIT licence).