BanglaKontho: Closing the Long-Form Gap in Bangla Text-to-Speech
cs.CL
Quick verdict
Releases a 20h single-speaker Bangla audiobook TTS corpus with careful QA; gains over IndicTTS-Bn look real but the domain-vs-scale attribution is untested.
Summary
The paper introduces BanglaKontho, a 20-hour single-speaker Bangla TTS corpus of 7,050 utterances at 24 kHz, segmented from professional audiobook recordings and released under CC BY-NC 4.0 together with a reusable Bangla text normalizer and the full preprocessing pipeline. Construction uses pyannote VAD plus ECAPA-TDNN narrator verification to remove music and non-narrator audio, pause-guided sentence segmentation (clips of 1–20 s), a Gemini 2.0 Flash-Lite first-pass transcription (2.7% WER on a 500-clip hand-transcribed subset), and 100% human verification by two native-speaker volunteers (Cohen's κ = 0.91 on the overlap subset). The normalizer covers Bangladeshi digit grouping on the lakh/koti scale, currency, dates, Danda punctuation, and Unicode composition. An MB-iSTFT-VITS baseline trained from scratch reaches 9.5% WER under IndicWav2Vec (natural-speech floor: 8%) and 4.46 naturalness MOS (15 raters, 95% CI ±0.06), versus 16.0% and 3.16 for the same architecture retrained on the 12-hour IndicTTS-Bn corpus; disabling the normalizer degrades results to 14.0% WER and 3.80 MOS. Evaluation uses a chapter-disjoint test split and reports a natural-speech MOS ceiling of 4.90. Main limitations: the domain-versus-scale explanation for the IndicTTS-Bn gap is asserted rather than experimentally isolated, training audio was band-limited (upsampled from the 16 kHz analysis copy), redistribution rights rest on the author's unverified authorization, and WER/MCD lack significance testing.
Review complete
Assessment at a glance
4
risk flags identified
● high 0
● medium 2
● low 2
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- BanglaKontho is the first released single-speaker, long-form Bangla audiobook TTS corpus with verified transcripts (20 hours, 7,050 utterances, 24 kHz, CC BY-NC 4.0).
- An MB-iSTFT-VITS trained from scratch on BanglaKontho substantially outperforms the same architecture retrained on the 12-hour IndicTTS-Bn corpus (WER 9.5% vs 16.0%; naturalness MOS 4.46 vs 3.16).
- The released Bangla text normalizer contributes a measurable, localized intelligibility gain (WER degrades from 9.5% to 14.0% when it is disabled).
- Synthesis quality approaches the natural-speech ceiling: 9.5% WER against an 8% ASR floor and 4.46 MOS against a 4.90 natural-speech ceiling.
Strengths
Genuine resource gap filled
Table 1 positions BanglaKontho against OpenSLR (~200 h, 505 speakers, ASR-style read prompts), Common Voice, Shrutilipi and IndicTTS-Bn (12 h); it is the first released single-speaker long-form Bangla audiobook corpus (20 h / 7,050 clips / 24 kHz) under an open license, exactly the regime VITS-class models need.
Above-average evaluation hygiene
Chapter-disjoint split (§4.1) guards prosodic leakage; ASR floor (8% WER on natural speech) and natural MOS ceiling (4.90) frame all synthesis numbers; Krippendorff's α = 0.74 across raters; Cohen's κ = 0.91 between verifiers; 95% CIs on every MOS.
Transparent AI-assisted pipeline with full human verification
Ablation isolates the normalizer's contribution
Honest self-reporting of weak spots
Weaknesses
Domain-vs-scale attribution is asserted, not tested
Baseline comparison has an uncontrolled text-processing confound
Rights chain rests on a bare assertion
QA is concentrated in the sole author
Point estimates without uncertainty for WER/MCD
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
BanglaKontho (arXiv 2609.29146, cs.CL) is a resource paper for Bangla text-to-speech. Its three claims are: (1) the first dedicated single-speaker, long-form Bangla audiobook TTS corpus with verified transcripts — 20 hours, 7,050 utterances at 24 kHz, released under CC BY-NC 4.0; (2) a reusable open-source Bangla text normalizer (Bangladeshi lakh/koti digit grouping, currency, dates, Danda handling, NFC Unicode, strict G2P-safe mode, §3.5); and (3) a from-scratch MB-iSTFT-VITS baseline that outperforms the same architecture retrained on the 12-hour IndicTTS-Bn corpus (WER 9.5% vs 16.0%; naturalness MOS 4.46 vs 3.16, Table 3), with a normalizer-disabled ablation at 14.0% WER / 3.80 MOS. Notably, all synthesis numbers are framed against an 8% ASR-floor WER (IndicWav2Vec on natural speech) and a 4.90 natural-speech MOS ceiling — a calibration practice most low-resource TTS papers omit.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 4 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is the BanglaKontho dataset?
A 20-hour, single-speaker Bangla text-to-speech corpus of 7,050 sentence-level clips at 24 kHz, cut from professional audiobook recordings, with human-verified transcripts. It is released openly under CC BY-NC 4.0 together with a reusable Bangla text normalizer and the full preprocessing pipeline.
Can I use BanglaKontho commercially?
No. The corpus carries CC BY-NC 4.0 plus a data-use agreement: redistribution is limited to non-commercial academic research, and commercial use, voice cloning for impersonation, fraud, and political-content synthesis are explicitly prohibited; downloaders must accept the agreement.
How good is the TTS baseline trained on BanglaKontho?
An MB-iSTFT-VITS trained from scratch reaches 9.5% WER as measured by IndicWav2Vec (natural Bangla speech scores 8% on the same ASR) and a 4.46 naturalness MOS from 15 native raters, versus a 4.90 MOS ceiling for natural speech. The same architecture retrained on the 12-hour IndicTTS-Bn corpus scores 16.0% WER and 3.16 MOS.
How was transcript quality ensured in BanglaKontho?
Google Gemini 2.0 Flash-Lite produced a first-pass transcript (2.7% WER on a 500-clip hand-transcribed subset), then two native Bangla speakers independently verified 100% of the clips, reaching Cohen's κ = 0.91 word-level agreement, with remaining disagreements resolved by discussion.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figure 1 and tables arrive as inline extracted text; visual layout and figure details were checked only against captions.
- The GitHub repository (data shards, normalizer, license documents) was not inspected; availability, the 'byte-equivalent filelists' claim, and per-stage discard statistics are taken as stated.
- No appendix appears in the extracted text; dataset composition details (source book list, per-chapter counts) are absent from the paper itself.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.