PaperlayerView exact revision on arXiv

Source-linked AI summary

BanglaBox: A Phonetically-Balanced Corpus and Data-Efficient Foundation-Model Adaptation for Bangla Text-to-Speech with Zero-Shot Voice Cloning

Emtiaz Uddin Ahmed, Araf Mahmud, Sajib Hossain, Tarikul Islam Tamiti, Sajid Fardin Dipto, Anomadarshi Barua

arXiv:2610.09211v1cs.SD

TL;DR

Bangla TTS has largely relied on small, single-speaker resources, with no open zero-shot cloning system identified for the Bangladeshi register. BanglaBox adapts an English-pretrained autoregressive model using a balanced corpus and targeted fine-tuning changes. It outperforms five baselines on perceptual metrics and reaches similarity comparable to prior few-shot results with substantially less Bangla audio.

  • Problem

    Existing Bangla TTS systems are tied to small single-speaker corpora, and the paper identifies no open zero-shot cloning system for the Bangladeshi register.

    Source: §1 · p. 1

  • Method

    BanglaBox adapts an English-pretrained autoregressive TTS model using a juktakkhor-aware corpus and tokenizer, text-normalization, and prompt-masked dual-loss changes.

    Source: §1 · p. 2, §3.1 · p. 5

  • Results

    BanglaBox leads all 5 baselines on every perceptual metric (p<0.01) and reaches speaker similarity comparable to prior few-shot results using 520.3 h versus 3.85k h of Bangla audio.

    Source: §1 · p. 2

  • Takeaways & Limitations

    The recipe is also evaluated on a seven-category stress set, public BnTTS benchmarks, and naturally occurring Bangla beyond the internal test split.

    Source: §6 · p. 8

    All Tier A scripts are LLM-generated, so possible translationese or LLM-register phrasing remains a concern; direct text-level register analysis is future work.

    Source: Limitations · p. 9

Abstract

from arXiv · show

We present a recipe for adapting English-pretrained autoregressive TTS foundation models to underrepresented languages, demonstrated on Bangladeshi Bangla. Existing Bangla TTS corpora are small and single-speaker, and to our knowledge no open zero-shot voice-cloning system is available for the Bangladeshi register. We contribute a phonetically- and gender-balanced two-tier Bangladeshi Bangla corpus balanced via a tiered Jensen-Shannon divergence objective over conjunct clusters (juktakkhor), together with three fine-tuning changes: a merge-consistent tokenizer extension, Bangla text normalization, and a prompt-masked dual-loss objective. These changes preserve zero-shot cloning across the language switch. Our BanglaEval protocol applies Wilcoxon signed-rank tests with Bonferroni correction over native-speaker ratings. BanglaBox attains near-natural Naturalness, outperforms commercial and open-source baselines on Naturalness, Speaker Similarity, and Clarity, and reaches speaker similarity comparable to prior few-shot results while using approximately 7x less Bangla fine-tuning audio. We further validate the recipe beyond our own test split using a seven-category stress set of difficult real-world text, the public BnTTS evaluation benchmarks, and naturally occurring Bangla that no language model wrote. All artifacts, including the corpus, weights, code, and complete evaluation materials, are released publicly and unconditionally.

1 Introduction

BanglaBox addresses limited Bangla TTS resources and the absence of an open zero-shot cloning system for the Bangladeshi register. It adapts an English-pretrained model and reports stronger perceptual results than five baselines.

Motivation

  • Existing Bangla TTS systems rely on small single-speaker corpora, and the largest prior balanced resource is also single-speaker; no prior work adapts Chatterbox to Bangla.The paper also notes that prior balancing designs do not model the conjunct-cluster (juktakkhor) inventory.Source: §1 · pp. 1–2

Results

  • BanglaBox leads all 5 baselines on every perceptual metric (p<0.01), reaches speaker similarity comparable to prior few-shot results, and uses 520.3 h versus 3.85k h of Bangla audio.The few-shot comparison is indicative because the encoders and model families differ; BanglaBox is within 0.20 MOS of the 4.76 human ceiling.Source: §1 · p. 2

Contributions

  • BanglaBox contributes a 520.3 h, 36-speaker, 7-domain corpus balanced with a juktakkhor-aware tiered JSD objective and three adaptation changes.The changes are a merge-consistent tokenizer extension, Bangla text normalization, and a prompt-masked dual-loss objective.Source: §1 · p. 2
  • BanglaEval uses 24 native raters and Wilcoxon signed-rank tests with Bonferroni correction, alongside a chunked long-form inference pipeline.The paper describes this as stronger statistical grounding than prior Bangla TTS evaluations.Source: §1 · p. 2

Validation

  • The validation suite extends beyond the internal split to a seven-category stress test, public BnTTS benchmarks, natural Bangla, and matched-hours and same-corpus comparisons.The corpus, weights, code, and complete evaluation materials are released publicly and unconditionally.Source: §1 · pp. 1–2

2 Dataset construction

The corpus combines studio-recorded and curated Bangla speech, with script selection guided by a tiered JSD objective over phonetic units through juktakkhor. Its statistics, split design, and recording checks document the resource’s scope and quality controls.

2.1 Two-tier corpus design

  • The corpus contains 520.3 h from 36 speakers across 2 tiers, with both tiers targeting a 7-domain distribution and gender balance.Tier A contributes 367.3 h from 13 studio-recorded speakers; Tier B contributes 153.0 h from 23 speakers curated from public datasets, after SNR and downstream-quality filtering.Source: §2.1 · p. 2

2.2 Tiered JSD balance objective

  • Tier A script selection minimizes weighted JSD over phone, diphone, triphone, and juktakkhor histograms, with a natural-plus-uniform target to retain coverage pressure on rare units.The selected scripts achieve ≥99% phone, ≥93% diphone, ≥80% valid triphone, and ≥85% juktakkhor coverage; top-150 juktakkhor coverage is ≥97%.Source: §2.2 · pp. 2–3

2.3 Statistics and splits

  • BanglaBox is about 9× Common Voice Bengali v9 and 17× the SUST TTS Corpus, and applies phonetic balancing across phone, diphone, triphone, and juktakkhor distributions.The comparison caption notes that the other listed non-English corpora do not apply phonetic balancing.Source: Table 2 · p. 3
  • The speaker-disjoint corpus split holds out 5 speakers (about 76 h) for zero-shot evaluation, within a roughly 312k-utterance corpus averaging 6.0 s across 7 domains.The duration and word-count histograms cover about 241k utterances of at least 2 s, while Figure 3’s summary statistics cover the full corpus.Source: §2.3 · p. 3, Figure 3 · p. 4

2.4 Tier A quality assurance

  • The coverage-feedback loop ran for 14 weekly iterations, lifting juktakkhor coverage from ∼70% (week 1) to the ≥85% reported above; measured per-gate rejection counts and rates are reported in Table 19.A Bangla-speaking linguist audits a 5% random sample/speaker for pronunciation correctness and prosodic naturalness, with ≥99% segment-level acceptance.Source: §2.4 · p. 4

3 Methodology

BanglaBox adapts Chatterbox to Bangladeshi Bangla through a three-stage fine-tuning pipeline, then uses chunked inference for long-form text. Training updates T3 while keeping the voice encoder and vocoder frozen.

3.1 Fine-tuning pipeline

  • A merge-consistent tokenizer extension grows the vocabulary from 704 to 4,240 tokens and filters invalid merges that could break inference tokenization.The 704 overlapping embedding rows are copied; 3,536 new rows are mean-initialized from the pretrained-row centroid.Source: §3.1 · p. 4
  • Bangla text normalization expands numerals and honorific abbreviations, applies Unicode normalization, and segments sentences through a six-step pipeline.The pipeline includes a 47-entry dictionary of religious honorific abbreviations absent from open packages.Source: §3.1 · p. 4
  • Only the 520M-parameter T3 is fine-tuned; the voice encoder and S3Gen remain frozen to preserve the speaker space and token-to-waveform mapping.Zero-shot cloning remains available at inference with Bangla references of 3–10 seconds.Source: §3.1 · p. 5
  • Prompt-masked dual cross-entropy trains text and speech tokens while excluding the first approximately 75 speech-prompt positions from speech loss.The prompt mask addresses decoder copying of prompt tokens rather than conditioning on the speaker embedding; only T3 is trainable.Source: §3.1 · p. 5

3.2 Long-form Bangla inference

  • Long-form inference splits paragraphs at Bangla sentence terminators, synthesizes each chunk with the same speaker embedding and prompt, trims silence, and inserts 200 ms pauses.Chunking addresses the 850-token per-utterance cap and premature EOS or drift in multi-sentence paragraphs.Source: §3.2 · p. 5
  • Bangla-calibrated inference lowers temperature from 0.8 to 0.3 and sets minimum-newtokens to 150 to counter English-phoneme hallucinations and premature EOS.The settings are used alongside the chunked pipeline.Source: §3.2 · p. 5

4 Evaluation Protocol – BanglaEval

BanglaEval combines native-speaker ratings, objective speech measures, and statistical testing to compare BanglaBox with commercial and open baselines. BanglaBox leads the reported perceptual metrics, while cloning and ablation results characterize its performance and components.

4 Evaluation Protocol – BanglaEval

  • The protocol adds a difficult-real-world-text stress test, public BnTTS benchmark evaluation, and a register control on naturally occurring Bangla.These three evaluations are reported in §5.1.Source: §4 · p. 5

4.4 Main results

  • BanglaBox achieves the highest Naturalness, SMOS, and Clarity; all pairwise differences are significant, and it leads cloning-capable baselines on SECS and ranks highest on NISQA.CER is competitive with Google and Azure and ahead of all open baselines; all 15 comparisons clear α/5=0.01.Source: Table 3 · p. 6, Table 4 · p. 6, §4.4 · p. 7

4.1 Training setup

  • BanglaBox is fine-tuned for 30–50 epochs on a speaker-disjoint training split, with Bangla-calibrated hyperparameters and chunked inference across conditions.The training configuration is specified in Table 9.Source: §4.1 · p. 6

4.2 Metrics and rater protocol

  • Twenty-four native Bangla speakers rate 76 utterances across three five-point scales, yielding 1,824 ratings per system; paired Wilcoxon tests use Bonferroni correction.Objective measures include CER, NISQA, and SECS over 436 test utterances.Source: §4.2 · pp. 6–7, Figure 5 · p. 6

4.3 Baselines

  • Comparisons include three commercial systems, IndicF5, YourTTS fine-tuned on the corpus, same-corpus XTTS-v2 fine-tuning, and human recordings as an upper bound.Direct BnTTS comparison is precluded by unreleased weights.Source: §4.3 · p. 7

4.5 Voice cloning

  • With held-out speakers, BanglaBox surpasses YourTTS-FT, IndicF5, and same-corpus XTTS-v2; its indicative BnTTS comparison is 0.627 vs. 0.548 SECS, using 520.3 h vs. 3.85k h.The BnTTS comparison is not head-to-head because the encoder and model family differ.Source: §4.5 · p. 7

4.6 Ablations

  • Removing text normalization causes the largest CER and Naturalness drop, while removing the prompt mask causes the largest SECS drop; naive tokenizer extension degrades all metrics.Un-freezing the voice encoder changes Naturalness and SECS by −0.01; the paper reports this observed change without characterizing it as noise.Source: §4.6 · p. 7
  • From 100 h to 520 h, all metrics improve, with diminishing returns beyond 200 h for CER and SECS while Naturalness continues to gain.The caption identifies Fig. 6 as showing Naturalness and CER scaling across 100/200/520 h subsets.Source: Figure 6 · p. 7, §4.6 · p. 8

5 Analysis

The analysis tests robustness beyond aggregate metrics, examining scaling, long-form synthesis, prosody, and persistent error categories. External evaluations cover difficult text, public benchmarks, and naturally occurring Bangla.

5 Analysis

  • Krippendorff’s α is 0.44 for Naturalness, 0.50 for SMOS, and 0.44 for Clarity; paired Wilcoxon tests can be significant despite overlapping marginal intervals.The intervals include rater and item variation, whereas paired tests cancel rater severity and item difficulty.Source: §5 · p. 8
  • As training data scales, CER saturates while Naturalness continues improving, supporting the paper’s phonetic-balance thesis against a 3.85k h Bangla regime.The analysis describes this contrast with BnTTS as a scaling comparison.Source: §5 · p. 8
  • Above the 850-token cap, single-pass synthesis is unreliable; chunking reduces premature truncation and length-driven degradation across 40 multi-sentence paragraphs, at a throughput cost.Below the cap, chunked and single-pass synthesis are directly comparable.Source: §5 · p. 8
  • F0 RMSE and duration deviation both worsen with longer utterances, and duration deviation doubles, consistent with timing drifting farther from the reference.The analysis splits results at the median reference duration.Source: §5 · p. 8
  • Three error categories persist: rare three-consonant conjuncts, code-switching boundaries, and onomatopoeia.The paper directs readers to Appendix D.4 for these failure modes.Source: §5 · p. 8

5.1 Robustness, external validation and register control

  • On the seven-category stress set, errors stay low and speaker similarity remains stable; addresses and spelling errors are hardest, near the in-domain reference.Each sentence was synthesized three times without selecting the best output.Source: Table 8 · p. 8, §5.1 · p. 8
  • On sampled public BnTTS evaluation sets, BanglaBox improves CER over reported BnTTS figures on all three; held-out naturally occurring Tier B text has CER 2.5% against 2.2%.The sampled BnTTS rows differ in size; Tier A scripts are LLM-generated, while Tier B provides the register control.Source: §5.1 · p. 8

6 Conclusion

The conclusion presents BanglaBox as a recipe for adapting English-pretrained TTS to an underrepresented language, combining tokenizer extension, text normalization, and prompt-masked fine-tuning over a 520.3 h JSD-balanced corpus. It reports gains across perceptual metrics and validation beyond the internal split.

6 Conclusion

  • The adaptation recipe combines tokenizer extension, text normalization, and prompt-masked fine-tuning over a 520.3 h JSD-balanced corpus.The conclusion frames the recipe as adapting English-pretrained TTS to an underrepresented language.Source: §6 · p. 8
  • BanglaBox beats five baselines on all three perceptual scales and matches prior few-shot speaker similarity using approximately one-seventh the Bangla fine-tuning audio.The conclusion reports validation on a seven-category stress set, public BnTTS benchmarks, and naturally occurring Bangla, with all artifacts public.Source: §6 · p. 8

Limitations

BanglaBox’s evaluation and training data have scope and measurement constraints, including limited speaker diversity, tier differences, and text-register concerns. The authors also identify boundaries in regional generalization, inference, and comparison with prior work.

Limitations

  • The 36-speaker corpus is small relative to VCTK (110) and LibriTTS (2,456); broader speaker diversity may improve cloning generalization.The authors identify this as a limitation on establishing cloning generalization.Source: Limitations · pp. 8–9
  • Tier A and Tier B differ in recording conditions and provenance, so residual tier variance may persist and warrants per-tier evaluation.Tier B was empirically curated after 16 of 39 initial speakers were dropped for degrading downstream metrics.Source: Limitations · p. 9
  • All Tier A scripts are LLM-generated, leaving possible translationese or LLM-register phrasing; Tier B’s 153.0 h is the only non-LLM training text.A direct text-level register analysis remains future work, although held-out natural Tier B text showed essentially unchanged error.Source: Limitations · p. 9
  • A direct BnTTS comparison remains unavailable because its weights are unreleased and the systems differ in speaker encoders, model families, and evaluation protocols.The authors describe prior BnTTS comparisons as indicative, despite evaluating public BnTTS sets under their own objective protocol.Source: Limitations · p. 9
  • BanglaBox occasionally produces English-like prosody in long utterances, with the measured gap in pitch range, pause length, and speaking rate growing with utterance length.Measurements are against held-out human recordings.Source: Limitations · p. 9
  • Medical and legal abbreviations and Banglish transliteration are not yet handled by the text-normalization system.Source: Limitations · p. 9
  • The system does not stream; an RTF of 0.42 under the stated single-utterance, unbatched H100 condition is described as sufficient for offline use only.The 0.66 value in Table 17 comes from one pass over 40 multi-sentence paragraphs, a different measurement condition.Source: Limitations · p. 9
  • Evaluation targets Standard Bangladeshi Bangla (Dhaka register); regional varieties in West Bengal, Assam, and Tripura remain untested.The authors identify a MUSHRA-style test with a fixed reference and anchor clips as a stronger next evaluation step.Source: Limitations · p. 9

Ethics statement

The paper documents consent, evaluator protections, and safeguards for its voice-cloning release, while noting potential misuse and intended beneficial applications.

Ethics statement

  • Every corpus speaker provided written informed consent for recordings to be used in TTS research.Source: Ethics statement · p. 9
  • The evaluation study had institutional ethics-board approval and evaluator consent, and collected no personally identifiable information.Source: Ethics statement · p. 9
  • Evaluators were paid USD 15/hour, above local prevailing-wage thresholds and crowd-work-guideline floors; corpus speakers were paid above local market rates in BDT.Source: Ethics statement · p. 9
  • To address impersonation and non-consensual synthetic-speech risks, generated audio is watermarked, corpus voices are consent-verified, and recording-use terms exclude commercial use.The backbone’s Perth mark is applied unconditionally to every generated waveform for third-party detection.Source: Ethics statement · pp. 9–10
  • The paper identifies accessibility, education, and information access as potential benefits of high-quality Bangla TTS when used responsibly.Source: Ethics statement · p. 10

A Implementation Notes

The appendix details the tokenizer, normalization, and prompt-masked training stages, alongside rater screening, listening controls, statistical testing, and baseline documentation.

A.1 Stage 1 — Tokenizer extension: full breakdown

  • The final vocabulary has 4,240 tokens: 704 copied English units, 1,750 multilingual units, and 1,786 Bangla units.The Bangla units comprise 73 Bengali characters, 1,712 corpus-trained BPE subwords, and a [bn] tag; Bengali dari is already in the base vocabulary.Source: §A.1 · p. 13
  • The merge-consistency filter drops imported BPE rules unless both inputs and their merged output exist in the target vocabulary; about 11% of Bangla-side merges are dropped.This prevents an invalid merge from producing an UNK at inference; the appendix contrasts the approach with byte fallback and wholesale tokenizer retraining.Source: §A.1 · p. 13, §A.2 · p. 13

A.2 Relation to alternative tokenizer-extension strategies

  • The appendix contrasts merge-consistent BPE with byte fallback, which roughly triples Bangla token length, and tokenizer retraining, which severs the link to pretrained embeddings.The 256-token input cap means ordinary Bangla sentences would not fit under byte fallback.Source: §A.2 · p. 13

A.3 Stage 2 — Text normalization: step-by-step examples

  • The normalization examples trace six steps, including a hand-curated dictionary of 47 honorific abbreviations, while code-switched tokens are left untouched.Medical, legal, and banking abbreviations and Banglish transliteration are out of scope; Table 10 presents a worked example.Source: §A.3 · p. 14, Table 10 · p. 14

A.4 Stage 3 — Prompt-masked loss: implementation notes

  • The prompt mask excludes prompt and padding positions from speech-token cross-entropy; without it, a control model parroted prompt audio and had lower SECS and SMOS.The control’s SECS fell from 0.627 to 0.51 and SMOS from 4.51 to 3.92; the prompt-mask checkpoint did not show the behavior.Source: §A.4 · p. 14

B.1 Rater pool: detailed screening and assignment

  • The rater protocol screened 71 applicants to form a 24-person pool and assigned every rater all 76 utterances across seven systems.Each system received 1,824 ratings per subjective dimension; screening included a Bangla transcription test and listening-environment checks.Source: §B.1 · pp. 14–15

B.2 Listening environment validation

  • Listening sessions standardized headphones, loudness, room noise, duration, and breaks; randomized catch trials tested similarity judgments and rejection of degraded audio.Raters failing either catch trial were excluded and replaced, with their playlist re-rated.Source: §B.2 · p. 15, §B.3 · p. 15

B.4 Statistical testing details

  • For Naturalness, SMOS, and Clarity, BanglaBox significantly exceeded all five baselines after Bonferroni correction, while the Human reference remained significantly higher; Table 14 reports two-sided p-values and rank-biserial effect sizes for 76 per-item differences.Source: §B.4 · p. 16

B.6 Juktakkhor pronunciation evaluation

Juktakkhor pronunciation remains weakest for conjuncts rarely seen in training, while matched-hours experiments connect balanced selection with broader rare-conjunct coverage and improved error rates. The section also reports the corpus-construction checks and measurement choices used to assess selection and synthesis.

B.6 Juktakkhor pronunciation evaluation

  • 88.2% phoneme-level accuracy on rare conjuncts, versus 97.7% on frequent conjuncts, leaves a 9.5-point gap; conjuncts in under 0.3% of training utterances remain weakest.The accuracy declines monotonically across frequent, moderate, and rare groups.Source: §B.6 · p. 16, Table 16 · p. 16

B.7 Balanced vs. random selection (matched hours)

  • At matched 100-hour training budgets, balanced selection raises rare-conjunct coverage from 79.8% to 91.7% and lowers CER from 7.1% to 4.3%.Across three seeds, Naturalness is 4.31 for balanced selection and 4.18 for random selection.Source: §B.7 · pp. 16–17

B.8 Objective prosody measurement

  • Prosody is measured against held-out ground truth for 56 utterances, split at the median reference duration.The measurements are F0 RMSE and duration deviation.Source: §B.8 · p. 17

B.9 NISQA measurement details

  • NISQA scores average predictions over 436 test utterances, but its English- and German-heavy training makes absolute Bangla predictions imperfectly calibrated.Its system ranking was consistent with the native-rater Naturalness ranking, so it is used as a complementary metric.Source: §B.9 · p. 17

C.1 Tier A studio recording protocol / C.2 Tier A quality assurance (extended)

  • Tier A recording used 13 professional Dhaka voice actors, while session checks and weekly linguist audits supported segment-level quality control.The recordings enforced an SNR floor of at least 44 dB; the linguist audit reported at least 99% segment-level acceptance.Source: §C.1 · pp. 17–18, §C.2 · p. 18

C.3 9 automated quality gates — full thresholds

  • The gates rejected 6,807 of 206,943 candidates (3.29%); gates 6 and 9 steer prompts against live coverage gaps rather than filtering individual candidates.Because those two gates are prompt-level, no per-candidate rejection rate is defined for them.Source: Table 19 · p. 17, §C.3 · pp. 18–19

C.4 Coverage-feedback loop — week-by-week trajectory

  • Juktakkhor coverage rose from approximately 70% in week 1 to at least 85% by week 14, with at least 97% coverage for the top 150 conjuncts.Tiered-JSD loss decreased monotonically over the 14 weekly iterations.Source: §C.4 · p. 19

C.5 Tiered-JSD selection algorithm / C.6 Weight-sensitivity of the tiered-JSD objective

  • The selection algorithm’s coverage figures are achieved outcomes, not preset stopping thresholds; coverage remained unchanged under every tested weight perturbation.The weight-sensitivity table reports unchanged coverage across all four tiers under every perturbation.Source: §C.5 · p. 19, Table 22 · p. 19

C.7 Tier B filter — ablation-driven criterion

Tier B curation filters public Bangla speech sources first by signal quality and then by each speaker’s measured contribution to development performance. The accompanying examples document honorific normalization, conjunct types, and utterances from several domains.

C.7 Tier B filter — ablation-driven criterion

  • 23 of 39 speakers were retained in Tier B, contributing 153.0 hours after SNR and ablation-driven filtering.The second pass dropped speakers whose inclusion did not improve, or degraded, development CER or SECS.Source: §C.7 · pp. 19–20, Table 23 · p. 20

D.1 Honorific abbreviation expansion

  • The honorific dictionary contains 47 terms, with Table 24 showing 12 representative entries used in Stage 2 normalization.These honorifics are frequent in Bangla religious and cultural literature but absent from open normalization toolchains.Source: §D.1 · p. 20, Table 24 · p. 20

D.2 Juktakkhor (conjunct) inventory examples

  • The conjunct inventory spans frequent two-consonant forms, moderate forms, and rare three-consonant forms occurring in under 0.3% of training utterances.The examples distinguish the inventory by training-frequency group.Source: §D.2 · p. 20

D.3 Cross-domain Bangla utterance samples

  • The text examples include utterances from news, healthcare, and IT domains.The excerpts illustrate the corpus’s cross-domain Bangla text.Source: §D.3 · p. 20

D.4 Failure modes with Bangla examples

The reported failure modes cluster around rare conjuncts, Bangla–English code-switch boundaries, and highly colloquial onomatopoeia. The paper also reports the compute used for training and evaluation.

D.4 Failure modes with Bangla examples

  • Conjuncts appearing in under 0.3% of training utterances can produce distorted phoneme sequences, such as a brief stutter or vowel insertion.The gap-unit gate up-weights these forms during script generation, but coverage remains uneven.Source: §D.4 · p. 20
  • Bangla–English code-switch boundaries can produce brief silences of 5–15 ms and slight pitch resets; explicit language-tag injection is absent.The paper identifies this as an opportunity for follow-up.Source: §D.4 · p. 20
  • Onomatopoeia with fewer than five training occurrences can yield unreliable output, including hallucinated near-neighbour words.The examples are regional sound-words such as jhomjhom, tuptap, and guṛguṛ.Source: §D.4 · p. 21

E Computing Infrastructure

  • The nine training runs used approximately 175 H100-hours, with about 4.5 additional H100-hours for inference.The reported total compute budget is approximately 180 H100-hours plus CPU preprocessing and evaluation work.Source: E Computing Infrastructure · p. 21

Checklist

The checklist records artifact documentation, evaluation and compute reporting, human-participant procedures, and disclosed AI use. It also summarizes the paper’s stated limitations and deployment risks.

Checklist

  • The stated limitations include modest speaker count, Tier A/Tier B heterogeneity, no BnTTS head-to-head comparison, long-utterance prosody bias, and Bangladesh-only evaluators.Other listed boundaries include text-normalization gaps, no streaming, and language-agnostic ECAPA-TDNN for speaker similarity.Source: Checklist · p. 21
  • The checklist reports corpus documentation across seven domains and a speaker-disjoint test split with five held-out speakers for zero-shot evaluation.The reported split includes 31 training speakers and approximately 76 hours for the held-out test speakers.Source: Checklist · p. 22
  • The checklist reports approximately 180 H100-hours total compute and documents model, hardware, training, and inference settings.The compute accounting includes approximately 175 H100-hours of training and 4.5 H100-hours of inference.Source: Checklist · p. 22
  • Subjective results report 95% confidence intervals and pairwise Wilcoxon signed-rank tests with Bonferroni correction; objective CER and SECS use 436 test utterances.The checklist states p<0.01 for the main results and describes 24 native Bangla-speaking evaluators.Source: Checklist · p. 22
  • The evaluation checklist describes three independent five-point scales, catch trials, evaluator compensation, consent, and institutional ethics review.The full BanglaEval protocol is identified as the source for the participant instructions and quality-control details.Source: Checklist · pp. 22–23
  • LLMs generated candidate Tier A scripts, which passed nine automated gates and two-pass human review before inclusion in the corpus.The checklist also discloses AI-assisted language polishing and states that technical claims and results were authored by the human authors.Source: Checklist · p. 23

Related papers

Loading 2610.09211v1…