Structure Before Sampling: Community-Aware Core-Set Selection for Data-Efficient Text-to-Speech
Mizbaul Haque Maruf, Mizbaul Haque Maruf, Muhammad Nur Yanhaona
cs.CL
Quick verdict
Structure-aware core-set selection beats random and entropy baselines for TTS at equal budgets; solid null tests, though ablations and speaker scope are thin.
Summary
The paper asks whether a TTS training corpus, viewed as a graph of utterances linked by phonemic similarity, has exploitable structure, and turns that structure into a core-set selector. Utterances become TF-IDF vectors over IPA phoneme unigrams and within-word bigrams; a k=15 nearest-neighbour graph is tested against Erdős–Rényi and degree-preserving nulls. On Bangla Common Voice recordings (40,422 utterances, 48.75 h) and LJSpeech (13,100 utterances, 23.92 h), clustering exceeds the random null by 199× and 56×, and modularity exceeds the configuration model (0.470 vs 0.172; 0.404 vs 0.177). The proposed Community Representative partitions the graph with Louvain, allocates the duration budget proportionally across communities, and spreads picks by farthest-point sampling seeded by rare-phoneme richness. It covers more rare phoneme bigrams than Phoneme Balance and Random at every budget (at 10% in Bangla: 89.5% vs 73.2% vs 49.9%), including on held-out utterances and a six-speaker corpus. MB-iSTFT-VITS models trained on 20% core-sets reach significantly lower CER than both equal-duration baselines in both languages; under matched epochs in Bangla the core-set model beats full-corpus training (3.93% vs 4.47% CER) with 4.5× less training time, while in English, matched on updates, full data stays best. Limitations include single-speaker corpora, one TTS architecture and ASR evaluator, no ablation isolating community stratification or the rarity seed, and ~25% more utterances in the core-set arm at equal duration.
Review complete
Assessment at a glance
4
risk flags identified
● high 0
● medium 2
● low 2
4 strengths · 4 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
- The phonotactic k-NN utterance graph has significant structure in both languages: clustering 199× (Bangla) and 56× (English) the Erdős–Rényi value, with modularity above a degree-preserving configuration model.
- Community Representative covers more rare phoneme bigrams than random and entropy-based selection at every budget in both languages, on held-out utterances and on a six-speaker corpus.
- TTS models trained on 20% core-sets achieve significantly lower CER than models trained on equal-duration random or Phoneme Balance subsets in both Bangla and English.
- Under matched epochs in Bangla, the 20% core-set model outperforms full-corpus training (3.93% vs 4.47% CER) with 4.5× less training time; this full-data advantage does not hold under matched updates in English.
Strengths
Structure is tested, not assumed
§4 tests the k-NN utterance graph against Erdős–Rényi and degree-preserving configuration nulls averaged over five realizations: clustering 0.1401 vs 0.0007 (Bangla, 199×) and 0.1211 vs 0.0022 (English, 56×); modularity 0.470 vs 0.172 and 0.404 vs 0.177 against the configuration model, which controls for the observed heavy degree tail (kmax 627/461, §4.2, Figure 3).
Cross-language replication under shared hyperparameters
Bangla and English share k=15, budgets, seeds, features and the rarity definition (§3.2, Table 1), so the replication of both the structural findings and the downstream selector ordering (Table 7) is not an artifact of per-language tuning.
Circularity controls in the intrinsic evaluation
Careful downstream evaluation hygiene
Weaknesses
No component ablation of the selector
Downstream confounds
Narrow evaluation surface
Metric–method alignment and fixed k
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper targets duration-budgeted core-set selection for TTS corpus construction. Its four central claims: (1) a phonotactic k-NN graph over utterances has statistically significant structure — clustering 199× (Bangla, 0.1401 vs 0.0007) and 56× (English, 0.1211 vs 0.0022) the Erdős–Rényi value, and modularity above a degree-preserving configuration model (0.470 vs 0.172; 0.404 vs 0.177, Table 2); (2) the proposed Community Representative selector covers more rare phoneme bigrams than random and entropy-based (Phoneme Balance) selection at every budget (10/20/40%) in both languages, on held-out utterances (Table 5) and on a six-speaker corpus (Table 6); (3) TTS models trained on 20% core-sets reach significantly lower CER than equal-duration random and Phoneme Balance subsets in both languages (Table 7); (4) under matched epochs in Bangla, the 20% core-set model beats full-corpus training (3.93% vs 4.47% CER, p = 0.011) with 4.5× less training time (11.3 h vs 50.4 h). The authors correctly scope claim (4): under matched gradient steps in English, full data remains best (2.84% vs 2.96%), and §8.2 explicitly refrains from claiming core-sets beat all data.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 4 strengths & 4 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 does the Community Representative method do?
It selects a small subset of a TTS corpus under a fixed audio-duration budget. The corpus is turned into a graph where utterances are linked to their most phonemically similar neighbours; the graph is split into communities with Louvain, each community gets a budget share proportional to its audio duration, and within each community the selector picks well-spread utterances by farthest-point sampling, starting from the utterance richest in rare phonemes.
How much TTS training data can this method save?
In the paper's experiments, models trained on a 20% core-set were more intelligible (lower character error rate) than models trained on random or phonetically balanced subsets of the same duration, in both Bangla and English. In Bangla, the 20% subset also beat full-corpus training (3.93% vs 4.47% CER) with 4.5× less training time — but that full-data comparison held only when models were trained for equal epochs, not under equal gradient updates in English.
Which datasets were used in the study?
Bangla: Common Voice Bengali sentences recorded by a single in-house speaker (40,422 utterances, 48.75 hours; recordings not publicly released). English: LJSpeech-1.1 (13,100 utterances, 23.92 hours, one speaker). A six-speaker Bangla corpus (OpenSLR SLR37) was used only to test that the coverage advantage transfers beyond single-speaker data.
Is code available for this paper?
The paper text does not mention a code release. The pipeline is described in enough detail to reimplement (epitran/CMUdict G2P, TF-IDF phoneme features, k=15 k-NN graph, Louvain, farthest-point sampling), and LJSpeech is public, but the Bangla recordings are private, so a full replication of the Bangla experiments is not possible from public data.
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.
- Figures 2–6 are referenced and their captions are in the text, but the figure content itself cannot be verified from the extraction; judgments rely on in-text numbers and tables.
- The extracted text appears complete (all sections, limitations, ethics, references), but any supplementary appendix would not have been captured.
- The Bangla corpus audio is described as in-house and unreleased, so external replication of the Bangla experiments is limited to the described pipeline.
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.