Deterministic Prompting for Speaker-Stable Low-Resource Greek TTS
Georgios Syllas, Efthymios Georgiou, Kosmas Kritsis, Alexandros Potamianos
cs.SD · cs.CL · cs.LG
Quick verdict
Claims deterministic prompts + LoRA on 3.5h yield stable Greek TTS near human quality; recipe is credible, but key stats are only trends.
Summary
The paper targets low-resource Modern Greek text-to-speech. The authors convert audiobook and Common Voice recordings into TTS-ready clips via WhisperX forced alignment, confidence/SNR filtering, and 1.5–10 s segmentation (~23 h for language adaptation). They fine-tune Parler-TTS (880M), replacing the standard LLM-generated style descriptions with deterministic, binned acoustic-label prompts after observing that prompt variety caused speaker drift; a speaker-specific LoRA stage (r=16, ~25M trainable parameters, 2 epochs) on 3.5 h of one male speaker then anchors identity. The best configuration reaches 10.7% WER on a 50-utterance held-out set, 2.9 points above the ASR floor on human audio, MOS-I 4.00, and MOS-C 4.24 versus 4.30 for human pairs. A listening study with 29 native Greek listeners found synthetic systems scored higher naturalness than ground truth; the authors attribute this to recording conditions and claim no superiority. Limitations include the single male speaker, privately licensed audio sources, small evaluation sets, and key statistical comparisons that remain non-significant after Holm correction.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
4 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
- Replacing stochastic LLM-generated style prompts with deterministic, binned acoustic-label prompts removes prompt-induced variance and, combined with speaker-specific LoRA, resolves speaker drift (abstract, §3.2).
- A speaker-specific LoRA stage on only 3.5 h of single-speaker data (~25M trainable parameters) anchors stable speaker identity without overwriting multilingual phonetic priors (§3.2).
- The best configuration (Det.+LoRA) reaches WER 10.7% — 2.9 pp above the 7.8% ASR floor on human audio — MOS-I 4.00, and near-human intra-system speaker consistency (MOS-C 4.24 vs 4.30) (abstract, Tables 3–4).
- Transcription accuracy outweighs raw data volume in low-resource TTS, evidenced by excluding the 7.5 h auto-filtered audiobook corpus whose ASR errors increased hallucinated syllables (§2).
Strengths
Honest, well-hedged statistical reporting
Section 5.2 reports Friedman/Wilcoxon with Holm correction and effect sizes, and explicitly labels non-significant results: naturalness 'inconclusive' (χ2=3.83, p=0.43), MOS-C omnibus non-significant (p=0.19), and 'non-significant differences from human recordings should not be interpreted as equivalence' (§6).
Transparent negative results and design decisions
The paper discloses the failed VITS baseline (§3.1, 'insufficient for formal evaluation'), exclusion of the 7.5 h auto-filtered Audiobook-2 due to ASR-error-induced hallucinations (§2, Table 1), and the failed Seed-VC augmentation attempt (§2).
Practical, compute-frugal recipe with full hyperparameters
Clear articulation of the speaker-averaging failure mode
Weaknesses
Very small evaluation sets
Core corpora cannot be redistributed
Headline wording outruns its own significance tests
Spectral fidelity does not improve for the best configuration
Single-run experiments with confounded comparisons
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
This Interspeech-style paper proposes a recipe for single-speaker Modern Greek TTS under data scarcity. Contributions (§1): (i) a WhisperX-based curation pipeline turning audiobook/community audio into TTS-ready clips; (ii) two-stage Parler-TTS adaptation — full fine-tuning on ~23 h of Greek (Table 1: 15.5 h filtered Common Voice, 4 h CSS10, 3.5 h verified audiobook) followed by a speaker-specific LoRA (r=16, ~25M params = ~5% of the 500M decoder, 2 epochs on 3.5 h); (iii) the central empirical claim that LLM-generated style prompts cause speaker drift and that deterministic, binned prompts fix it. Headline numbers: WER 10.7% vs 7.8% ASR floor, MOS-I 4.00 vs 4.36 for human, MOS-C 4.24 vs 4.30 (Tables 3–4).
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 4 strengths & 5 weaknesses, analyzed in detail
- 5 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 paper 'Deterministic Prompting for Speaker-Stable Low-Resource Greek TTS' achieve?
It builds a Modern Greek text-to-speech system from limited data: about 23 hours of curated audio for language adaptation plus 3.5 hours from a single male speaker for a LoRA adaptation stage. The best model transcribes with 10.7% word error rate (about 3 points above what an ASR system achieves on real human Greek audio) and achieves near-human voice consistency in listener ratings (4.24 vs 4.30 for human recordings).
Why did the authors replace LLM-generated voice descriptions with deterministic prompts in Parler-TTS?
During development they found that varied, LLM-written style prompts made the synthesized voice drift between utterances. Replacing them with fixed, binned acoustic labels (pitch, speed, clarity, etc.) removed that variance. Notably, deterministic prompts alone actually scored worse on word error rate (18.8% vs 15.2%); the large benefit appeared only after the speaker-specific LoRA stage (10.7% vs 21.1%).
How much data is needed to replicate this Greek TTS recipe?
About 3.5 hours of clean single-speaker recordings for the speaker-specific LoRA stage, and roughly 23 hours total (including ~15.5 hours of filtered Common Voice and 4 hours of CSS10) for the earlier full fine-tuning stage. A second 7.5-hour audiobook corpus was excluded because automatic transcription errors increased hallucinated syllables. The audiobook sources are privately licensed, so the exact datasets cannot be redistributed; only preparation scripts, code and model weights are released.
What are the main limitations of this Greek TTS work?
It was validated on one language (Greek), one male speaker, and a reading style. The audiobook sources are private. Objective evaluation used small sets (50 utterances for WER, 20 for speaker metrics). Several perceptual comparisons, including the speaker-consistency advantage of deterministic prompts, were not statistically significant after multiple-comparison correction (marginal trend, padj = 0.076), which the authors themselves acknowledge. Speaker similarity to the target voice also remained modest (SIM-S about 0.61).
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.
- The paper's Table 1, Table 4 and Figure 1 are reproduced from the extracted text; the figure itself was not inspected, so any statement about Fig. 1 relies on its caption only.
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.