Life sciences · Preprint

Deterministic Prompting for Speaker-Stable Low-Resource Greek TTS

arXiv · September 9, 2026

Preliminary Why this matters to you Expert discussion

Early or partial results. Treat as a signal, not a conclusion.

AI Summary

This is an unreviewed system development report describing a deterministic-prompt approach to single-speaker Greek text-to-speech synthesis using fine-tuned Parler-TTS and LoRA adaptation on limited audiobook data. The authors report acceptable word error rate (10.7%) and speaker consistency (MOS-C 4.24) but near-human quality (MOS-I 4.00) with no comparative arms or peer review.

What was studied

Single-arm system development and technical validation study. Low-resource Greek language; single-speaker synthesis task. No human subjects enrolled.. Intervention: Deterministic prompting strategy combined with speaker-specific LoRA fine-tuning on 3.5 h audiobook data..

Key findings

Word error rate (WER) of 10.7%, 2.9 points above ASR floor Mean opinion score for identity (MOS-I) of 4.00 versus 4.36 for human speech Speaker consistency (MOS-C) of 4.24 versus 4.30 for human reference

Safety

Safety was not reported in the material analysed. Check the source before drawing any conclusion about harm.

Who this may help

The source did not state who this applies to in practice.

Plain-language summary

A proof-of-concept system for low-resource Greek TTS with engineering validation (WER, MOS scores) but no comparative trial, no peer review, and no clinical application.

Study design

As stated by the source record.

MethodsAudiobook recordings were curated using WhisperX alignment and filtering. Parler-TTS (880M parameters), a prompt-based multilingual model, was fine-tuned on the curated Greek data. During development, LLM-generated style prompts were found to induce speaker drift at inference; deterministic prompts were substituted. A speaker-specific LoRA (Low-Rank Adaptation) stage was trained on 3.5 hours of single-speaker data, updating approximately 5% of model parameters to anchor speaker identity. System performance was evaluated using word error rate (WER), mean opinion scores for identity (MOS-I) and…

DesignSingle-arm system development and technical validation study

PopulationLow-resource Greek language; single-speaker synthesis task. No human subjects enrolled.

InterventionDeterministic prompting strategy combined with speaker-specific LoRA fine-tuning on 3.5 h audiobook data.

Reported figures

Quoted from the source exactly as published.

Word error rate10.7%

WER gap above ASR floor2.9

Mean opinion score (identity)4.00

Human speech MOS reference4.36

Mean opinion score (speaker consistency)4.24

Human MOS-C reference4.30

LoRA training data duration3.5 h

LoRA parameter update proportion~5%

Caveats

Not peer reviewed; published on arXiv as preprint.
Evaluation sample size and rater count not disclosed.
No comparative arms (e.g., LLM-prompt baseline, other multilingual models, or human performance on identical sentences).
Metrics are technical and subjective (MOS); generalizability to multi-speaker or other low-resource languages not tested.

Evidence assessment

Graded across the dimensions that decide whether you should act, each from what the source actually supports. There is no single score, and where a dimension was not assessed it says so.

Study design

limited

Single-arm system development without RCT or comparative control arms.

Directness

limited

Metrics are technical (WER, MOS) rather than clinical or user outcomes.

Consistency

insufficient

No inter-rater reliability or cross-dataset validation reported.

Risk of bias

insufficient

No blinding, rater independence, or bias mitigation methods described.

Sample adequacy

limited

Training on 3.5 h single-speaker data; evaluation sample size not stated.

Source abstract

Modern TTS systems approach human quality for high-resource languages but degrade when clean speech data is scarce. Modern Greek exemplifies this, lacking the curated corpora behind state-of-the-art synthesis. We propose a data curation recipe that transforms audiobook recordings into TTS-ready data via WhisperX alignment and filtering. Then we fine-tune Parler-TTS (880M), a prompt-based multilingual model whose pre-training encodes phonetic priors transferable to Greek. During development, we find that LLM-generated style prompts introduce speaker drift at inference. Replacing them with deterministic prompts resolves this, and a speaker-specific LoRA stage trained on 3.5 h of single-speaker data anchors identity while updating ~5% of parameters. Our system achieves WER 10.7% (2.9 above the ASR floor), MOS-I 4.00 (vs. 4.36 human speech), and near-human speaker consistency (MOS-C 4.24 vs. 4.30), showing that robust single-speaker Greek TTS is achievable with limited curated data.

Follow from this

Taken from the source record, never inferred. Follow any of these and new work involving them reaches your briefing.

Georgios SyllasKosmas KritsisEfthymios GeorgiouAlexandros Potamianos