Greek TTS from 3.5h Data: Deterministic Prompts Fix Speaker Drift, Near-Human MOS

kastnerkyle · x · 2026-09-13

An Interspeech 2026 paper shows how to build high-quality single-speaker TTS for a low-resource language.

Data curation: WhisperX alignment and filtering turns audiobook recordings into TTS-ready data
Base model: fine-tuned Parler-TTS (880M), whose pre-training encodes phonetic priors transferable to Greek
Key finding: LLM-generated style prompts cause speaker drift at inference; replacing them with deterministic prompts resolves it
LoRA anchoring: a speaker-specific LoRA trained on just 3.5h of single-speaker data (5% of parameters) anchors identity
Results: WER 10.7% (2.9 above the ASR floor), MOS-I 4.00 vs 4.36 for human speech, speaker consistency MOS-C 4.24 vs 4.30

Conclusion: robust single-speaker Greek TTS is achievable with limited curated data.

More from Multimodal