FraPiz/xtts-v2-moldovan-romanian

XTTS-v2 for Moldovan Dialectal Romanian

This checkpoint adapts coqui/XTTS-v2 to Romanian educational speech with Moldovan dialectal characteristics. XTTS-v2 is a multilingual voice-cloning model and requires a short reference recording at inference time.

Training data and configuration

TTS preparation started from a 35,760-segment Romanian educational-speech corpus. Segments exceeding the configured maximum waveform length were removed, leaving 31,472 eligible utterances (55.05 hours). With seed 42, these were split as follows:

Partition	Utterances	Duration
Train	25,178	44.06 h
Evaluation	3,147	5.47 h
Test	3,147	5.52 h

Training used an NVIDIA A100-SXM4-80GB GPU for 30 epochs, batch size 4, gradient accumulation 63 (effective batch size 252), and learning rate 5e-6. The recorded execution lasted 8 hours and 6 minutes.

Hugging Face dataset: https://huggingface.co/datasets/FraPiz/moldovan-dialectal-romanian-speech-corpus

Original Google Drive release: https://drive.google.com/drive/folders/1Va0tmpp-Q8A2gGbFwLignYVtf0KD6Uth

Files

• best_model.pth: fine-tuned XTTS-v2 checkpoint.
• config.json: model and inference configuration.
• vocab.json: tokenizer vocabulary.
• dvae.pth and mel_stats.pth: XTTS-v2 acoustic assets.
• generate_tts.py: local inference example, including Romanian text normalization used during evaluation.

No reference-speaker recording is included. Supply a reference WAV for a voice that you have permission to use.

Installation and usage

Use Python 3.10 or 3.11 and install the dependencies from requirements.txt. Download the repository, place a reference recording at reference_voice/reference.wav, edit TEXT in generate_tts.py, and run:

python generate_tts.py

The included script loads best_model.pth explicitly, applies Romanian cedilla-to-comma-below normalization, and synthesizes at 24 kHz. A CUDA-capable GPU is strongly recommended.

Evaluation

Automatic intelligibility proxies on 10 generated prompts produced 1.68% WER (95% CI 0.00-4.13), 0.36% CER (0.00-0.86), and generation RTF 0.388. These scores were obtained by transcribing generated speech with the accompanying fine-tuned ASR model and are not independent measures of speech quality.

A randomized, model-blinded pilot evaluation involved 10 listeners (8 native Moldovan listeners), 10 matched prompts, five female and five male reference voices, and 20 samples per listener. Naturalness and timbre similarity favored the fine-tuned model after listener-level multiplicity correction. Moldovan dialectal adequacy showed a positive but statistically uncertain trend; pronunciation favored the Romanian baseline, and sentence-ending stability was nearly tied.

Limitations and responsible use

• The listening study is preliminary and contains only 10 listeners.
• The training data emphasize planned educational speech, not spontaneous conversation.
• The automatic WER/CER comparison contains only 10 prompts.
• Output quality depends strongly on reference-audio quality, text normalization, and inference parameters.
• Voice cloning must only be performed with informed consent or another valid legal basis. Do not use this model for impersonation, deception, fraud, or unauthorized biometric processing.

Authors and project

Marius Cerescu, Alexandr Parahonco, Olesea Caftanatov, Tudor Bumbu, Nichita Degteariov, and Ion Bostan. Vladimir Andrunachievici Institute of Mathematics and Computer Science, Moldova State University, Chisinau, Republic of Moldova.

This work was elaborated within the project Platforma Educationala bazata pe Inteligenta Artificiala "Guguta", registered in the State Register of projects in science and innovation, code 25.80012.0807.21TC, project leader Tudor Bumbu, Dr.

Citation

@inproceedings{cerescu2026moldovan,
  title     = {Fine-Tuning Neural Speech Models for Romanian Speech Recognition and Synthesis with Moldovan Dialectal Data},
  author    = {Cerescu, Marius and Parahonco, Alexandr and Caftanatov, Olesea and Bumbu, Tudor and Degteariov, Nichita and Bostan, Ion},
  booktitle = {International Conference on System Analysis and Intelligent Information Technologies (SAIIT)},
  year      = {2026}
}

License

This derivative model is distributed under the Coqui Public Model License 1.0. Review LICENSE.txt before downloading or using the checkpoint.