FraPiz/whisper-large-v3-turbo-moldovan-romanian

Whisper Large-v3-Turbo for Moldovan Dialectal Romanian

This model is a fine-tuned version of openai/whisper-large-v3-turbo for automatic speech recognition of Romanian educational speech containing Moldovan dialectal characteristics.

Training data

The model was adapted on the training partition of the Moldovan Dialectal Romanian Educational Speech Corpus. The complete corpus contains 35,760 validated audio-text segments (70.67 hours), derived from 418 educational lesson videos and 56 inferred speakers. The split was produced at segment level with seed 42 using an 80/10/10 protocol.

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

Evaluation

Evaluation used the complete held-out test partition of 3,576 segments (approximately 6.99 hours). References and hypotheses were normalized to Unicode NFC, stripped, and lowercased; punctuation was retained. Decoding used Romanian language conditioning and the transcription task.

Model	WER (95% CI)	CER (95% CI)	RTF
Base model, zero-shot	18.85% (17.71-19.99)	7.20% (6.60-7.81)	0.020
This fine-tuned model	3.18% (2.83-3.56)	0.98% (0.80-1.20)	0.021

Confidence intervals use 10,000 paired bootstrap replicates clustered by source video. The test partition contains unseen segments, but not unseen speakers or source videos. These results therefore measure matched-domain segment generalization and must not be interpreted as speaker-independent performance.

Usage

import torch
from transformers import AutoModelForSpeechSeq2Seq, AutoProcessor, pipeline

model_id = "FraPiz/whisper-large-v3-turbo-moldovan-romanian"
device = 0 if torch.cuda.is_available() else -1
dtype = torch.float16 if torch.cuda.is_available() else torch.float32

model = AutoModelForSpeechSeq2Seq.from_pretrained(
    model_id,
    torch_dtype=dtype,
    low_cpu_mem_usage=True,
)
processor = AutoProcessor.from_pretrained(model_id)

asr = pipeline(
    "automatic-speech-recognition",
    model=model,
    tokenizer=processor.tokenizer,
    feature_extractor=processor.feature_extractor,
    torch_dtype=dtype,
    device=device,
)

result = asr(
    "audio.wav",
    generate_kwargs={"language": "ro", "task": "transcribe"},
)
print(result["text"])

Limitations

• The training material is dominated by planned educational speech.
• Evaluation does not establish performance on unseen speakers, independent source videos, spontaneous conversation, or other recording domains.
• Accuracy may decrease for code-switching, heavy noise, distant microphones, and dialectal varieties not represented in the corpus.
• Speaker gender metadata in the corpus was inferred from names and is only approximate descriptive metadata.

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

The model is released under the Apache License 2.0, consistently with the base Whisper model. Users remain responsible for compliance with applicable laws and with the terms governing any audio processed with the model.