GLM-5.2 is live. #1 throughput on OpenRouter, pay-per-token on FriendliAI. Try it today ➜
FraPiz
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Run this model inference with full control and performance in your environment.
Model Details
Model Provider
FraPiz
Model Tree
openai/whisper-large-v3-turbo
Input Modalities
Output Modalities
Supported Functionality
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 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.
python
import torchfrom transformers import AutoModelForSpeechSeq2Seq, AutoProcessor, pipelinemodel_id = "FraPiz/whisper-large-v3-turbo-moldovan-romanian"device = 0 if torch.cuda.is_available() else -1dtype = torch.float16 if torch.cuda.is_available() else torch.float32model = 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"])
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.
bibtex
@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}}
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.