HiTZ/whisper-large-v2-eu

Whisper Large-V2 Basque

Model summary

Whisper Large-V2 Basque is an automatic speech recognition (ASR) model for Basque (eu) speech. It is fine-tuned from [openai/whisper-large-v2] on the Basque portion of Mozilla Common Voice 13.0, achieving a Word Error Rate (WER) of 11.34% on the Common Voice evaluation split.

This model provides state-of-the-art transcription quality for Basque speech, delivering improved accuracy over previous large Whisper variants while remaining suitable for offline and batch processing.

---

Model description

• Architecture: Transformer-based encoder–decoder (Whisper)
• Base model: openai/whisper-large-v2
• Language: Basque (eu)
• Task: Automatic Speech Recognition (ASR)
• Output: Text transcription in Basque
• Decoding: Autoregressive sequence-to-sequence decoding

Leveraging Whisper’s multilingual pretraining, this large-v2 model is fine-tuned on Basque speech data to provide highly accurate transcription for a low-resource language, appropriate for research, media, and archival use cases.

---

Intended use

Primary use cases

• High-quality transcription of Basque audio recordings
• Offline or batch ASR pipelines
• Research and development in Basque ASR
• Media, educational, and archival transcription tasks

Intended users

• Researchers working on Basque or low-resource ASR
• Developers building Basque speech applications
• Academic and institutional users

Out-of-scope use

• Real-time or low-latency ASR without optimization
• Speech translation tasks
• Safety-critical applications without validation

---

Limitations and known issues

• Performance may degrade on:
  • Noisy or low-quality recordings
  • Conversational or spontaneous speech
  • Accents underrepresented in Common Voice
• While highly accurate, transcription errors may still occur under challenging acoustic conditions
• Dataset biases from Common Voice may be reflected in outputs

Users are encouraged to evaluate the model on their own data before deployment.

---

Training and evaluation data

Training data

• Dataset: Mozilla Common Voice 13.0 (Basque subset)
• Data type: Crowd-sourced, read speech
• Preprocessing:
  • Audio resampled to 16 kHz
  • Text normalized using Whisper tokenizer
  • Filtering of invalid or problematic samples

Evaluation data

• Dataset: Mozilla Common Voice 13.0 (Basque evaluation split)
• Metric: Word Error Rate (WER)

---

Evaluation results

Metric	Value
WER (eval)	11.34%

These results demonstrate state-of-the-art transcription performance for Basque ASR using a large-v2 Whisper model.

---

Training procedure

Training hyperparameters

• Learning rate: 1e-5
• Optimizer: Adam (β1=0.9, β2=0.999, ε=1e-8)
• LR scheduler: Linear
• Warmup steps: 500
• Training steps: 20,000
• Train batch size: 32
• Gradient accumulation steps: 2
• Total effective batch size: 64
• Evaluation batch size: 16
• Seed: 42

Training results (summary)

Training Loss	Epoch	Step	Validation Loss	WER
0.0355	4.01	1000	0.2616	14.8224
0.0079	9.01	2000	0.2777	13.5202
0.0041	14.01	3000	0.2764	12.7364
0.0047	19.0	4000	0.2932	12.6939
0.004	24.0	5000	0.2969	12.7992
0.0019	29.0	6000	0.3066	12.6008
0.004	33.01	7000	0.2973	12.6696
0.0007	38.01	8000	0.3253	12.2686
0.0006	43.01	9000	0.3391	12.5319
0.0009	48.01	10000	0.3303	12.2767
0.0004	53.0	11000	0.3383	12.0195
0.0003	58.0	12000	0.3398	11.7441
0.0005	63.0	13000	0.3396	11.8778
0.0001	67.01	14000	0.3544	11.6469
0.0	72.01	15000	0.3752	11.4160
0.0	77.01	16000	0.3860	11.3411
0.0	82.01	17000	0.3943	11.3391
0.0	87.0	18000	0.4013	11.3532
0.0	92.0	19000	0.4063	11.3613
0.0	97.0	20000	0.4086	11.3512

---

Framework versions

• Transformers 4.33.0.dev0
• PyTorch 2.0.1+cu117
• Datasets 2.14.4
• Tokenizers 0.13.3

---

How to use

from transformers import pipeline

hf_model = "HiTZ/whisper-large-v2-eu"  # replace with actual repo ID
device = 0  # set to -1 for CPU

pipe = pipeline(
    task="automatic-speech-recognition",
    model=hf_model,
    device=device
)

result = pipe("audio.wav")
print(result["text"])

---

Ethical considerations and risks

• This model transcribes speech and may process personal data.
• Users should ensure compliance with applicable data protection laws (e.g., GDPR).
• The model should not be used for surveillance or non-consensual audio processing.

---

Citation

If you use this model in your research, please cite:

@misc{dezuazo2025whisperlmimprovingasrmodels,
  title={Whisper-LM: Improving ASR Models with Language Models for Low-Resource Languages},
  author={Xabier de Zuazo and Eva Navas and Ibon Saratxaga and Inma Hernáez Rioja},
  year={2025},
  eprint={2503.23542},
  archivePrefix={arXiv},
  primaryClass={cs.CL}
}

Please, check the related paper preprint in arXiv:2503.23542 for more details.

---

License

This model is available under the Apache-2.0 License. You are free to use, modify, and distribute this model as long as you credit the original creators.

---

Contact and attribution

• Fine-tuning and evaluation: HiTZ/Aholab - Basque Center for Language Technology
• Base model: OpenAI Whisper
• Dataset: Mozilla Common Voice

For questions or issues, please open an issue in the model repository.

Funding

This project with reference 2022/TL22/00215335 has been parcially funded by the Ministerio de Transformación Digital and by the Plan de Recuperación, Transformación y Resiliencia – Funded by the European Union – NextGenerationEU ILENIA and by the project IkerGaitu funded by the Basque Government. This model was trained at Hyperion, one of the high-performance computing (HPC) systems hosted by the DIPC Supercomputing Center.