Whisper Large v3 Basque

This model is a fine-tuned version of openai/whisper-large-v3 specifically for Basque (eu) language Automatic Speech Recognition (ASR). It was trained on the asierhv/composite_corpus_eu_v2.1 dataset, which is a composite corpus designed to improve Basque ASR performance.

Key improvements and results compared to the base model:

• Significant WER reduction: The fine-tuned model achieves a Word Error Rate (WER) of 6.5443 on the validation set of the asierhv/composite_corpus_eu_v2.1 dataset, demonstrating a substantial improvement in accuracy for Basque speech.
• Exceptional performance on Common Voice: When evaluated on the Mozilla Common Voice 18.0 dataset, the model achieved a WER of 4.84. This showcases the model's outstanding ability to generalize to diverse Basque speech datasets, and highlights the high accuracy achievable with the large-v3 model.

Model description

This model leverages the whisper-large-v3 architecture, the most powerful variant of the Whisper models, known for its exceptional accuracy in multilingual speech recognition. By fine-tuning this model on a dedicated Basque speech corpus, it achieves state-of-the-art performance in Basque ASR. The whisper-large-v3 model offers the highest capacity and therefore the highest accuracy, but requires significantly more computational resources.

Intended uses & limitations

Intended uses:

• Ultra-high-accuracy automatic transcription of Basque speech for critical applications.
• Development of cutting-edge Basque speech-based applications demanding the highest possible precision.
• Research in Basque speech processing requiring the most accurate transcriptions.
• Professional transcription services and applications where accuracy is paramount and computational resources are available.
• Use in scenarios where the highest possible accuracy is required, and the computational cost is justifiable.

Limitations:

• Performance is still influenced by audio quality, with challenges arising from background noise and poor recording conditions.
• Accuracy may be affected by highly dialectal or informal Basque speech, although the large model mitigates this to a great degree.
• Despite its high performance, the model may still produce errors, particularly with complex linguistic structures or rare words.
• The large-v3 model demands substantial computational resources, making it less suitable for real-time or resource-constrained applications.

Training and evaluation data

• Training dataset: asierhv/composite_corpus_eu_v2.1. This dataset is a comprehensive and meticulously curated collection of Basque speech data, designed to maximize the performance of Basque ASR systems.
• Evaluation Dataset: The test split of asierhv/composite_corpus_eu_v2.1.

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 4.375e-06
• train_batch_size: 16
• eval_batch_size: 8
• seed: 42
• optimizer: Use OptimizerNames.ADAMW_TORCH with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 1000
• training_steps: 20000
• mixed_precision_training: Native AMP

Training results

Framework versions

• Transformers 4.49.0.dev0
• Pytorch 2.6.0+cu124
• Datasets 3.3.1.dev0
• Tokenizers 0.21.0