adriszmar's picture
Update
b32b704 verified
metadata
library_name: transformers
language:
  - es
license: mit
base_model: openai/whisper-large-v3-turbo
tags:
  - whisper-event
  - generated_from_trainer
datasets:
  - mozilla-foundation/common_voice_17_0
model-index:
  - name: Whisper Large V3 Turbo - Spanish
    results: []

Whisper Large V3 Turbo - Spanish

This model is a fine-tuned version of openai/whisper-large-v3-turbo on the Common Voice 17.0 dataset - spanish subset.

The fine-tuning process reduced the Word Error Rate (WER) from 6.91% to 5.34%, demonstrating significant improvement in transcription accuracy for spanish audios.

Model description

More information needed

Intended uses & limitations

More information needed

Training and evaluation data

The model was trained using the Common Voice 17.0 dataset - spanish subset (mozilla-foundation/common_voice_17_0). Both the base model, whisper-large-v3-turbo, and the fine-tuned model, whisper-large-v3-turbo-es, were evaluated using Word Error Rate (WER) on the test split of the same dataset. The results are as follows:

  • WER for whisper-large-v3-turbo (base): 6.91%
  • WER for whisper-large-v3-turbo-es (fine-tuned): 5.34%

This significant reduction in WER shows that fine-tuning the model for spanish audio led to improved transcription accuracy compared to the original base model.

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

  • learning_rate: 1e-05
  • train_batch_size: 64
  • eval_batch_size: 8
  • seed: 42
  • optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
  • lr_scheduler_type: linear
  • lr_scheduler_warmup_steps: 500
  • training_steps: 5000
  • mixed_precision_training: Native AMP

Framework versions

  • Transformers 4.44.2
  • Pytorch 2.4.1+cu121
  • Tokenizers 0.19.1