hosszu/whisper-small-hu-SzL

Whisper Small Hungarian (SzL_Dataset)

This model is a fine-tuned version of openai/whisper-small on the Hungarian SzL_Dataset.

All the Python scripts applied at creating and checking the dataset and performing the training have been created with the help of Gemini, the Google's AI assistant.

Model Details

• Language: Hungarian
• Task: Transcribe
• Base Model: openai/whisper-small
• Training Steps: 5000

Evaluation Results

• Best WER: 118.56 (at step 2000)
• Final WER: 128.10 (at step 5000)

Training Hyperparameters

• learning_rate: 1e-05
• train_batch_size: 8
• eval_batch_size: 8
• gradient_accumulation_steps: 2
• fp16: True