tauri19/whisper-tiny-minds14-en-us

Whisper Tiny — Audio Course MINDS-14 en-US exercise

Actual fine-tuning of openai/whisper-tiny for the Hugging Face Audio Course Unit 5 assignment.

Training and evaluation data

The dataset is PolyAI/minds14, configuration en-US, original train order. The first 450 examples were used for training and the remaining 113 were held out for evaluation. Preprocessing used num_proc=1; audio was resampled to 16 kHz. No evaluation examples were used for gradient updates.

Results

• Normalized WER: 0.3040141676505313 (30.40%).
• Orthographic WER: 0.3362122146822949 (33.62%).
• Evaluation loss: 1.2425181865692139.

WER values in the model-index are fractions, as required by the assignment. Text normalization uses Whisper's BasicTextNormalizer; WER is computed with JIWER. The required normalized WER threshold is below 0.37.

Intended uses and limitations

Educational speech recognition practice on banking queries. This small dataset evaluation does not demonstrate general-purpose, production, or multilingual performance. The artifact has been evaluated locally; the course provider has not yet assessed the uploaded model.

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 1e-05
• train_batch_size: 4
• eval_batch_size: 4
• seed: 42
• gradient_accumulation_steps: 2
• total_train_batch_size: 8
• optimizer: Use OptimizerNames.ADAMW_TORCH_FUSED 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: 10
• training_steps: 180

Training results

Framework versions

• Transformers 4.57.6
• Pytorch 2.14.1
• Datasets 3.6.0
• Tokenizers 0.22.2