gemini_whisper_ipa

Fine-tuned openai/whisper-small for Santali IPA/Sanlish transliteration (speech-to-text).

Last updated: 2026-08-31T09:16:04.453718Z — after epoch 13.0

Dataset split (80/10/10)

  • train: 3094
  • validation: 193
  • test: 193

Latest evaluation metrics (epoch 13.0)

  • eval_loss: 0.8340893387794495
  • eval_wer: 36.113837095191364
  • eval_cer: 9.316446911866759
  • eval_runtime: 83.7586
  • eval_samples_per_second: 2.304
  • eval_steps_per_second: 0.298
  • epoch: 13.0

Training hyperparameters

  • per_device_train_batch_size: 8
  • gradient_accumulation_steps: 2
  • effective_batch_size: 16
  • learning_rate: 2e-05
  • warmup_steps: 150
  • num_train_epochs: 15
  • gradient_checkpointing: True
  • fp16: True
  • eval_strategy: IntervalStrategy.EPOCH
  • save_strategy: SaveStrategy.EPOCH
  • save_total_limit: 2
  • load_best_model_at_end: True
  • metric_for_best_model: wer
  • predict_with_generate: True
  • generation_max_length: 225
Downloads last month
-
Safetensors
Model size
0.2B params
Tensor type
F32
·
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for thunderboltc/whisper_sanlish_augmented_test1

Finetuned
(3746)
this model