---
language: []
tags:
- whisper
- automatic-speech-recognition
- santali
- sanlish
- ipa
base_model: openai/whisper-small
---

# gemini_whisper_ipa

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

Last updated: 2026-08-29T23:56:15.284469Z — after epoch 20.0

## Dataset split (80/10/10)
- train: 1547
- validation: 193
- test: 194

## Latest evaluation metrics (epoch 20.0)
- **test_loss**: 0.4051657021045685
- **test_wer**: 40.782828282828284

## Training hyperparameters
- per_device_train_batch_size: 8
- gradient_accumulation_steps: 2
- effective_batch_size: 16
- learning_rate: 0.0001
- warmup_steps: 100
- num_train_epochs: 20
- 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