Modelpublic

thunderboltc/gemini_whisper_ipa

sourceHugging Faceupdated 23d agoView on Hugging Face

0likes741downloads

Model Card

geminiwhisperipa

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

—perdevicetrainbatchsize: 8
—gradientaccumulationsteps: 2
—effectivebatchsize: 16
—learning_rate: 0.0001
—warmup_steps: 100
—numtrainepochs: 20
—gradient_checkpointing: True
—fp16: True
—eval_strategy: IntervalStrategy.EPOCH
—save_strategy: SaveStrategy.EPOCH
—savetotallimit: 2
—loadbestmodelatend: True
—metricforbest_model: wer
—predictwithgenerate: True
—generationmaxlength: 225