hriteshMaikap/gemma-3n-odia-asr

Fine-tuned Gemma 3N for Odia Automatic Speech Recognition (ASR)

This repository contains a Gemma 3N model fine-tuned on the IndicVoices Odia dataset for Automatic Speech Recognition (ASR).

Model Details

• Base Model: unsloth/gemma-3n-E4B-it
• Fine-tuning Method: LoRA
• Language: Odia
• Dataset: ai4bharat/IndicVoices (Odia split)
• Training Data Size: 3000 samples
• Validation Data Size: 1000 samples

Fine-tuning Summary

The model was fine-tuned for 1 epoch with a batch size of 2 and a learning rate of 5e-5.

• Training Runtime: 58.06 minutes
• Peak Reserved Memory: 12.592 GB
• Peak Reserved Memory for Training: 85.422 %

Evaluation

The fine-tuned model was evaluated on a subset of the validation set using the Word Error Rate (WER) metric.

• Final WER: 1.2887

Outcome

Initially the model was not so good with Odia input (text and audio), but now can generate texts in odia for conversation but audio transcription is still very basic with limited to no accuracy due to high WER.

How to Use

You can load and use this model for Odia ASR tasks using the transformers library.