README.md

---
license: apache-2.0
base_model: openai/whisper-small
tags:
- generated_from_trainer
metrics:
- wer
model-index:
- name: Whisper-Anuj-small-Malyalam-final
  results: []
---

<!-- This model card has been generated automatically according to the information the Trainer had access to. You
should probably proofread and complete it, then remove this comment. -->

# Whisper-Anuj-small-Malyalam-final

This model is a fine-tuned version of [openai/whisper-small](https://huggingface.co/openai/whisper-small) on the None dataset.
It achieves the following results on the evaluation set:
- Loss: 0.1640
- Wer: 45.0607
- Cer: 9.4661

## Model description

More information needed

## Intended uses & limitations

More information needed

## Training and evaluation data

More information needed

## Training procedure

### Training hyperparameters

The following hyperparameters were used during training:
- learning_rate: 1e-05
- train_batch_size: 6
- eval_batch_size: 4
- seed: 42
- gradient_accumulation_steps: 4
- total_train_batch_size: 24
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
- lr_scheduler_type: linear
- lr_scheduler_warmup_steps: 500
- training_steps: 1800

### Training results

| Training Loss | Epoch   | Step | Validation Loss | Wer     | Cer     |
|:-------------:|:-------:|:----:|:---------------:|:-------:|:-------:|
| 0.0563        | 4.3243  | 600  | 0.1279          | 55.3846 | 12.5049 |
| 0.006         | 8.6486  | 1200 | 0.1527          | 48.6640 | 10.2313 |
| 0.0004        | 12.9730 | 1800 | 0.1640          | 45.0607 | 9.4661  |


### Framework versions

- Transformers 4.42.4
- Pytorch 2.3.0+cu121
- Datasets 2.20.0
- Tokenizers 0.19.1