README.md

---
license: apache-2.0
base_model: openai/whisper-tiny
tags:
- generated_from_trainer
datasets:
- google/fleurs
metrics:
- wer
model-index:
- name: whisper-tiny-finetune-hindi-fleurs
  results:
  - task:
      name: Automatic Speech Recognition
      type: automatic-speech-recognition
    dataset:
      name: google/fleurs
      type: google/fleurs
      config: hi_in
      split: train+test
      args: hi_in
    metrics:
    - name: Wer
      type: wer
      value: 0.42621638924455824
language:
- hi
---

# whisper-tiny-finetune-hindi-fleurs

This model is a fine-tuned version of [openai/whisper-tiny](https://huggingface.co/openai/whisper-tiny) on the google/fleurs dataset.
It achieves the following results on the evaluation set:
- Loss: 0.8315
- Wer Ortho: 0.4313
- Wer: 0.4262

A working Hugging Face Space can be found [here](https://huggingface.co/spaces/Aryan-401/whisper-tiny-finetune-hindi)

## Model description

This model is a fine-tuned version of [openai/whisper-tiny](https://huggingface.co/openai/whisper-tiny) on the google/fleurs dataset. It improves the WER from 102.3 as stated in the [Whisper Paper](https://cdn.openai.com/papers/whisper.pdf) to 0.42 on the Hindi Subset of google/fleurs

## Intended uses & limitations

This model is intended to be used on Edge Low Compute Devices such as the Raspbery Pi Pico/3/3B/4 and offers real time transcription of Hindi audio into the English Lexicon.

## Training and evaluation data

The model was trained on `google/fleurs`'s `hi_in` Subset and used WER as the evaluation criteria

## Training procedure

### Training hyperparameters

The following hyperparameters were used during training:
- learning_rate: 1e-05
- train_batch_size: 32
- eval_batch_size: 32
- seed: 42
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
- lr_scheduler_type: constant_with_warmup
- lr_scheduler_warmup_steps: 50
- training_steps: 500
- mixed_precision_training: Native AMP

### Training results

| Training Loss | Epoch | Step | Validation Loss | Wer Ortho | Wer    |
|:-------------:|:-----:|:----:|:---------------:|:---------:|:------:|
| 1.8112        | 1.39  | 100  | 1.7274          | 0.6323    | 0.6258 |
| 1.0387        | 2.78  | 200  | 1.1194          | 0.5130    | 0.5072 |
| 0.7671        | 4.17  | 300  | 0.9671          | 0.4665    | 0.4613 |
| 0.5283        | 5.56  | 400  | 0.8840          | 0.4494    | 0.4440 |
| 0.4458        | 6.94  | 500  | 0.8315          | 0.4313    | 0.4262 |


### Framework versions

- Transformers 4.35.2
- Pytorch 2.1.0+cu121
- Datasets 2.16.0
- Tokenizers 0.15.0

## Citations

```
@inproceedings{Bhat:2014:ISS:2824864.2824872,
 author = {Bhat, Irshad Ahmad and Mujadia, Vandan and Tammewar, Aniruddha and Bhat, Riyaz Ahmad and Shrivastava, Manish},
 title = {IIIT-H System Submission for FIRE2014 Shared Task on Transliterated Search},
 booktitle = {Proceedings of the Forum for Information Retrieval Evaluation},
 series = {FIRE '14},
 year = {2015},
 isbn = {978-1-4503-3755-7},
 location = {Bangalore, India},
 pages = {48--53},
 numpages = {6},
 url = {http://doi.acm.org/10.1145/2824864.2824872},
 doi = {10.1145/2824864.2824872},
 acmid = {2824872},
 publisher = {ACM},
 address = {New York, NY, USA},
 keywords = {Information Retrieval, Language Identification, Language Modeling, Perplexity, Transliteration},
}
```
```
@misc{radford2022whisper,
  doi = {10.48550/ARXIV.2212.04356},
  url = {https://arxiv.org/abs/2212.04356},
  author = {Radford, Alec and Kim, Jong Wook and Xu, Tao and Brockman, Greg and McLeavey, Christine and Sutskever, Ilya},
  title = {Robust Speech Recognition via Large-Scale Weak Supervision},
  publisher = {arXiv},
  year = {2022},
  copyright = {arXiv.org perpetual, non-exclusive license}
}
```