README.md

---
language:
- uz
license: apache-2.0
tags:
- whisper-event
- generated_from_trainer
datasets:
- mozilla-foundation/common_voice_11_0
- google/fleurs
metrics:
- wer
model-index:
- name: Whisper Small Uzbek
  results:
  - task:
      type: automatic-speech-recognition
      name: Automatic Speech Recognition
    dataset:
      name: mozilla-foundation/common_voice_11_0
      type: mozilla-foundation/common_voice_11_0
      config: uz
      split: test
      args: da
    metrics:
    - type: wer
      value: 23.650914047642605
      name: Wer
  - task:
      type: automatic-speech-recognition
      name: Automatic Speech Recognition
    dataset:
      name: google/fleurs
      type: google/fleurs
      config: uz_uz
      split: test
    metrics:
    - type: wer
      value: 47.15
      name: WER
---

<!-- Disclaimer: I've never written a model card before. I'm probably not correctly following standard practices on how they should be written. 
     I'm new to this. I'm sorry -->

# Whisper Small Uzbek

This model is a fine-tuned version of [openai/whisper-small](https://huggingface.co/openai/whisper-small) trained and evaluated on the mozilla-foundation/common_voice_11_0 uz and google/fleurs uz_uz datasets.

It achieves the following results on the common_voice_11_0 evaluation set:
- Loss: 0.3872
- Wer: 23.6509

It achieves the following results on the FLEURS evaluation set:
- Wer: 47.15

## Model description

This model was created as part of the Whisper fine-tune sprint event. 

Based on eval, this model achieves a WER of 23.6509 against the Common Voice 11 dataset and 47.15 against the FLEURS dataset.

This is a significant improvement over the smallest reported WER of 90.2 for the Uzbek language recorded on the [Whisper article](https://cdn.openai.com/papers/whisper.pdf):

![A part of Table 13 from the paper "Robust Speech Recognition via Large-Scale Weak Supervision", which shows the WER achieved by the Whisper model under the FLEURS dataset. Highlighted is the best score it achieved under for the Uzbek language, which was 90.2.](https://huggingface.co/BlueRaccoon/whisper-small-uz/resolve/main/uzbektable13.png)

## Intended uses & limitations

More information needed

## Training and evaluation data

Training was performed using the train and evaluation splits from [Mozilla's Common Voice 11](https://huggingface.co/mozilla-foundation/common_voice_11_0) and [Google's FLEURS](https://huggingface.co/google/fleurs) datasets.

Testing was performed using the test splits from the same datasets.

## Training procedure

Training and CV11 testing was performed using a modified version of Hugging Face's [run_speech_recognition_seq2seq_streaming.py](https://github.com/kamfonas/whisper-fine-tuning-event/blob/e0377f55004667f18b37215d11bf0e54f5bda463/run_speech_recognition_seq2seq_streaming.py) script by Michael Kamfonas.

FLEURS testing was performed using the standard [run_eval_whisper_streaming.py](https://github.com/huggingface/community-events/blob/main/whisper-fine-tuning-event/run_eval_whisper_streaming.py) script.

### Training hyperparameters

The following hyperparameters were used during training:
- learning_rate: 1e-05
- train_batch_size: 64
- eval_batch_size: 32
- seed: 42
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
- lr_scheduler_type: linear
- lr_scheduler_warmup_steps: 400
- training_steps: 5000
- mixed_precision_training: Native AMP

### Training results

| Training Loss | Epoch | Step | Validation Loss | Wer     |
|:-------------:|:-----:|:----:|:---------------:|:-------:|
| 0.1542        | 0.2   | 1000 | 0.4711          | 30.8413 |
| 0.0976        | 0.4   | 2000 | 0.4040          | 26.6464 |
| 0.1088        | 1.0   | 3000 | 0.3765          | 24.4952 |
| 0.0527        | 1.21  | 4000 | 0.3872          | 23.6509 |
| 0.0534        | 1.41  | 5000 | 0.3843          | 23.6817 |


### Framework versions

- Transformers 4.26.0.dev0
- Pytorch 1.13.0+cu117
- Datasets 2.7.1.dev0
- Tokenizers 0.13.2