|
--- |
|
library_name: transformers |
|
license: mit |
|
base_model: openai/whisper-large-v3-turbo |
|
tags: |
|
- generated_from_trainer |
|
metrics: |
|
- wer |
|
model-index: |
|
- name: whisper-large-v3-turbo-FLEURS-GL |
|
results: [] |
|
datasets: |
|
- juanjucm/FLEURS-SpeechT-GL-EN |
|
language: |
|
- gl |
|
--- |
|
|
|
# whisper-large-v3-turbo-FLEURS-GL |
|
|
|
This model is a fine-tuned version of [openai/whisper-large-v3-turbo](https://huggingface.co/openai/whisper-large-v3-turbo) trained on [juanjucm/FLEURS-SpeechT-GL-EN](https://huggingface.co/datasets/juanjucm/FLEURS-SpeechT-GL-EN) for **Galician Text to Speech** task. It takes galician speech audios as input and generates the correspondant transcription. |
|
|
|
This Automatic Speech Recognition model, was developed to be the first stage of a Speech Translation cascade system for transcribing and translating Galician audios into English texts. After this first STT step, this [Galician-to-English MT model](https://huggingface.co/juanjucm/nllb-200-distilled-600M-FLEURS-GL-EN) can be applied over the generated Galician transcriptions to get English text translations. |
|
|
|
The motivation behind this work is to increase the visibility of the Galician language, making it more accessible for non-Galician speakers to understand and engage with Galician audio content. |
|
|
|
This model was developed during a 3-week Speech Translation workshop organised by [Yasmin Moslem](https://huggingface.co/ymoslem). |
|
|
|
### Performance and training details |
|
|
|
Baseline model achieved a WER score of **16.6** on the evaluation dataset. |
|
|
|
It achieves the following results on the evaluation set: |
|
- Loss: 0.2522 |
|
- **WER: 9.1731** |
|
|
|
The following hyperparameters were used during training: |
|
- learning_rate: 5e-06 |
|
- train_batch_size: 16 |
|
- eval_batch_size: 8 |
|
- seed: 42 |
|
- distributed_type: multi-GPU |
|
- num_devices: 2 |
|
- total_train_batch_size: 32 |
|
- total_eval_batch_size: 16 |
|
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08 |
|
- lr_scheduler_type: linear |
|
- num_epochs: 10 |
|
- mixed_precision_training: Native AMP |
|
|
|
### Training results |
|
|
|
We used [WER (Word Error Rate)](https://en.wikipedia.org/wiki/Word_error_rate) as our reference transcription metric for selecting the best checkpoint after training. |
|
|
|
| Training Loss | Epoch | Step | Validation Loss | Wer | |
|
|:-------------:|:-----:|:----:|:---------------:|:-------:| |
|
| 0.125 | 1.0 | 86 | 0.2128 | 10.1464 | |
|
| 0.0835 | 2.0 | 172 | 0.2006 | 9.4315 | |
|
| 0.0322 | 3.0 | 258 | 0.2091 | 9.6985 | |
|
| 0.0115 | 4.0 | 344 | 0.2325 | 9.8880 | |
|
| 0.0127 | 5.0 | 430 | 0.2313 | 9.2506 | |
|
| 0.0022 | 7.0 | 602 | 0.2498 | 9.2679 | |
|
| 0.0021 | 6.0 | 516 | 0.2412 | 9.3885 | |
|
| 0.0011 | 8.0 | 688 | 0.2522 | 9.1731 | |
|
| 0.0017 | 9.0 | 774 | 0.2538 | 9.3023 | |
|
| 0.0011 | 10.0 | 860 | 0.2556 | 9.2937 | |
|
|
|
|
|
### Framework versions |
|
|
|
- Transformers 4.45.1 |
|
- Pytorch 2.4.1+cu121 |
|
- Datasets 3.0.1 |
|
- Tokenizers 0.20.0 |
