|
--- |
|
license: apache-2.0 |
|
tags: |
|
- generated_from_trainer |
|
metrics: |
|
- wer |
|
model-index: |
|
- name: whisper-large-v2-kangri |
|
results: |
|
- task: |
|
type: automatic-speech-recognition |
|
name: Speech Recognition |
|
dataset: |
|
type: bridgeconn/snow-mountain |
|
name: snow-moutain-Kangri |
|
config: Kangri |
|
split: train_500 |
|
metrics: |
|
- type: wer |
|
value: 17.40 |
|
name: WER |
|
lower_is_better: true |
|
--- |
|
<!-- 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-large-v2-kangri |
|
|
|
This model is a fine-tuned version of [vasista22/whisper-hindi-large-v2](https://huggingface.co/vasista22/whisper-hindi-large-v2) on the [bridgeconn/snow-mountain](https://huggingface.co/datasets/bridgeconn/snow-mountain) dataset for the low resource Indian language- Kangri. |
|
It achieves the following results on the evaluation set: |
|
- Loss: 0.2967 |
|
- Wer: 0.1740 |
|
|
|
## Usage |
|
|
|
In order to evaluate this model on an entire dataset, the evaluation codes available in the whisper-finetune repository can be used. |
|
|
|
The same repository also provides the scripts for faster inference using whisper-jax. |
|
|
|
|
|
## Training and evaluation data |
|
|
|
Training Data: |
|
- [Snow Mountain Dataset for Kangri Language](https://huggingface.co/datasets/bridgeconn/snow-mountain) |
|
|
|
Evaluation Data: |
|
- [Snow Mountain Dataset for Kangri Language](https://huggingface.co/datasets/bridgeconn/snow-mountain) |
|
- [Kangri Translators Dataset ](https://drive.google.com/drive/folders/16BdOieekGRAo2bFOQDd4YhE2LpgiRnqQ?usp=share_link) |
|
|
|
## Training procedure |
|
|
|
We implemented Cross-Lingual Phoneme Recognition - a process that leverages patterns in resource-rich languages such as Hindi to recognize utterances in resource-poor languages |
|
such as Kangri. By fine-tuning a pre-trained model of the Whisper-Hindi-Large-V2 on a customised dataset - we have achieved SoTa accuracy. |
|
A customised dataset - consisting of the brigdeconn/snow-mountain and sentences collected from Kangri translators was created. This was then split using the 80/20 |
|
split rule. The results were evaluated with 5000 steps. The model decreases the word error rate by 0.6% after the initial 1000 steps. The Validation Loss increases due to |
|
more data being introduced. |
|
|
|
|
|
### Training hyperparameters |
|
|
|
The following hyperparameters were used during training: |
|
- learning_rate: 1e-05 |
|
- train_batch_size: 8 |
|
- eval_batch_size: 8 |
|
- seed: 42 |
|
- distributed_type: multi-GPU |
|
- num_devices: 2 |
|
- total_train_batch_size: 16 |
|
- total_eval_batch_size: 16 |
|
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08 |
|
- lr_scheduler_type: linear |
|
- lr_scheduler_warmup_steps: 500 |
|
- training_steps: 5000 |
|
- mixed_precision_training: Native AMP |
|
|
|
### Training results |
|
|
|
| Training Loss | Epoch | Step | Validation Loss | Wer | |
|
|:-------------:|:-----:|:----:|:---------------:|:------:| |
|
| 0.0001 | 40.0 | 1000 | 0.2442 | 0.1800 | |
|
| 0.0 | 80.0 | 2000 | 0.2752 | 0.1764 | |
|
| 0.0 | 120.0 | 3000 | 0.2870 | 0.1747 | |
|
| 0.0 | 160.0 | 4000 | 0.2940 | 0.1745 | |
|
| 0.0 | 200.0 | 5000 | 0.2967 | 0.1740 | |
|
|
|
|
|
### Framework versions |
|
|
|
- Transformers 4.28.0.dev0 |
|
- Pytorch 2.0.0+cu117 |
|
- Datasets 2.11.0 |
|
- Tokenizers 0.13.3 |
|
|