|
--- |
|
language: |
|
- ur |
|
license: apache-2.0 |
|
library_name: transformers |
|
tags: |
|
- automatic-speech-recognition |
|
- robust-speech-event |
|
- hf-asr-leaderboard |
|
datasets: |
|
- mozilla-foundation/common_voice_8_0 |
|
metrics: |
|
- wer |
|
- cer |
|
pipeline_tag: automatic-speech-recognition |
|
base_model: Harveenchadha/vakyansh-wav2vec2-urdu-urm-60 |
|
model-index: |
|
- name: wav2vec2-urdu |
|
results: |
|
- task: |
|
type: automatic-speech-recognition |
|
name: Urdu Speech Recognition |
|
dataset: |
|
name: Common Voice (Urdu) |
|
type: common_voice |
|
metrics: |
|
- type: wer |
|
value: 57.47 |
|
name: WER |
|
config: load_metric("wer") |
|
verified: true |
|
- type: cer |
|
value: 32.68 |
|
name: CER |
|
config: load_metric("cer") |
|
verified: 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. --> |
|
|
|
# wav2vec2-large-xls-r-300m-Urdu |
|
|
|
This model is a fine-tuned version of [Harveenchadha/vakyansh-wav2vec2-urdu-urm-60](https://huggingface.co/Harveenchadha/vakyansh-wav2vec2-urdu-urm-60) on the common_voice dataset. |
|
It achieves the following results on the evaluation set: |
|
- Wer: 0.5747 |
|
- Cer: 0.3268 |
|
|
|
## Model description |
|
The training and valid dataset is 0.58 hours. It was hard to train any model on lower number of so I decided to take vakyansh-wav2vec2-urdu-urm-60 checkpoint and finetune the wav2vec2 model. |
|
|
|
## Training procedure |
|
Trained on Harveenchadha/vakyansh-wav2vec2-urdu-urm-60 due to lesser number of samples. |
|
|
|
### Training hyperparameters |
|
|
|
The following hyperparameters were used during training: |
|
- learning_rate: 0.0003 |
|
- train_batch_size: 64 |
|
- eval_batch_size: 8 |
|
- seed: 42 |
|
- gradient_accumulation_steps: 2 |
|
- total_train_batch_size: 128 |
|
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08 |
|
- lr_scheduler_type: linear |
|
- lr_scheduler_warmup_steps: 100 |
|
- num_epochs: 100 |
|
- mixed_precision_training: Native AMP |
|
|
|
### Training results |
|
|
|
| Training Loss | Epoch | Step | Validation Loss | Wer | Cer | |
|
|:-------------:|:-----:|:----:|:---------------:|:------:|:------:| |
|
| 4.3054 | 16.67 | 50 | 9.0055 | 0.8306 | 0.4869 | |
|
| 2.0629 | 33.33 | 100 | 9.5849 | 0.6061 | 0.3414 | |
|
| 0.8966 | 50.0 | 150 | 4.8686 | 0.6052 | 0.3426 | |
|
| 0.4197 | 66.67 | 200 | 12.3261 | 0.5817 | 0.3370 | |
|
| 0.294 | 83.33 | 250 | 11.9653 | 0.5712 | 0.3328 | |
|
| 0.2329 | 100.0 | 300 | 7.6846 | 0.5747 | 0.3268 | |
|
|
|
|
|
### Framework versions |
|
|
|
- Transformers 4.16.0.dev0 |
|
- Pytorch 1.10.1+cu102 |
|
- Datasets 1.17.1.dev0 |
|
- Tokenizers 0.11.0 |