|
--- |
|
language: |
|
- ug |
|
license: apache-2.0 |
|
datasets: |
|
- THUGY20 |
|
metrics: |
|
- cer |
|
model-index: |
|
- name: Whisper Small Fine-tuned with THUYG20 Uyghur Dataset |
|
results: |
|
- task: |
|
name: Automatic Speech Recognition |
|
type: automatic-speech-recognition |
|
dataset: |
|
name: 'THUGY20: A free Uyghur speech database' |
|
type: THUGY20 |
|
metrics: |
|
- name: Cer |
|
type: cer |
|
value: 7.21 |
|
--- |
|
# Uyghur Automatic Speech Recognition |
|
|
|
Uyghur ASR using CTC loss trained with THUYG20 dataset |
|
It achieves the following results on the evaluation set: |
|
- Best CER: 7.21% |
|
|
|
Reference: https://github.com/gheyret/uyghur-asr-ctc |
|
|
|
## Training procedure |
|
|
|
### Training hyperparameters |
|
|
|
The following hyperparameters were used during training: |
|
- learning_rate: 0.0001 |
|
- train_batch_size: 20 |
|
- eval_batch_size: 20 |
|
- seed: 42 |
|
- optimizer: Adam with weight_decay=0.000001 |
|
|
|
### Training results |
|
|
|
Best CER: 7.21% |
|
Trained: 473 epochs |
|
The model has 26,389,282 trainable parameters |
