Whisper Medium ID - FLEURS-CV-LBV - Augmented
This model is a fine-tuned version of openai/whisper-medium on the following datasets:
It achieves the following results on the evaluation set (Common Voice 11.0):
- Loss: 0.2788
- Wer: 7.6132
- Cer: 2.3332
Model description
More information needed
Intended uses & limitations
More information needed
Training and evaluation data
Training:
- mozilla-foundation/common_voice_11_0 (train+validation)
- google/fleurs (train+validation)
- indonesian-nlp/librivox-indonesia (train)
Evaluation:
- mozilla-foundation/common_voice_11_0 (test)
- google/fleurs (test)
- indonesian-nlp/librivox-indonesia (test)
Training procedure
Datasets were augmented on-the-fly using audiomentations via PitchShift, AddGaussianNoise and TimeStretch transformations at p=0.3
.
Training hyperparameters
The following hyperparameters were used during training:
- learning_rate: 1e-05
- train_batch_size: 32
- eval_batch_size: 16
- seed: 42
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
- lr_scheduler_type: linear
- lr_scheduler_warmup_steps: 500
- training_steps: 10000
- mixed_precision_training: Native AMP
Training results
Training Loss | Epoch | Step | Validation Loss | Wer | Cer |
---|---|---|---|---|---|
0.3002 | 1.9 | 1000 | 0.1659 | 8.1850 | 2.5333 |
0.0514 | 3.8 | 2000 | 0.1818 | 8.0559 | 2.5244 |
0.0145 | 5.7 | 3000 | 0.2150 | 7.8945 | 2.5281 |
0.0037 | 7.6 | 4000 | 0.2248 | 7.7100 | 2.3738 |
0.0016 | 9.51 | 5000 | 0.2402 | 7.6224 | 2.3591 |
0.0009 | 11.41 | 6000 | 0.2525 | 7.7654 | 2.3952 |
0.0005 | 13.31 | 7000 | 0.2609 | 7.5994 | 2.3487 |
0.0008 | 15.21 | 8000 | 0.2682 | 7.5855 | 2.3347 |
0.0002 | 17.11 | 9000 | 0.2756 | 7.6178 | 2.3288 |
0.0002 | 19.01 | 10000 | 0.2788 | 7.6132 | 2.3332 |
Framework versions
- Transformers 4.26.0.dev0
- Pytorch 1.13.1+cu117
- Datasets 2.7.1.dev0
- Tokenizers 0.13.2
- Downloads last month
- 5
This model does not have enough activity to be deployed to Inference API (serverless) yet. Increase its social
visibility and check back later, or deploy to Inference Endpoints (dedicated)
instead.
Datasets used to train Scrya/whisper-medium-id-augmented
Evaluation results
- WER on google/fleurstest set self-reported7.170
- CER on google/fleurstest set self-reported2.390
- WER on mozilla-foundation/common_voice_11_0test set self-reported7.590
- CER on mozilla-foundation/common_voice_11_0test set self-reported2.330
- WER on indonesian-nlp/librivox-indonesiatest set self-reported6.070
- CER on indonesian-nlp/librivox-indonesiatest set self-reported1.840