Whisper Medium TW - Augmented
This model is a fine-tuned version of openai/whisper-medium on the mozilla-foundation/common_voice_11_0 dataset. It achieves the following results on the evaluation set:
- eval_loss: 0.0951
- eval_wer: 7.4865
- eval_runtime: 2823.6824
- eval_samples_per_second: 1.668
- eval_steps_per_second: 0.834
- epoch: 1.7
- step: 600
Model description
More information needed
Intended uses & limitations
More information needed
Training and evaluation data
Training:
- mozilla-foundation/common_voice_11_0 (train+validation)
Evaluation:
Training procedure
- Datasets were augmented on-the-fly using audiomentations via PitchShift and TimeStretch transformations at
p=0.3
. - A space is added between each Chinese character, as demonstrated in the original paper. Effectively, WER == CER in this case.
Training hyperparameters
The following hyperparameters were used during training:
- learning_rate: 1e-05
- train_batch_size: 2
- eval_batch_size: 2
- seed: 42
- gradient_accumulation_steps: 16
- total_train_batch_size: 32
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
- lr_scheduler_type: linear
- lr_scheduler_warmup_steps: 100
- training_steps: 1000
- mixed_precision_training: Native AMP
Framework versions
- Transformers 4.26.0.dev0
- Pytorch 1.13.0+cu117
- Datasets 2.7.1.dev0
- Tokenizers 0.13.2
- Downloads last month
- 4
Inference Providers
NEW
This model is not currently available via any of the supported third-party Inference Providers, and
the model is not deployed on the HF Inference API.
Dataset used to train Scrya/whisper-medium-zh-TW-augmented
Evaluation results
- WER on mozilla-foundation/common_voice_11_0test set self-reported7.486