GIanlucaRub/whisper-tiny-it-2

Whisper Tiny It 2 - Gianluca Ruberto

This model is a fine-tuned version of openai/whisper-tiny on the Common Voice 11.0 dataset. It achieves the following results on the evaluation set:

• Loss: 0.711485
• Wer: 43.392956

Model description

This model is the openai whisper small transformer adapted for Italian audio to text transcription. This model has weight decay set to 0.3 to cope with overfitting.

Intended uses & limitations

The model is available through its HuggingFace web app

Training and evaluation data

Data used for training is the initial 10% of train and validation of Italian Common Voice 11.0 from Mozilla Foundation. The dataset used for evaluation is the initial 10% of test of Italian Common Voice. Unfortunately weight decay showed to have slightly worse result also on the evaluation dataset.

Training procedure

After loading the pre trained model, it has been trained on the dataset.

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 1e-05
• train_batch_size: 16
• eval_batch_size: 8
• 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: 4000
• mixed_precision_training: Native AMP
• weight_decay: 0.3

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
0.5837	0.95	1000	0.790046	50.6032
0.4186	1.91	2000	0.730115	46.0067
0.3154	2.86	3000	0.712776	44.114
0.2676	3.82	4000	0.711485	43.393

Framework versions

• Transformers 4.26.0.dev0
• Pytorch 1.12.1+cu113
• Datasets 2.7.1
• Tokenizers 0.13.2