iulik-pisik/all_data_model_tiny

Whisper Tiny - finetuned on weather and horoscope

This model is a fine-tuned version of openai/whisper-tiny on the Vreme ProTV and Horoscop Neti datasets. It achieves the following results on the evaluation set:

• Loss: 0.0053
• Wer: 17.14

Model description

This is a fine-tuned version of the Whisper Tiny model, specifically adapted for Romanian language Automatic Speech Recognition (ASR) in the domains of weather forecasts and horoscopes. The model has been trained on two custom datasets to improve its performance in transcribing Romanian speech in these specific contexts.

Training procedure

The model was fine-tuned using transfer learning techniques on the pre-trained Whisper Tiny model. Two custom datasets were used: audio recordings of weather forecasts and horoscopes in Romanian.

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: 3000
• mixed_precision_training: Native AMP

Training results

Epoch	Step	Validation Loss	WER
3.75	1000	0.1388	18.7298
7.69	2000	0.036	17.5637
11.63	3000	0.0089	17.3574
15.38	4000	0.0053	17.1421

Framework versions

• Transformers 4.39.2
• Pytorch 2.2.1+cu121
• Datasets 2.18.0
• Tokenizers 0.15.2