Whisper Small Ko

🚀 whisper-small-ko

This model is a fine - tuned version of Whisper Small on the following AI hub dataset. Due to the large size of the dataset, the data was randomly shuffled and divided into 5 parts for training.

🚀 Quick Start

This section provides a brief overview of the model and its training process.

✨ Features

• The model is trained on multiple datasets from AI hub simultaneously, which helps to improve its generalization ability across different domains.
• Future use of adapters may lead to better performance on dialects, children's voices, and the elderly's voices.

📚 Documentation

Training results

Dataset

The model is characterized by training on many datasets from AI hub at once. ASR has a high dependency on the domain. Therefore, even if it is trained on one dataset, its performance will drop significantly when tested on other datasets. To avoid this, the model was trained on as many datasets as possible at once. In the future, using adapters for dialects, children's voices, and the elderly's voices may yield good performance.

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 1e - 05
• train_batch_size: 32
• eval_batch_size: 16
• gradient_accumulation_steps: 2
• warmup_ratio: 0.01
• num_train_epoch: 1

📄 License

This project is licensed under the Apache - 2.0 license.