Fine-tuning speech-to-text

In this scenario, we fine-tune Whisper, a model for speech-to-text recognition.

Requirements

You'll need a HuggingFace token that has access to the voice library we will use. Enable your token to access this repository.

Create a workspace on Coder for Jupyter. When the workspace is up, access Jupyter and create a new notebook.

Set-up

Import the platform's library and create a project:

import digitalhub as dh

project = dh.get_or_create_project("whisper-fine-tuning")

Create a secret as follows, make sure you replace the value with the token that has access to the aforementioned repository:

project.new_secret(name="HF_TOKEN", secret_value="my-token")

The functions we will run use code from a Python file. Due to the many lines this file has, instead of presenting it in this documentation, we invite you to download it from the Whisper fine-tuning scenario in the tutorials repository.

Use the src directory and ensure it is at the same level of the notebook you're using.