Uploading fine-tuning data for a speech recognition model

To upload fine-tuning data for a speech recognition model, collect them into an archive and send the archive to the SpeechKit team.

Getting started

1. Prepare TSV files with text templates and glossaries.
2. Make sure that they meet the requirements; otherwise, they will not pass the check and you will not be able to collect an archive.
3. Create a community in Yandex DataSphere. This is where you will work from.
4. Link a billing account to the community.

Preparing an archive with fine-tuning data

1. Create a project on the DataSphere home page.

2. Open the notebook containing the data check function. If the data meets the requirements, it will be packed into an archive you will need to forward to the SpeechKit team to fine-tune the model.

3. Click Run in Yandex DataSphere.

4. Select the project you created earlier and click Add.

   The notebook will open in JupyterLab in the selected project.

5. Drag files with text templates and glossaries to the project directory in JupyterLab.

6. In the notebook, invoke the data check function. This will form an archive to be loaded to SpeechKit.

   To invoke the function, select the following cell in the notebook and click :

   prepare_stt_templates(
      templates_path="<template_file_name>.tsv",
      variables_path="<glossary_file_name>",
      output_path="<output_file_name>_tar.gz",
   )
   

7. Download the archive. To do this, right-click it and select Download.

8. Contact support and provide the archive to it.