Speech-to-text datasets

A dataset is a set of recordings with the words spoken in each one. You can run many evaluations on the same dataset.

See your datasets

In the sidebar, click Speech-to-Text, then click the Datasets tab. Each row shows how many recordings the dataset has (Items) and how many evaluations used it (Evals).

Create a dataset

1. On the Datasets tab, click New dataset.
2. Under Dataset name, type a name, for example Tutor questions in Hindi, and click Create.
3. On the dataset page, click Upload .wav and type the words spoken in the box next to it.
4. Click Add another sample for each other recording. To add many at once, click Choose ZIP file in the Upload ZIP box.
5. Click Save.

Each recording must be a .wav file of under 60 seconds and under 5 MB. A dataset is also created when you upload recordings on the Dataset tab of a new evaluation.

Change a dataset

Click a dataset on the Datasets tab to open it. On the dataset page you can:

• Add recordings with Add another sample or Choose ZIP file.
• Fix the words for a recording by typing in its box.
• Remove a recording with the bin button on its row. The button shows once your other changes are saved, and only while the dataset has more than one recording.
• Rename the dataset with the pencil button next to its name.

Click Save to keep your changes. To change the audio of a recording, remove that row and add the recording again.

Run an evaluation on a dataset

On the dataset page, click New evaluation. The dataset is already chosen, so you only pick the language and models. Save any changes first.

Delete a dataset

On the Datasets tab, click the bin button at the end of the dataset’s row, then click Remove. Evaluations that used it keep their results.

Next steps