Dataset Builder

Dataset builder

The Dataset tab builds a LoRA training dataset one sample at a time, letting you listen to each sample before you commit.

When to use it

• You want to adjust voice descriptions and emotions until every sample sounds right.
• You want to build a dataset over several sessions.
• You want to hear how different emotion-and-text combinations change the voice.

It is also what the Training tab's Build New Dataset button opens.

Workflow

1. Create a project

Click New Dataset and enter a name. This creates a workspace in dataset_builder/ that is kept between sessions.

2. Describe the voice

• Root Voice Description — the voice's identity, for example "A warm, deep male narrator with a calm baritone". The Voice reference explains which words work.
• Global Seed — optional, for reproducible generation; leave it empty for random.

3. Add samples

Click Add Row for each sample. Each row has:

field	meaning
Text	what is spoken in this sample
Emotion	a style direction added to the voice description ("cheerful", "tense whisper", "barking orders")

The emotion is only a prompt for generating the audio with VoiceDesign; it is not stored as a training label. (If you are preparing real recordings instead, see the Training guide.)

Tips:

• Cover a range of emotions: neutral, happy, angry, sad, tense, commanding.
• Mix short exclamations ("Oh!", "Right.") with longer passages.
• Include at least one long neutral passage to use as the reference sample.
• The Training guide has more on what makes a good dataset.

4. Generate and listen

• Generate on a row — render just that sample.
• Generate Pending — render every sample that has no audio yet.
• Regen All — re-render every sample.
• Cancel — stop a running batch; finished samples are kept.

Listen to each sample and regenerate any that don't sound right, adjusting the emotion or the voice description as needed.

5. Save as a dataset

When every sample sounds right:

1. Choose the reference sample. It becomes ref.wav, the speaker embedding used in training — pick a clear, representative line, usually a long neutral passage.
2. Click Save as Training Dataset. The project is copied to lora_datasets/ and appears in the Training tab.

You can also Import / Export JSON to move a project's rows between machines.

Where projects are kept

• Each project lives in dataset_builder/{name}/ on the server.
• The rows, description and seed are saved in the project's state.json, so you can close the browser and come back later.
• Saving a dataset copies the data; the project stays, and you can keep working on it.

API

The API reference shows how to drive the Dataset Builder over HTTP.