Dataset Builder

Dataset Builder

The Dataset Builder tab provides an interactive workflow for creating LoRA training datasets with per-sample audio preview.

When to Use

Use the Dataset Builder when you want fine-grained control over each training sample before committing to a dataset. It's ideal for:

• Iterating on voice descriptions and emotions until each sample sounds right
• Building datasets incrementally over multiple sessions
• Previewing how different emotion+text combinations affect the generated voice

For quick dataset creation without per-sample preview, use the Generate Dataset form in the Training tab instead.

Workflow

1. Create a Project

Click New Project and enter a name. This creates a workspace in dataset_builder/ that persists across sessions.

2. Configure Voice

• Voice Description — The base voice identity description (e.g., "A warm, deep male narrator with a calm baritone")
• Global Seed — Optional seed for reproducible generation (leave empty for random)

3. Define Samples

Add rows to the sample table. Each row has:

Field	Description
Text	The spoken content for this sample
Emotion	Emotion/style direction appended to the voice description (e.g., "cheerful", "tense whisper", "barking orders"). This is a VoiceDesign generation prompt — it controls how the synthetic audio is produced, not a training label. If you're preparing real audio for ZIP upload instead, see Training Guide#Method 2: Upload Real Audio (ZIP).

Tips for good training data:

• Include 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 for the reference sample
• See Training Guide for detailed dataset guidance

4. Generate and Preview

• Generate single — Click the generate button on any row to preview that sample
• Generate All — Batch-generate all samples at once
• Cancel — Stop a running batch without losing completed samples

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

5. Save as Dataset

When satisfied with all samples:

1. Select a Reference Sample — This becomes ref.wav for speaker embedding during training. Choose a clear, representative line (typically a long neutral passage).
2. Click Save as Dataset — The project is exported to lora_datasets/ and appears in the Training tab.

Project Persistence

• Projects are saved to dataset_builder/{name}/ on the server
• The UI state (rows, description, seed) is stored in the project's state.json
• You can close the browser and return later — the project loads from the server
• Projects persist independently of training datasets: saving a dataset copies the data, it doesn't consume the project

API

See API Reference#Dataset Builder for programmatic access to all Dataset Builder endpoints.