Dataset Maker

Multi-purpose dataset maker for various TTS models:

• Tortoise TTS/XTTS
• StyleTTS 2 ~ Webui
• Higgs Audio ~ Base - My fork
• VibeVoice ~ Base - My fork
• IndexTTS 2 ~ My Trainer

Main entrypoint

The primary script is gradio_interface.py. It starts the Gradio UI that orchestrates project creation, audio uploads, transcription, and dataset export.

Quick start (Windows)

1. Make sure you have Astral uv installed.
2. Run:

   git clone https://github.com/JarodMica/dataset-maker.git
   cd dataset-maker
   uv sync
   

3. Launch the UI:

   uv run .\gradio_interface.py
   

Core workflow

1. Create a project in the UI (creates datasets_folder/<project>).
2. Upload audio into wavs/.
3. Optionally combine short clips into longer files for faster transcription.
4. Transcribe using:
   • Silence slicer (local slicer)
   • WhisperX timestamps (Silero VAD hardcoded for pyannote 4.0 compatibility)
   • Emilia pipe (full multi-stage pipeline)
5. Review/correct transcripts and export to the target dataset format.

Outputs

Below are examples based on the export code in gradio_interface.py.

Base (Tortoise/XTTS) exports a train.txt with file_id | transcript and a wavs/ folder:

train.txt
seg000001.wav|Hello world.
seg000002.wav|Another line.

StyleTTS format (from the Adjust train.txt tab) adds a speaker tag:

train.txt
seg000001.wav|Hello world.|speaker_01

GPTSoVITS format (from the Adjust train.txt tab) adds slicer_opt + language:

train.txt
seg000001.wav|slicer_opt|en|Hello world.

Higgs Audio export writes per-sample files and metadata.json:

speaker_01_000000.wav
speaker_01_000000.txt
metadata.json

metadata.json entries include speaker and quality fields (e.g. speaker_id, gender, duration, quality_score).

VibeVoice export writes a JSONL manifest with a "Speaker 0: " prefix:

{"text":"Speaker 0: Hello world.","audio":"<dataset_folder>/vibevoice_000000.wav"}

Qwen 3 TTS export writes a JSONL manifest with audio + ref_audio paths (relative to the export folder). Audio files are copied from the transcribe outputs into data/ (prefixed if collisions), and the reference clip is copied as data/ref_audio.<ext>:

{"audio":"./data/seg_00000001.wav","text":"Hello world.","ref_audio":"./data/ref_audio.wav"}

In the UI, the Qwen 3 TTS reference audio picker lists files from the project's transcribe/ folder. Selecting the Qwen 3 TTS export format refreshes the reference list and resets to the first available item.

Emilia pipe (Originally intended for IndexTTS 2, not exclusive) writes a project-scoped JSONL and audio folder:

{"id":"<clip_id>","text":"Hello world.","audio":"<project>_emilia_dataset/audio/<clip_id>.mp3","speaker":"<base>_SPEAKER_X","language":"en","duration":1.23,"source":"original.wav"}

Emilia configuration

The Emilia path in the UI expects Emilia/config.json. Copy Emilia/config_example.json to Emilia/config.json and update:

• huggingface_token
• entrypoint.input_folder_path (optional override from UI)
• model paths in separate.step1.model_path and mos_model.primary_model_path

Onnx Runtime Issue CUDA

CUDAExecution provider may not be found even when using uv. The fix is to remove and then add optimum[onnxruntime-gpu] in the terminal.

Problematic

uv run python
>>> import onnxruntime as ort
>>> print("Available providers:", ort.get_available_providers())
Available providers: ['AzureExecutionProvider', 'CPUExecutionProvider']

Fixed

uv run python
>>> import onnxruntime as ort
>>> print("Available providers:", ort.get_available_providers())
Available providers: ['TensorrtExecutionProvider', 'CUDAExecutionProvider', 'CPUExecutionProvider']

More documentation

See docs/CONFIG.md, docs/ARCHITECTURE.md, docs/OPERATIONS.md, and docs/TROUBLESHOOTING.md for details.