Notebooks

Each notebook follows the same top-to-bottom pattern:

1. install VoiceHub;
2. edit one configuration cell;
3. run lightweight checks;
4. enable the expensive stage; and
5. inspect the output before continuing.

Notebook	Purpose	GitHub	Colab
Inference	TTS, ASR, VAD, and duration checks	View	Run
Data preparation	Validate records and create group-disjoint splits	View	Run
Training	Inspect support and start with one optimizer step	View	Run
Dia workflow	Baseline, data, fine-tune, export, and reload	View	Run

Every registry entry with a Hugging Face checkpoint also has a dedicated, opt-in notebook. The model guides pair each notebook with model information, data preparation, and training steps.

Safe defaults

Real inference and training stay off until their flags are changed:

RUN_TTS = False
RUN_ASR = False
RUN_VAD = False
RUN_TRAINING = False

This lets readers inspect the registry, request types, data contracts, and one-step settings without downloading checkpoints.

Recommended order

1. Start with inference.ipynb and run one model in eager mode.
2. Confirm that TTS audio is at least 10 seconds using the returned waveform length and sample rate.
3. Run data_preparation.ipynb with a few authorized recordings.
4. Run training.ipynb with max_steps=1.
5. Use tts_workflow.ipynb only when the separate stages are understood.

Use a GPU runtime for model inference and fine-tuning. CPU is sufficient for the lightweight discovery and contract cells.

Before recording results

• Pin the VoiceHub revision, checkpoint revision, dataset revision, and seed.
• Record the device, PyTorch version, precision, warm-up count, and measured audio duration.
• Compare optimization modes with identical inputs.
• Report both latency and peak memory; never infer percentages from a config.
• Listen to every compared sample and keep the eager result as the quality baseline.

See Inference, Training, and Data preparation for detailed contracts.