Workflows

VoiceHub shares one public lifecycle across TTS, ASR, and VAD models while keeping task- and architecture-specific semantics explicit. Choose a workflow below. The notebook gallery provides focused inference, data, and training examples plus the complete Dia lifecycle.

1. 01 Discover a model Inspect registry support and select a compatible checkpoint.
2. 02 Run baseline inference Generate a reference sample before changing any weights.
3. 03 Prepare consented data Validate audio, provenance, transcripts, and split boundaries.
4. 04 Build the model batch Create the tokens, codes, masks, or flow targets the model expects.
5. 05 Train and evaluate Run the verified recipe and measure held-out behavior.
6. 06 Save the artifact Write portable weights, metadata, and any native export.
7. 07 Compare inference Reload the result and compare it with the baseline.

Inference

Use the TTS inference guide to:

• discover models without importing their ML runtimes;
• load Hub checkpoints and local artifacts;
• configure deterministic generation;
• provide voice, language, style, and reference conditioning;
• consume the normalized TTSOutput; and
• keep serving optimizations separate from the training graph.

Use the speech recognition guide and voice activity detection guide to:

• discover providers through task-filtered registry metadata;
• load file, array, tensor, mapping, or AudioInput audio;
• consume normalized ASROutput and VADOutput values;
• configure timestamps, long-form decoding, thresholds, and segmentation; and
• distinguish native, optimized, and upstream-managed runtimes.

Data preparation

Use the data preparation guide to:

• design auditable JSON Lines manifests;
• validate sample rate, channels, finite samples, transcripts, and consent;
• prevent speaker and recording-session leakage;
• understand raw-data versus preprocessed routes; and
• inspect model-owned codec, mask, and target layouts before training.

The ASR and VAD data guide adds transcript and speech-region manifests, CTC/seq2seq/transducer batch shapes, clip/frame classification, and schema-aware variable-length audio collation.

Training

Use the training guide to:

• verify whether the exact model variant can be fine-tuned;
• select the differentiable checkpoint instead of GGUF or serving exports;
• run verified native LM and flow recipes, and understand how the trainer represents model-dependent VITS, GAN, and hybrid phases;
• start with a one-step gradient smoke test;
• resume complete VoiceHub checkpoints exactly; and
• distinguish a portable inference artifact from optimizer-bearing state.

After the one-step baseline is correct, use TTS optimization by architecture. It keeps VITS adversarial windows, codec/LLM token budgets, and diffusion frame/EMA policies separate, adds architecture-compatible compile/Triton/CUDA/FA4 plans, and ties every default to a pinned original implementation.

Use the focused diffusion and flow optimization and neural codec optimization guides for the active model inventories, safe repeated graph boundaries, native codec component catalog, fixed-shape CUDA Graph rules, and exact versus approximate policies.

For production LLM-based synthesis, use the vLLM and SGLang serving guide. It separates flat token generation from complete Omni speech pipelines, lists every verified model/backend pairing, and keeps incompatible engine dependencies outside the VoiceHub environment.

Training support is not universal

TTS support remains checkpoint-specific; read the TTS training matrix. ASR and VAD providers separately declare a native, preprocessed, custom-component, or inference-only boundary in the speech-input support matrix.

Artifact boundaries

Artifact	Purpose
Source manifest	Provenance, consent, transcript, speaker/session identity, and audio path
Prepared dataset	Normalized audio and immutable, versioned split records
Dataset and collator	Model-specific tokens, codes, masks, flow targets, or phase inputs
checkpoint-N/	Exact resume: model, optimizers, scheduler, RNG, sampler, recipe state
trainer.save_model()	Portable VoiceHub artifact for reload and weight warm-start
native_export/	Source-native export with semantics declared by the model adapter

A safetensors file is a weight container, not an exact training checkpoint. GGUF, ONNX, TensorRT, JIT, vLLM, and other optimized serving artifacts are not automatically differentiable.