Skip to content

VoiceHub: One Speech Model Lifecycle

A task-aware Python library for inference, data preparation, and model-specific fine-tuning across modern TTS, ASR, and VAD families.

VoiceHub continuous integration status VoiceHub documentation build status VoiceHub supports Python 3.10 and later VoiceHub is licensed under Apache 2.0

What is VoiceHub?

VoiceHub presents text-to-speech, automatic speech recognition, and voice activity detection through shared configuration, processor, model, typed output, and trainer APIs. Implementations remain architecture-aware: codec language models, CTC and transducer ASR, speech encoder-decoders, flow-matching and diffusion models, audio/frame classifiers, VITS-style adversarial systems, and upstream-native pipelines keep their own conditioning, objectives, parameter ownership, and export rules.

The registry contains 68 integrations: 34 TTS backends, 23 ASR providers, and 11 VAD providers. Fine-tuning support is checkpoint- and runtime-specific; an inference integration does not imply that its current VoiceHub artifact is differentiable. Use the TTS catalog, TTS training matrix, and ASR/VAD support matrix to select an integration.

Model source and every built-in TTS, ASR, and VAD inference runtime are installed with VoiceHub. Checkpoint weights are downloaded lazily or provided as local paths. Add only voicehub[training] for fine-tuning and reporting. The Apache-2.0 license covers VoiceHub itself; integrated source, checkpoints, codecs, datasets, and generated audio may have separate terms.

  • Getting started

    Install VoiceHub from the current source tree and run the first generation request through the shared model factory.

    Quick start

  • Inference

    Discover integrations, load Hub or local checkpoints, configure reproducible generation, and consume normalized audio.

    Inference guide

  • Speech recognition

    Transcribe files or in-memory audio through native CTC, transducer, encoder-decoder, and Whisper-family graphs with normalized timestamps.

    ASR guide

  • Voice activity detection

    Detect ordered speech regions with native Wav2Vec2, Silero, PyanNet, WebRTC, SpeechBrain, NeMo, or FunASR FSMN.

    VAD guide

  • Data preparation

    Build auditable manifests, validate audio, prevent speaker or session leakage, and create model-specific training inputs.

    Data preparation guide

  • Training

    Validate checkpoint boundaries, run native objectives, evaluate, resume complete checkpoints, and save portable artifacts.

    Training guide

  • Models

    Compare TTS registry entries, default checkpoints, capabilities, source provenance, and constraints.

    Model catalog

  • ASR and VAD support

    Compare provider families, default runtime coverage, output capabilities, and the exact native-trainable or inference-only boundary.

    Speech-input matrix

  • Training support

    Check the exact raw-data, preprocessed, specialized, or unavailable fine-tuning boundary for every integration.

    Training matrix

  • Notebooks

    Run focused inference, data, and training examples or follow the complete Dia workflow through export and fresh-runtime reload.

    Open the notebook gallery

  • API reference

    Look up factories, outputs, trainer arguments, callbacks, collators, strategies, artifacts, and extension registries.

    Browse the API

  • Architecture

    Understand the registry, model wrappers, adapters, runtime strategies, checkpoints, and portable artifact boundaries.

    Library architecture

  • Add a model

    Implement and test a lazy wrapper, training specification, specialized adapter when required, and export contract.

    Model integration guide