Copy page

Public exports

VoiceHub exposes one explicit package-root surface for discovery, loading, inference, training, optimization, serving, and serialization. This page is generated from voicehub.__all__ and complements the task-oriented full API reference. Its grouping follows the role of Transformers' Main Classes while keeping speech-specific contracts and names.

Every row resolves from voicehub, points to repository source, records the canonical defining or re-export module, and includes a callable signature or an explicit constant/type-alias marker. Generation fails on duplicate, unresolved, undocumented, source-less, or stale exports.

Current inventory: 261 public exports.

Package metadata

Export	Kind	Canonical module	Signature	Summary	Lazy
__version__	constant	voicehub	constant	Installed VoiceHub package version.	no

Configuration, factories, and models

Export	Kind	Canonical module	Signature	Summary	Lazy
AudioProcessor	class	voicehub.processing_utils	(**kwargs)	Validate the public audio-input envelope without loading audio.	no
AutoConfig	class	voicehub.auto	()	Instantiate the registered configuration class for a model type.	no
AutoInferenceModel	class	voicehub.automodel	()	Factory class that dynamically loads and instantiates TTS model backends.	no
AutomaticSpeechRecognitionPipeline	class	voicehub.pipelines	(model)	Pipeline that maps audio to :class:~voicehub.ASROutput.	no
AutoModel	class	voicehub.auto	()	Task-aware entry point for every registered speech model.	no
AutoModelForSpeechRecognition	class	voicehub.auto	()	Load a registered automatic speech-recognition model.	no
AutoModelForTextToSpeech	class	voicehub.auto	()	Load a registered text-to-speech model.	no
AutoModelForVoiceActivityDetection	class	voicehub.auto	()	Load a registered voice-activity-detection model.	no
AutoProcessor	class	voicehub.auto	()	Create the processor paired with a VoiceHub speech configuration.	no
BatchFeature	class	voicehub.processing_utils	inherited constructor	Dictionary of processor values with a tensor-like to helper.	no
get_default_model_spec	callable	voicehub.models.registry	(task: 'SpeechTask | str') -> 'ModelSpec | None'	Return the registry-declared default for a speech task, if present.	no
get_model_spec	callable	voicehub.models.registry	(model_type: 'str') -> 'ModelSpec'	Return registry metadata or raise an error containing valid choices.	no
list_model_specs	callable	voicehub.models.registry	(*, task: 'SpeechTask | str | None' = None, native: 'bool | None' = None) -> 'tuple[ModelSpec, ...]'	Return registered models with task and native-runtime filters.	no
MODEL_CATALOG	constant	voicehub.models.registry	constant	Read-only facade over the built-in model registry.	no
ModelRegistry	class	voicehub.models.registry	(specs: 'Iterable[ModelSpec]' = (), *, aliases: 'Mapping[str, str] | None' = None) -> 'None'	Thread-safe catalogue of lazily imported speech models.	no
ModelSpec	class	voicehub.models.registry	(model_type: 'str', module: 'str', class_name: 'str', default_model_path: 'str', install_extra: 'str | None' = None, capabilities: 'tuple[str, ...]' = ('text-to-speech',), config_module: 'str' = 'voicehub.configuration_utils', config_class: 'str' = 'VoiceHubConfig', task: 'SpeechTask | str' = <SpeechTask.TEXT_TO_SPEECH: 'text-to-speech'>, architecture: 'str | None' = None, components: 'tuple[str, ...]' = (), default_for_task: 'bool' = False, processor_module: 'str | None' = None, processor_class: 'str | None' = None) -> None	Metadata required to discover and lazily import a backend.	no
Pipeline	class	voicehub.pipelines	(model)	Small task adapter around a loaded or lazily constructed speech model.	no
pipeline	callable	voicehub.pipelines	(task: 'SpeechTask | str', model=None, *, model_type: 'str | None' = None, config=None, device: 'str | None' = None, inference_strategy=None, config_kwargs: 'Mapping[str, Any] | None' = None, model_kwargs: 'Mapping[str, Any] | None' = None) -> 'Pipeline'	Create a task pipeline from a checkpoint source or existing model.	no
PreTrainedSpeechModel	class	voicehub.modeling_utils	()	Marker base shared by task-specific pretrained speech wrappers.	no
PreTrainedTTSModel	class	voicehub.modeling_utils	(config: 'VoiceHubConfig', *, device: 'str' = 'auto', lazy_load: 'bool' = True)	Base lifecycle shared by source-integrated VoiceHub architectures.	no
register_model_alias	callable	voicehub.models.registry	(alias: 'str', model_type: 'str', *, exist_ok: 'bool' = False) -> 'None'	Register a public alias, allowing idempotence when requested.	no
register_model_spec	callable	voicehub.models.registry	(spec: 'ModelSpec', *, aliases: 'Iterable[str]' = (), exist_ok: 'bool' = False) -> 'None'	Register or explicitly replace one lazily imported model backend.	no
SpeechTask	enum	voicehub.tasks	(value)	Canonical public tasks supported by VoiceHub.	no
TextToSpeechPipeline	class	voicehub.pipelines	(model)	Pipeline that maps text to :class:~voicehub.TTSOutput.	no
unregister_model_alias	callable	voicehub.models.registry	(alias: 'str', *, missing_ok: 'bool' = False) -> 'str | None'	Remove a public alias and return its former canonical target.	no
unregister_model_spec	callable	voicehub.models.registry	(model_type: 'str', *, missing_ok: 'bool' = False) -> 'ModelSpec | None'	Remove a model backend and every alias that resolves to it.	no
VoiceActivityDetectionPipeline	class	voicehub.pipelines	(model)	Pipeline that maps audio to :class:~voicehub.VADOutput.	no
VoiceHubConfig	class	voicehub.configuration_utils	(, sample_rate: 'int' = 24000, architectures: 'list[str] | None' = None, name_or_path: 'str | Path' = '', return_dict: 'bool' = True, output_hidden_states: 'bool' = False, output_attentions: 'bool' = False, generation_config: 'dict[str, Any] | None' = None, *kwargs)	Serializable configuration shared by all VoiceHub speech architectures.	no
VoiceHubProcessor	class	voicehub.processing_utils	(**kwargs)	Transform raw synthesis inputs into model-ready values.	no

Inputs and normalized outputs

Export	Kind	Canonical module	Signature	Summary	Lazy
ASRInferenceConfig	class	voicehub.inference_configuration	(, language: 'str | None' = None, task: 'str' = 'transcribe', return_timestamps: 'bool | str' = False, chunk_length_s: 'float | None' = None, stride_length_s: 'float | tuple[float, float] | None' = None, batch_size: 'int | None' = None, num_beams: 'int | None' = None, max_new_tokens: 'int | None' = None, hotwords: 'str | tuple[str, ...] | list[str] | None' = None, *kwargs)	Decoding and long-audio controls shared by ASR providers.	no
ASROutput	class	voicehub.modeling_outputs	(text: 'str', segments: 'tuple[ASRSegment, ...]' = (), language: 'str | None' = None, duration: 'float | None' = None, metadata: 'dict[str, Any]' = <factory>) -> None	Normalized transcription produced by any ASR backend.	no
ASRSegment	class	voicehub.modeling_outputs	(text: 'str', start: 'float | None' = None, end: 'float | None' = None, confidence: 'float | None' = None, language: 'str | None' = None, speaker: 'str | None' = None, words: 'tuple[ASRWord, ...]' = (), metadata: 'dict[str, Any]' = <factory>) -> None	A timestamped transcription segment.	no
ASRWord	class	voicehub.modeling_outputs	(text: 'str', start: 'float | None' = None, end: 'float | None' = None, confidence: 'float | None' = None, speaker: 'str | None' = None) -> None	One recognized word with optional timing and confidence metadata.	no
AudioInput	class	voicehub.audio	(waveform: 'Any', sampling_rate: 'int', path: 'Path | None' = None) -> None	Materialized mono waveform with an explicit sampling rate.	no
default_data_collator	callable	voicehub.data_collator	(features: 'list[Any]', return_tensors: 'str' = 'pt') -> 'dict[str, Any]'	Collate mapping-like samples while preserving strings and metadata.	no
DefaultDataCollator	class	voicehub.data_collator	(return_tensors: 'str' = 'pt') -> None	Callable object form of :func:default_data_collator.	no
load_audio	callable	voicehub.audio	(audio: 'AudioInput | Mapping[str, Any] | str | Path | Any', *, sampling_rate: 'int | None' = None, target_sampling_rate: 'int | None' = None) -> 'AudioInput'	Load, downmix, and optionally resample one waveform.	no
PreTrainedASRModel	class	voicehub.audio_modeling_utils	(config: 'VoiceHubConfig', *, device: 'str' = 'auto', lazy_load: 'bool' = True)	Base class for speech-recognition models.	no
PreTrainedAudioModel	class	voicehub.audio_modeling_utils	(config: 'VoiceHubConfig', *, device: 'str' = 'auto', lazy_load: 'bool' = True)	Shared lazy lifecycle for ASR and VAD wrappers.	no
PreTrainedVADModel	class	voicehub.audio_modeling_utils	(config: 'VoiceHubConfig', *, device: 'str' = 'auto', lazy_load: 'bool' = True)	Base class for voice-activity-detection models.	no
SpeechInferenceConfig	class	voicehub.inference_configuration	(**kwargs)	Base for task-specific, extensible inference configuration.	no
SpeechSegment	class	voicehub.modeling_outputs	(start: 'float', end: 'float', score: 'float | None' = None, label: 'str' = 'speech', channel: 'int | str | None' = None, metadata: 'dict[str, Any]' = <factory>) -> None	One detected interval of speech or non-speech.	no
SpeechTrainingOutput	class	voicehub.modeling_outputs	(loss: 'Any | None' = None, logits: 'Any | None' = None, predictions: 'Any | None' = None, audio_values: 'Any | None' = None, hidden_states: 'Any | None' = None, attentions: 'Any | None' = None, training_phase: 'str | None' = None, optimizer_names: 'tuple[str, ...]' = (), losses: 'dict[str, Any]' = <factory>, metadata: 'dict[str, Any]' = <factory>) -> None	Task-neutral differentiable output consumed by :class:voicehub.Trainer.	no
TTSGenerationConfig	class	voicehub.generation_configuration	(, output_file: 'str | Path | None' = None, seed: 'int | None' = None, speed: 'float | None' = None, temperature: 'float | None' = None, top_p: 'float | None' = None, max_new_tokens: 'int | None' = None, *kwargs)	Serializable, extensible generation options for every TTS model.	no
TTSOutput	class	voicehub.modeling_outputs	(audio: 'Any', sample_rate: 'int', file_path: 'str | Path | None' = None, metadata: 'dict[str, Any]' = <factory>) -> None	Audio output with sampling metadata and optional backend details.	no
TTSTrainingOutput	class	voicehub.modeling_outputs	(loss: 'Any | None' = None, logits: 'Any | None' = None, predictions: 'Any | None' = None, audio_values: 'Any | None' = None, hidden_states: 'Any | None' = None, attentions: 'Any | None' = None, training_phase: 'str | None' = None, optimizer_names: 'tuple[str, ...]' = (), losses: 'dict[str, Any]' = <factory>, metadata: 'dict[str, Any]' = <factory>) -> None	Backward-compatible name for the shared speech training output.	no
VADInferenceConfig	class	voicehub.inference_configuration	(, threshold: 'float' = 0.5, onset: 'float | None' = None, offset: 'float | None' = None, min_speech_duration_ms: 'int' = 250, min_silence_duration_ms: 'int' = 100, speech_pad_ms: 'int' = 30, max_speech_duration_s: 'float | None' = None, window_size_samples: 'int | None' = None, return_frames: 'bool' = False, *kwargs)	Thresholding and segmentation controls shared by VAD providers.	no
VADOutput	class	voicehub.modeling_outputs	(segments: 'tuple[SpeechSegment, ...]', duration: 'float | None' = None, sample_rate: 'int | None' = None, probabilities: 'Any | None' = None, metadata: 'dict[str, Any]' = <factory>) -> None	Normalized speech regions produced by any VAD backend.	no
VADSegment	class	voicehub.modeling_outputs	(start: 'float', end: 'float', score: 'float | None' = None, label: 'str' = 'speech', channel: 'int | str | None' = None, metadata: 'dict[str, Any]' = <factory>) -> None	One detected interval of speech or non-speech.	no

Inference and serving

Export	Kind	Canonical module	Signature	Summary	Lazy
bridge_vllm_omni_tts_config	callable	voicehub.diffusion_serving.bridge	(model_type: 'str', config: 'LLMBackendConfig') -> 'tuple[DiffusionTTSServingPlan, LLMBackendConfig]'	Validate and reuse an existing vLLM-Omni speech configuration.	yes
detect_vllm_omni_features	callable	voicehub.diffusion_serving.vllm_omni	(*, probe_registry: 'bool' = True) -> 'VLLMOmniFeatureStatus'	Detect the optional engine and its public plugin registration API.	yes
DiffusionServingBackend	enum	voicehub.diffusion_serving.capabilities	(value)	Serving runtimes kept distinct by both engine and modality.	yes
DiffusionServingCapability	class	voicehub.diffusion_serving.capabilities	(backend: 'DiffusionServingBackend', engine: 'str', diffusion_modalities: 'tuple[str, ...]', supports_tts: 'bool', supports_tts_diffusion: 'bool', verified_tts_models: 'tuple[str, ...]' = (), supports_custom_plugins: 'bool' = False, notes: 'str' = '') -> None	Static, dependency-free facts about one serving runtime.	yes
DiffusionServingCompatibilityError	exception	voicehub.diffusion_serving.capabilities	inherited constructor	Raised when an engine cannot preserve a TTS diffusion pipeline.	yes
DiffusionTTSServingPlan	class	voicehub.diffusion_serving.capabilities	(model_type: 'str', capability: 'DiffusionServingCapability', verified: 'bool', experimental: 'bool' = False, plugin: 'VLLMOmniDiffusionPlugin | None' = None) -> None	Resolved TTS diffusion support without constructing an engine client.	yes
EagerInferenceStrategy	class	voicehub.inference_strategy	()	Default no-op strategy using each model's native eager runtime.	no
get_diffusion_serving_capability	callable	voicehub.diffusion_serving.capabilities	(backend: 'str | DiffusionServingBackend') -> 'DiffusionServingCapability'	Return one dependency-free backend capability record.	yes
get_inference_strategy	callable	voicehub.inference_strategy	(strategy: 'str | InferenceStrategy | None' = None) -> 'InferenceStrategy'	Resolve a strategy name or validate an existing strategy instance.	no
get_llm_backend_support	callable	voicehub.llm_serving.support	(model_type: 'str', backend: 'str | LLMBackend', *, transport: 'str | LLMBackendTransport' = <LLMBackendTransport.AUTO: 'auto'>) -> 'tuple[LLMBackendSupport, LLMBackendTransport]'	Resolve one pairing and its concrete transport, or fail clearly.	no
InferenceStrategy	class	voicehub.inference_strategy	()	Lifecycle hooks for an inference optimization runtime.	no
list_diffusion_serving_capabilities	callable	voicehub.diffusion_serving.capabilities	(*, supports_tts: 'bool | None' = None, supports_visual_diffusion: 'bool | None' = None) -> 'tuple[DiffusionServingCapability, ...]'	List serving facts without importing any optional engine.	yes
list_inference_strategies	callable	voicehub.inference_strategy	() -> 'tuple[str, ...]'	Return registered strategy names in deterministic order.	no
list_llm_backend_support	callable	voicehub.llm_serving.support	(*, backend: 'str | LLMBackend | None' = None, model_type: 'str | None' = None) -> 'tuple[LLMBackendSupport, ...]'	List verified pairings without importing either serving engine.	no
LLMBackend	enum	voicehub.llm_serving.configuration	(value)	Language-model serving engines understood by VoiceHub.	no
LLMBackendConfig	class	voicehub.llm_serving.configuration	(backend: 'LLMBackend | str', endpoint: 'str | None' = None, transport: 'LLMBackendTransport | str' = <LLMBackendTransport.AUTO: 'auto'>, model: 'str | None' = None, api_key: 'str | None' = None, timeout: 'float' = 300.0, headers: 'Mapping[str, str] | None' = None, extra_body: 'Mapping[str, Any] | None' = None, max_response_bytes: 'int' = 536870912) -> None	Connection settings for a separately managed vLLM/SGLang server.	no
LLMBackendSupport	class	voicehub.llm_serving.support	(model_type: 'str', backend: 'LLMBackend', transports: 'tuple[LLMBackendTransport, ...]', default_transport: 'LLMBackendTransport', engine: 'str', checkpoint_family: 'str', notes: 'str' = '', task_type_without_reference: 'str | None' = None, task_type_with_reference: 'str | None' = None, task_type_aliases: 'tuple[tuple[str, str], ...]' = (), reference_format: 'str' = 'flat', speech_string_options: 'tuple[str, ...]' = ()) -> None	One verified model/backend protocol pairing.	no
LLMBackendTransport	enum	voicehub.llm_serving.configuration	(value)	Protocol used between a VoiceHub wrapper and an engine server.	no
register_inference_strategy	callable	voicehub.inference_strategy	(name: 'str', factory: 'InferenceStrategyFactory | type[InferenceStrategy]', *, exist_ok: 'bool' = False) -> 'None'	Register a zero-argument strategy factory.	no
register_llm_backend_support	callable	voicehub.llm_serving.support	(support: 'LLMBackendSupport', *, exist_ok: 'bool' = False) -> 'None'	Register one process-local model/backend capability record.	no
resolve_diffusion_tts_backend	callable	voicehub.diffusion_serving.capabilities	(model_type: 'str', backend: 'str | DiffusionServingBackend', *, plugin: 'VLLMOmniDiffusionPlugin | None' = None) -> 'DiffusionTTSServingPlan'	Resolve complete TTS diffusion support, failing closed by default.	yes
TokenGenerationRequest	class	voicehub.llm_serving.protocol	(prompt_token_ids: 'Sequence[int]', max_new_tokens: 'int', temperature: 'float' = 1.0, top_p: 'float | None' = None, top_k: 'int | None' = None, min_p: 'float | None' = None, repetition_penalty: 'float' = 1.0, stop_token_ids: 'Sequence[int]' = (), seed: 'int | None' = None) -> None	Tokenizer-free request sent to a flat causal-LM server.	no
TokenGenerationResult	class	voicehub.llm_serving.protocol	(token_ids: 'Sequence[int]', finish_reason: 'str | None' = None, prompt_tokens: 'int | None' = None, completion_tokens: 'int | None' = None) -> None	Generated suffix IDs and optional engine accounting.	no
TorchCompileInferenceStrategy	class	voicehub.inference_strategy	(*, backend: 'str' = 'inductor', mode: 'str | None' = None, fullgraph: 'bool' = False, dynamic: 'bool | None' = True, options: 'dict[str, Any] | None' = None, requirement: 'str' = 'required') -> 'None'	Opt-in, reversible torch.compile inference preparation.	no
unregister_inference_strategy	callable	voicehub.inference_strategy	(name: 'str') -> 'None'	Remove a custom strategy registration.	no
unregister_llm_backend_support	callable	voicehub.llm_serving.support	(model_type: 'str', backend: 'str | LLMBackend', *, missing_ok: 'bool' = False) -> 'LLMBackendSupport | None'	Remove one process-local capability record and return it.	no
VLLMOmniDiffusionPlugin	class	voicehub.diffusion_serving.vllm_omni	(model_type: 'str', model_arch: 'str', module_name: 'str', class_name: 'str', complete_tts_pipeline: 'bool' = False, pre_process_func_name: 'str | None' = None, post_process_func_name: 'str | None' = None, action_post_process_func_name: 'str | None' = None, ir_op_priority_func_name: 'str | None' = None) -> None	Experimental contract for a vLLM-Omni out-of-tree diffusion model.	yes
VLLMOmniFeatureStatus	class	voicehub.diffusion_serving.vllm_omni	(installed: 'bool', version: 'str | None', register_diffusion_model: 'bool | None', error: 'str | None' = None) -> None	Installed-version and registry API status for vLLM-Omni.	yes

Training

Export	Kind	Canonical module	Signature	Summary	Lazy
AcousticTrainingAdapter	class	voicehub.training.adapters	(model, spec: 'ModelTrainingSpec')	Mel, codec, or waveform reconstruction objective.	no
ALL_MODEL_TRAINING_SPECS	constant	voicehub.training.specs	constant	Immutable catalog of every built-in model training specification.	no
ASRDataArchitecture	enum	voicehub.training.asr_data_contracts	(value)	Canonical source-data layouts used by ASR fine-tuning recipes.	no
ASRDataReadiness	enum	voicehub.training.asr_data_contracts	(value)	How far VoiceHub owns a model's ASR dataset preparation path.	no
ASRDataset	class	voicehub.training.asr_datasets	(records: 'Iterable[Mapping[str, Any]]', *, model_type: 'str | None' = None, architecture: 'ASRDataArchitecture | str | None' = None, root: 'str | PathLike[str] | None' = None, aliases: 'Mapping[str, str] | None' = None, validate: 'bool' = True, validate_files: 'bool' = False, transform: 'Callable[[dict[str, Any]], Mapping[str, Any]] | None' = None, transform_fingerprint: 'str | None' = None) -> 'None'	Validated ASR records from mappings, manifests, or WAV folders.	no
ASRDatasetSpec	class	voicehub.training.asr_data_contracts	(architecture: 'ASRDataArchitecture', variants: 'tuple[ASRRecordVariant, ...]', model_type: 'str | None' = None, sample_rate: 'int | None' = None, description: 'str' = '', readiness: 'ASRDataReadiness | None' = None, training_support: 'str | None' = None, homogeneous_batch_fields: 'tuple[tuple[str, ...], ...]' = (), field_aliases: 'Mapping[str, str] | tuple[tuple[str, str], ...]' = (), record_normalizer: 'str | None' = None, record_normalizer_phase: 'str' = 'after-aliases') -> None	Inspectable source-data contract for one ASR architecture or model.	no
ASRRecordVariant	class	voicehub.training.asr_data_contracts	(name: 'str', required_fields: 'tuple[str, ...]' = (), one_of: 'tuple[tuple[str, ...], ...]' = (), at_most_one_of: 'tuple[tuple[str, ...], ...]' = (), forbidden_fields: 'tuple[str, ...]' = (), requires: 'tuple[tuple[str, tuple[str, ...]], ...]' = (), requires_one_of: 'tuple[tuple[str, tuple[str, ...]], ...]' = (), description: 'str' = '', preprocessed: 'bool' = False) -> None	One accepted raw or preprocessed ASR record shape.	no
AudioClassificationTrainingAdapter	class	voicehub.training.adapters	(model, spec: 'ModelTrainingSpec')	Audio classifier with explicit CE/BCE-with-logits fallbacks.	no
AudioFieldSchema	class	voicehub.training.collators	(sequence_dim: 'int' = 0, padding_value: 'float | int | None' = None, padding_side: 'str' = 'right', length_field: 'str | None' = None, mask_field: 'str | None' = None, pad_to_multiple_of: 'int | None' = None, allow_missing: 'bool' = False) -> None	Describe the variable-length dimension of one training field.	no
AutoTrainingAdapter	class	voicehub.training.auto	()	Resolve the adapter paired with a VoiceHub model or future family.	no
BaseTrainingAdapter	class	voicehub.training.adapters	(model, spec: 'ModelTrainingSpec')	Expose an inference wrapper's source modules through a training API.	no
build_diffusion_training_pair	callable	voicehub.training.objectives.diffusion	(samples: 'Any', *, coefficient_fn: 'Callable[[Any, Any], tuple[Any, Any]]', prediction_type: 'str' = 'epsilon', timesteps: 'Any | None' = None, noise: 'Any | None' = None, generator: 'Any | None' = None, timestep_sampler: 'Callable[..., Any] | None' = None, noise_sampler: 'Callable[..., Any] | None' = None, num_train_timesteps: 'int | None' = None) -> 'DiffusionTrainingPair'	Construct a discrete diffusion input and epsilon/v/sample target.	no
build_flow_matching_training_pair	callable	voicehub.training.objectives.diffusion	(samples: 'Any', *, prediction_type: 'str' = 'velocity', timesteps: 'Any | None' = None, noise: 'Any | None' = None, generator: 'Any | None' = None, timestep_sampler: 'Callable[..., Any] | None' = None, noise_sampler: 'Callable[..., Any] | None' = None) -> 'DiffusionTrainingPair'	Construct a linear flow-matching path and its supervised target.	no
CausalLMTrainingAdapter	class	voicehub.training.adapters	(model, spec: 'ModelTrainingSpec')	Autoregressive codec-token objective with shifted cross entropy.	no
CompositeTrainingAdapter	class	voicehub.training.adapters	(model, spec: 'ModelTrainingSpec')	Multi-component adapter that prefers phase-specific native losses.	no
CTCTrainingAdapter	class	voicehub.training.adapters	(model, spec: 'ModelTrainingSpec')	CTC adapter that preserves backend blank and alignment semantics.	no
DataCollatorForAudioTraining	class	voicehub.training.collators	(padding_value: 'float' = 0.0, label_pad_token_id: 'int' = -100, return_attention_mask: 'bool' = True, return_input_lengths: 'bool' = False, field_schemas: 'Mapping[str, AudioFieldSchema | Mapping[str, Any]] | None' = None) -> None	Collate heterogeneous audio examples without guessing task semantics.	no
DataCollatorForTTSTraining	class	voicehub.training.collators	(padding_value: 'float' = 0.0, label_pad_token_id: 'int' = -100, return_attention_mask: 'bool' = True, return_input_lengths: 'bool' = False, field_schemas: 'Mapping[str, AudioFieldSchema | Mapping[str, Any]] | None' = None) -> None	Collate heterogeneous audio examples without guessing task semantics.	no
diffusion_tts_acceleration_plan	callable	voicehub.training.tts_acceleration	(*, kernel_backend: 'str' = 'auto', attention_policy: 'str' = 'auto', use_torch_compile: 'bool' = True, compile_backend: 'str' = 'inductor', compile_mode: 'str | None' = 'max-autotune-no-cudagraphs', compile_fullgraph: 'bool' = False, compile_dynamic: 'bool | None' = True, compile_requirement: 'str' = 'auto') -> 'tuple[OptimizationPass, ...]'	Build the diffusion-TTS bias-GELU, FA4, and compile plan.	no
DiffusionTrainingPair	class	voicehub.training.objectives.diffusion	(noisy_inputs: 'Any', targets: 'Any', timesteps: 'Any', noise: 'Any', alpha: 'Any', sigma: 'Any') -> None	Noisy model input and supervised target for a diffusion recipe.	no
DiffusionTTSOptimizationConfig	class	voicehub.training.tts_optimization	(learning_rate: 'float' = 7.5e-05, weight_decay: 'float' = 0.01, adam_beta1: 'float' = 0.9, adam_beta2: 'float' = 0.999, adam_epsilon: 'float' = 1e-08, warmup_steps: 'int' = 20000, max_grad_norm: 'float' = 1.0, per_device_train_batch_size: 'int' = 64, frame_budget: 'int' = 38400, length_field: 'str' = 'num_frames', use_bf16: 'bool' = True, gradient_checkpointing: 'bool' = True, fused_adamw: 'bool' = True, ema_decay: 'float' = 0.9999, ema_update_after_step: 'int' = 0, ema_update_every: 'int' = 1, source_url: 'str' = 'https://github.com/SWivid/F5-TTS/tree/9c614e9657089213efc6a7421b30630be138a3f5', techniques: 'tuple[str, ...]' = ('mel-frame-budget batching', 'activation checkpointing', 'scaled dot-product attention', 'fused AdamW when CUDA supports it', 'linear warmup and decay', 'optimizer-update-coupled EMA', 'bfloat16 mixed precision')) -> None	F5-style diffusion/flow-matching optimization profile.	no
EarlyStoppingCallback	class	voicehub.trainer_callback	(early_stopping_patience: 'int' = 1, early_stopping_threshold: 'float' = 0.0)	Stop after a metric fails to improve for a configured patience.	no
EpochGroupedBatchSampler	class	voicehub.training.asr_datasets	(dataset: 'ASRDataset', *, batch_size: 'int', seed: 'int', shuffle: 'bool', drop_last: 'bool') -> 'None'	Deterministic batches that never mix model-incompatible metadata.	no
EpochLengthBatchSampler	class	voicehub.training.tts_batching	(records: 'Sequence[Mapping[str, Any]]', config: 'TTSBatchingConfig | Mapping[str, Any]', *, batch_size: 'int', seed: 'int', shuffle: 'bool', drop_last: 'bool') -> 'None'	Epoch-addressable bucket or budget batches with exact-resume state.	no
EvalPrediction	class	voicehub.trainer_utils	(predictions: 'Any', label_ids: 'Any', inputs: 'Any | None' = None)	Container passed to a user-provided compute_metrics function.	no
FlowMatchingTrainingAdapter	class	voicehub.training.adapters	(model, spec: 'ModelTrainingSpec')	Continuous flow objective with strict native-loss preference.	no
FrameClassificationTrainingAdapter	class	voicehub.training.adapters	(model, spec: 'ModelTrainingSpec')	Frame classifier using the classification fallback and explicit mask.	no
get_asr_dataset_spec	callable	voicehub.training.asr_data_contracts	(model_type: 'str | None' = None, *, architecture: 'ASRDataArchitecture | str | None' = None) -> 'ASRDatasetSpec'	Return the inspectable ASR dataset contract for a model or architecture.	no
get_last_checkpoint	callable	voicehub.trainer_utils	(folder: 'str | Path') -> 'str | None'	Return the checkpoint with the greatest numeric global step.	no
get_training_spec	callable	voicehub.training.specs	(model_type: 'str') -> 'ModelTrainingSpec'	Resolve inference aliases and return one registered training profile.	no
get_training_strategy	callable	voicehub.training.strategy	(strategy: 'str | TrainingStrategy | None' = None) -> 'TrainingStrategy'	Resolve a strategy name or validate an existing strategy instance.	no
get_tts_dataset_spec	callable	voicehub.training.data_contracts	(model_type: 'str | None' = None, *, architecture: 'TTSDataArchitecture | str | None' = None) -> 'TTSDatasetSpec'	Return the inspectable dataset contract for a model or architecture.	no
get_tts_training_optimization_profile	callable	voicehub.training.tts_optimization	(model_type_or_architecture: 'str | TTSDataArchitecture') -> 'TTSTrainingOptimizationProfile'	Resolve the applicable special profile for a model or architecture.	no
IntervalStrategy	enum	voicehub.trainer_utils	(value)	When a recurring Trainer action should run.	no
list_asr_dataset_specs	callable	voicehub.training.asr_data_contracts	() -> 'tuple[ASRDatasetSpec, ...]'	Return one model-specific dataset contract for every ASR profile.	no
list_training_specs	callable	voicehub.training.specs	(*, task: 'SpeechTask | str | None' = <SpeechTask.TEXT_TO_SPEECH: 'text-to-speech'>, support: 'TrainingSupport | str | None' = None) -> 'tuple[ModelTrainingSpec, ...]'	List profiles by task and, optionally, support boundary.	no
list_training_strategies	callable	voicehub.training.strategy	() -> 'tuple[str, ...]'	Return registered execution strategy names.	no
list_tts_dataset_specs	callable	voicehub.training.data_contracts	() -> 'tuple[TTSDatasetSpec, ...]'	Return one model-specific dataset contract for every TTS profile.	no
llm_tts_acceleration_plan	callable	voicehub.training.tts_acceleration	(*, kernel_backend: 'str' = 'auto', attention_policy: 'str' = 'auto', use_torch_compile: 'bool' = True, compile_backend: 'str' = 'inductor', compile_mode: 'str | None' = 'max-autotune-no-cudagraphs', compile_fullgraph: 'bool' = False, compile_dynamic: 'bool | None' = True, compile_requirement: 'str' = 'auto') -> 'tuple[OptimizationPass, ...]'	Build the LLM-TTS SwiGLU, FlashAttention-4, and compile plan.	no
LLMTTSOptimizationConfig	class	voicehub.training.tts_optimization	(recipe: 'str' = 'conversationtts', learning_rate: 'float' = 1e-05, weight_decay: 'float' = 0.05, adam_beta1: 'float' = 0.9, adam_beta2: 'float' = 0.95, adam_epsilon: 'float' = 1e-08, warmup_ratio: 'float' = 0.03, max_grad_norm: 'float' = 1.0, gradient_accumulation_steps: 'int' = 1, per_device_train_batch_size: 'int' = 64, token_budget: 'int' = 7500, max_sequence_length: 'int' = 2048, length_field: 'str' = 'num_tokens', use_bf16: 'bool' = True, gradient_checkpointing: 'bool' = False, fused_adamw: 'bool' = True, lr_scheduler_type: 'str' = 'cosine', source_url: 'str' = 'https://github.com/Audio-Foundation-Models/ConversationTTS/tree/b3851f70c2dc0d35ba609734b08915637fe2a733', techniques: 'tuple[str, ...]' = ('offline frozen-codec tokenization', 'token-budget length batching', 'scaled dot-product attention', 'norm-excluded weight decay', 'fused AdamW when CUDA supports it', 'cosine decay with warmup', 'bfloat16 mixed precision')) -> None	Codec-language-model optimization profile.	no
masked_diffusion_regression_loss	callable	voicehub.training.objectives.diffusion	(predictions: 'Any', targets: 'Any', *, mask: 'Any | None' = None, weights: 'Any | None' = None, loss_type: 'str' = 'mse', reduction: 'str' = 'mean')	Compute exact-shape masked diffusion or flow regression.	no
ModelTrainingSpec	class	voicehub.training.specs	(model_type: 'str', family: 'TrainingFamily | str', module_paths: 'tuple[str, ...]' = ('model', 'model.model'), component_paths: 'tuple[str, ...]' = (), label_names: 'tuple[str, ...]' = ('labels', 'targets', 'target'), prediction_keys: 'tuple[str, ...]' = ('logits', 'predictions', 'audio_values', 'waveform'), loss_keys: 'tuple[str, ...]' = ('loss', 'total_loss'), loss_weights: 'tuple[tuple[str, float], ...]' = (), regression_loss: 'str' = 'mse', source_entrypoints: 'tuple[str, ...]' = (), native_training: 'bool' = False, separate_optimizers: 'bool' = False, support: 'TrainingSupport' = <TrainingSupport.PREPROCESSED: 'preprocessed'>, phases: 'tuple[TrainingPhaseSpec, ...]' = (), default_phase: 'str | None' = None, fallback_objective: 'str | None' = None, recipe_kind: 'TrainingRecipeKind' = <TrainingRecipeKind.SINGLE_PHASE: 'single-phase'>, allow_module_discovery: 'bool' = False, training_default_model_name_or_path: 'str | None' = None, field_schemas: 'Mapping[str, Any]' = <factory>, task: 'SpeechTask | str' = <SpeechTask.TEXT_TO_SPEECH: 'text-to-speech'>, adapter_factory: 'str | None' = None, dataset_factory: 'str | None' = None, dataset_spec_factory: 'str | None' = None, tokenizer_paths: 'tuple[str, ...]' = ('tokenizer', 'model.tokenizer'), optimization_profile_factory: 'str | None' = None) -> None	Everything a training adapter needs to expose a source runtime.	no
multi_codebook_cross_entropy	callable	voicehub.training.objectives.token	(logits: 'Any', labels: 'Any', *, loss_mask: 'Any | None' = None, ignore_index: 'int' = -100, causal_shift: 'bool' = False, sequence_dim: 'int' = -1, codebook_weights: 'Any | None' = None, codebook_dim: 'int' = 1, reduction: 'str' = 'mean')	Compute exact-shape token cross entropy for one or more codebooks.	no
OptimizerBundle	class	voicehub.training.optimization	(optimizers: 'Mapping[str, object]')	Present multiple named optimizers through a routed optimizer API.	no
PredictionOutput	class	voicehub.trainer_utils	(predictions: ForwardRef('Any'), label_ids: ForwardRef('Any'), metrics: ForwardRef('dict[str, float]'))	Predictions, references, and metrics returned by Trainer.predict.	no
register_training_alias	callable	voicehub.training.specs	(alias: 'str', model_type: 'str', *, exist_ok: 'bool' = False) -> 'None'	Register a training-only alias for a canonical profile.	no
register_training_spec	callable	voicehub.training.specs	(spec: 'ModelTrainingSpec', *, exist_ok: 'bool' = False, aliases: 'Iterable[str]' = ()) -> 'None'	Register or explicitly replace a training profile.	no
register_training_strategy	callable	voicehub.training.strategy	(name: 'str', factory: 'Callable[[], TrainingStrategy] | type[TrainingStrategy]', *, exist_ok: 'bool' = False) -> 'None'	Register a lazily constructed execution strategy.	no
RNNTTrainingAdapter	class	voicehub.training.adapters	(model, spec: 'ModelTrainingSpec')	RNN-T adapter that requires the backend's transducer objective.	no
SchedulerBundle	class	voicehub.training.optimization	(schedulers: 'Mapping[str, object]')	Keep one scheduler aligned with each named optimizer.	no
SchedulerType	enum	voicehub.trainer_utils	(value)	Learning-rate schedules implemented without external trainer packages.	no
Seq2SeqTrainingAdapter	class	voicehub.training.adapters	(model, spec: 'ModelTrainingSpec')	Teacher-forced sequence objective without causal label shifting.	no
set_seed	callable	voicehub.trainer_utils	(seed: 'int') -> 'None'	Seed Python and the native PyTorch compute backend.	no
SpeechDataset	class	voicehub.training.dataset_base	(records: 'Iterable[Mapping[str, Any]]', *, required_fields: 'Iterable[str]' = (), transform: 'Callable[[dict[str, Any]], Mapping[str, Any]] | None' = None)	An immutable-indexed view over validated speech records.	no
SpeechSeq2SeqTrainingAdapter	class	voicehub.training.adapters	(model, spec: 'ModelTrainingSpec')	Speech encoder-decoder adapter with an optional token CE fallback.	no
TDTTrainingAdapter	class	voicehub.training.adapters	(model, spec: 'ModelTrainingSpec')	Token-and-duration transducer adapter requiring its native objective.	no
TorchTrainingStrategy	class	voicehub.training.strategy	()	Default lazy PyTorch execution strategy.	no
Trainer	class	voicehub.trainer	(model=None, args: 'TrainingArguments | None' = None, data_collator: 'Callable[[list[Any]], dict[str, Any]] | None' = None, train_dataset=None, eval_dataset=None, processing_class=None, model_init: 'Callable[[], Any] | None' = None, compute_loss_func: 'Callable[[Any, Any, int | None], Any] | None' = None, compute_metrics: 'Callable[[EvalPrediction], dict[str, float]] | None' = None, callbacks: 'list[TrainerCallback | type[TrainerCallback]] | None' = None, optimizers: 'tuple[Any | None, Any | None]' = (None, None), optimizer_cls_and_kwargs: 'tuple[type, dict[str, Any]] | None' = None, preprocess_logits_for_metrics: 'Callable[[Any, Any], Any] | None' = None, training_adapter: 'BaseTrainingAdapter | None' = None, optimizer_factory: 'Callable[[str, list[tuple[str, Any]], TrainingArguments], Any] | None' = None, scheduler_factory: 'Callable[[str, Any, int, TrainingArguments], Any] | None' = None, training_strategy: 'str | TrainingStrategy | None' = None, optimization_plan: 'str | OptimizationPass | Iterable[str | OptimizationPass] | None' = None, optimization_config=None, optimization_context: 'OptimizationContext | None' = None, optimization_pass_registry: 'OptimizationPassRegistry | None' = None)	Complete single-process PyTorch train/evaluate/checkpoint loop.	no
TrainerCallback	class	voicehub.trainer_callback	()	Base class for non-invasive Trainer customizations.	no
TrainerControl	class	voicehub.trainer_callback	(should_training_stop: 'bool' = False, should_epoch_stop: 'bool' = False, should_save: 'bool' = False, should_evaluate: 'bool' = False, should_log: 'bool' = False) -> None	Boolean signals returned by callbacks to control the loop.	no
TrainerState	class	voicehub.trainer_callback	(epoch: 'float | None' = None, global_step: 'int' = 0, max_steps: 'int' = 0, logging_steps: 'int' = 500, eval_steps: 'int' = 500, save_steps: 'int' = 500, num_train_epochs: 'float' = 0.0, total_flos: 'float' = 0.0, log_history: 'list[dict[str, Any]]' = <factory>, best_metric: 'float | None' = None, best_model_checkpoint: 'str | None' = None, train_epoch: 'int | None' = None, train_batch_cursor: 'int | None' = None, is_local_process_zero: 'bool' = True, is_world_process_zero: 'bool' = True, is_hyper_param_search: 'bool' = False, trial_name: 'str | None' = None) -> None	Mutable, serializable progress state shared with callbacks.	no
TrainingArguments	class	voicehub.training_args	(output_dir: 'str' = 'trainer_output', overwrite_output_dir: 'bool' = False, do_train: 'bool' = False, do_eval: 'bool' = False, eval_strategy: 'IntervalStrategy | str' = <IntervalStrategy.NO: 'no'>, prediction_loss_only: 'bool' = False, per_device_train_batch_size: 'int' = 8, per_device_eval_batch_size: 'int' = 8, gradient_accumulation_steps: 'int' = 1, eval_accumulation_steps: 'int | None' = None, learning_rate: 'float' = 5e-05, weight_decay: 'float' = 0.0, adam_beta1: 'float' = 0.9, adam_beta2: 'float' = 0.999, adam_epsilon: 'float' = 1e-08, adamw_fused: 'bool' = False, adamw_torch_compile: 'bool' = False, max_grad_norm: 'float' = 1.0, num_train_epochs: 'float' = 3.0, max_steps: 'int' = -1, lr_scheduler_type: 'SchedulerType | str' = <SchedulerType.LINEAR: 'linear'>, warmup_ratio: 'float' = 0.0, warmup_steps: 'int' = 0, lr_scheduler_gamma: 'float' = 1.0, logging_strategy: 'IntervalStrategy | str' = <IntervalStrategy.STEPS: 'steps'>, logging_steps: 'int' = 500, logging_first_step: 'bool' = False, eval_steps: 'int | None' = None, save_strategy: 'IntervalStrategy | str' = <IntervalStrategy.STEPS: 'steps'>, save_steps: 'int' = 500, save_total_limit: 'int | None' = None, seed: 'int' = 42, data_seed: 'int | None' = None, dataloader_drop_last: 'bool' = False, dataloader_num_workers: 'int' = 0, dataloader_pin_memory: 'bool' = True, remove_unused_columns: 'bool' = True, label_names: 'list[str]' = <factory>, load_best_model_at_end: 'bool' = False, metric_for_best_model: 'str | None' = None, greater_is_better: 'bool | None' = None, gradient_checkpointing: 'bool' = False, fp16: 'bool' = False, bf16: 'bool' = False, use_cpu: 'bool' = False, disable_tqdm: 'bool' = True, report_to: 'list[str] | str' = <factory>, run_name: 'str | None' = None, wandb_project: 'str | None' = None, wandb_entity: 'str | None' = None, wandb_group: 'str | None' = None, wandb_tags: 'list[str]' = <factory>, wandb_notes: 'str | None' = None, wandb_mode: 'str | None' = None, wandb_log_model: 'str | bool' = False, evaluation_strategy: 'IntervalStrategy | str | None' = None) -> None	Arguments controlling a single-process PyTorch training run.	no
TrainingContext	class	voicehub.training.contracts	(phase: 'TrainingPhaseSpec', inputs: 'Mapping[str, Any]', step: 'int | None' = None, epoch: 'float | None' = None, is_training: 'bool' = True, metadata: 'Mapping[str, Any]' = <factory>) -> None	Runtime context passed through phase preparation and execution hooks.	no
TrainingFamily	enum	voicehub.training.specs	(value)	Built-in objective and optimization shapes.	no
TrainingPhaseKind	enum	voicehub.training.contracts	(value)	Semantic role of a phase within a training recipe.	no
TrainingPhaseSpec	class	voicehub.training.contracts	(name: 'str', component_paths: 'tuple[str, ...]' = (), optimizer_names: 'tuple[str, ...]' = (), forward_component: 'str | None' = None, forward_method: 'str' = 'forward', label_names: 'tuple[str, ...]' = ('labels', 'targets', 'target'), prediction_keys: 'tuple[str, ...]' = ('logits', 'predictions', 'audio_values', 'waveform'), loss_keys: 'tuple[str, ...]' = ('loss', 'total_loss'), loss_weights: 'tuple[tuple[str, float], ...]' = (), input_aliases: 'tuple[tuple[str, str], ...]' = (), required_inputs: 'tuple[str, ...]' = (), frequency: 'int' = 1, offset: 'int' = 0, fallback_objective: 'str | None' = None, kind: 'TrainingPhaseKind' = <TrainingPhaseKind.OBJECTIVE: 'objective'>, detach_inputs: 'tuple[str, ...]' = (), frozen_component_paths: 'tuple[str, ...]' = (), optimizer_step_after_phase: 'bool' = False) -> None	One independently callable and optimizable phase of a TTS recipe.	no
TrainingRecipeKind	enum	voicehub.training.contracts	(value)	High-level orchestration shape of a model training profile.	no
TrainingStrategy	class	voicehub.training.strategy	()	Framework-facing operations used by :class:voicehub.Trainer.	no
TrainingSupport	enum	voicehub.training.contracts	(value)	How completely a model's source integration supports fine-tuning.	no
TrainOutput	class	voicehub.trainer_utils	(global_step: ForwardRef('int'), training_loss: ForwardRef('float'), metrics: ForwardRef('dict[str, float]'))	Return value of :meth:voicehub.Trainer.train.	no
TTSBatchingConfig	class	voicehub.training.tts_batching	(strategy: 'TTSBatchingStrategy | str', length_field: 'str', length_multiplier: 'float' = 1.0, bucket_boundaries: 'tuple[int, ...]' = (), max_batch_units: 'int | None' = None, max_samples: 'int | None' = None, max_sequence_length: 'int | None' = None, budget_mode: 'str' = 'sum') -> None	Serializable length batching settings attached to a TTSDataset.	no
TTSBatchingStrategy	enum	voicehub.training.tts_batching	(value)	Supported variable-length batching policies.	no
TTSDataArchitecture	enum	voicehub.training.data_contracts	(value)	Canonical source-data layouts used by TTS training recipes.	no
TTSDataReadiness	enum	voicehub.training.data_contracts	(value)	How far VoiceHub owns a model's dataset preparation path.	no
TTSDataset	class	voicehub.training.tts_datasets	(records: 'Iterable[Mapping[str, Any]]', *, model_type: 'str | None' = None, architecture: 'TTSDataArchitecture | str | None' = None, root: 'str | PathLike[str] | None' = None, aliases: 'Mapping[str, str] | None' = None, validate: 'bool' = True, validate_files: 'bool' = False, batching: 'TTSBatchingConfig | Mapping[str, Any] | None' = None, transform: 'Callable[[dict[str, Any]], Mapping[str, Any]] | None' = None, transform_fingerprint: 'str | None' = None)	Portable TTS records loaded from mappings, JSON/JSONL, CSV, or TSV.	no
TTSDatasetSpec	class	voicehub.training.data_contracts	(architecture: 'TTSDataArchitecture', variants: 'tuple[TTSRecordVariant, ...]', model_type: 'str | None' = None, sample_rate: 'int | None' = None, description: 'str' = '', readiness: 'TTSDataReadiness | None' = None, training_support: 'str | None' = None, field_aliases: 'Mapping[str, str] | tuple[tuple[str, str], ...]' = ()) -> None	Inspectable source-data contract for one TTS architecture or model.	no
TTSFieldSchema	class	voicehub.training.collators	(sequence_dim: 'int' = 0, padding_value: 'float | int | None' = None, padding_side: 'str' = 'right', length_field: 'str | None' = None, mask_field: 'str | None' = None, pad_to_multiple_of: 'int | None' = None, allow_missing: 'bool' = False) -> None	Describe the variable-length dimension of one training field.	no
TTSOptimizationProfile	type alias	voicehub.training.tts_optimization	type alias	Union of the public TTS training optimization profiles.	no
TTSRecordVariant	class	voicehub.training.data_contracts	(name: 'str', required_fields: 'tuple[str, ...]' = (), one_of: 'tuple[tuple[str, ...], ...]' = (), at_most_one_of: 'tuple[tuple[str, ...], ...]' = (), forbidden_fields: 'tuple[str, ...]' = (), requires: 'tuple[tuple[str, tuple[str, ...]], ...]' = (), requires_one_of: 'tuple[tuple[str, tuple[str, ...]], ...]' = (), description: 'str' = '', preprocessed: 'bool' = False) -> None	One accepted source or preprocessed record shape.	no
TTSTrainingOptimizationProfile	type alias	voicehub.training.tts_optimization	type alias	Union of model-specific TTS training optimization profiles.	no
unregister_training_alias	callable	voicehub.training.specs	(alias: 'str', *, missing_ok: 'bool' = False) -> 'str | None'	Remove an alias and return its former canonical target.	no
unregister_training_spec	callable	voicehub.training.specs	(model_type: 'str', *, missing_ok: 'bool' = False) -> 'ModelTrainingSpec | None'	Remove and return a dynamically registered (or built-in) profile.	no
unregister_training_strategy	callable	voicehub.training.strategy	(name: 'str') -> 'None'	Remove a custom strategy registration.	no
UpstreamNativeTrainingAdapter	class	voicehub.training.adapters	(model, spec: 'ModelTrainingSpec')	Require the integrated source runtime to return its native objective.	no
vits_acceleration_plan	callable	voicehub.training.tts_acceleration	(*, kernel_backend: 'str' = 'auto', use_torch_compile: 'bool' = True, compile_backend: 'str' = 'inductor', compile_mode: 'str | None' = None, compile_fullgraph: 'bool' = False, compile_dynamic: 'bool | None' = None, compile_requirement: 'str' = 'auto', cuda_graphs: 'VITSCUDAGraphPolicy | str | bool' = <VITSCUDAGraphPolicy.DISABLED: 'disabled'>) -> 'tuple[OptimizationPass, ...]'	Build VITS's gated-WaveNet-kernel and compile training plan.	no
vits_discriminator_loss	callable	voicehub.training.objectives.vits	(real_scores: 'Any', fake_scores: 'Any', *, masks: 'Any | None' = None) -> 'VITSDiscriminatorLoss'	Compute the standard VITS least-squares discriminator objective.	no
vits_feature_matching_loss	callable	voicehub.training.objectives.vits	(real_features: 'Any', fake_features: 'Any', *, masks: 'Any | None' = None, scale: 'float' = 2.0, detach_real: 'bool' = True)	Compute VITS feature matching across discriminator feature pyramids.	no
vits_generator_adversarial_loss	callable	voicehub.training.objectives.vits	(fake_scores: 'Any', *, masks: 'Any | None' = None)	Compute the standard VITS least-squares generator objective.	no
vits_kl_loss	callable	voicehub.training.objectives.vits	(posterior_latents: 'Any', posterior_log_scale: 'Any', prior_mean: 'Any', prior_log_scale: 'Any', *, mask: 'Any | None' = None)	Compute the diagonal-Gaussian KL component used by standard VITS.	no
VITSCUDAGraphPolicy	enum	voicehub.training.tts_acceleration	(value)	CUDA-graph policy for shape-bucketed VITS execution.	no
VITSDiscriminatorLoss	class	voicehub.training.objectives.vits	(loss: 'Any', real_losses: 'tuple[Any, ...]', fake_losses: 'tuple[Any, ...]') -> None	Total and per-discriminator least-squares GAN losses.	no
VITSOptimizationConfig	class	voicehub.training.tts_optimization	(learning_rate: 'float' = 0.0002, weight_decay: 'float' = 0.01, adam_beta1: 'float' = 0.8, adam_beta2: 'float' = 0.99, adam_epsilon: 'float' = 1e-09, lr_decay_per_epoch: 'float' = 0.999875, max_grad_norm: 'float' = 1.0, per_device_train_batch_size: 'int' = 64, length_field: 'str' = 'num_frames', bucket_boundaries: 'tuple[int, ...]' = (32, 300, 400, 500, 600, 700, 800, 900, 1000), use_fp16: 'bool' = True, fused_adamw: 'bool' = True, compile_adamw: 'bool' = False, source_url: 'str' = 'https://github.com/jaywalnut310/vits/tree/2e561ba58618d021b5b8323d3765880f7e0ecfdb', techniques: 'tuple[str, ...]' = ('posterior-latent segment decoding', 'sequential discriminator and generator updates', 'spectrogram-length bucket batching', 'separate fused AdamW optimizers when CUDA supports them', 'optional torch.compile optimizer-step fusion', 'epoch-normalized exponential learning-rate decay', 'automatic mixed precision')) -> None	Original-VITS optimizer, segment, and length-bucket profile.	no
VITSTrainingAdapter	class	voicehub.training.adapters	(model, spec: 'ModelTrainingSpec')	Phase-aware VITS/GAN adapter.	no
WandbCallback	class	voicehub.integrations	() -> 'None'	Log VoiceHub training runs to Weights & Biases.	no

Optimization and codecs

Export	Kind	Canonical module	Signature	Summary	Lazy
AudioAutoencoderView	class	voicehub.components.audio.codecs.base	(codec: 'Any', encoder: 'Any | None', bottleneck: 'Any | None', decoder: 'Any | None', encoder_attribute: 'str | None' = None, bottleneck_attribute: 'str | None' = None, decoder_attribute: 'str | None' = None) -> None	Non-owning encoder/bottleneck/decoder references for one codec.	yes
AudioCodec	class	voicehub.components.audio.codecs.base	(args, *kwargs)	Minimal structural protocol implemented by waveform codecs.	yes
AudioCodecComponentView	class	voicehub.components.audio.codecs.base	(codec: 'Any', encoder: 'Any | None', bottleneck: 'Any | None', decoder: 'Any | None', encoder_attribute: 'str | None' = None, bottleneck_attribute: 'str | None' = None, decoder_attribute: 'str | None' = None) -> None	Non-owning encoder/bottleneck/decoder references for one codec.	yes
capture_codec_cuda_graph	callable	voicehub.optimization.codecs	(codec: 'Any', example_args: 'tuple[Any, ...] | Any', *, example_kwargs: 'Mapping[str, Any] | None' = None, target: 'str | Callable[..., Any]' = 'auto', decoder_only: 'bool | None' = None, epsilon: 'Tensor | None' = None, stochastic_vae: 'bool | None' = None, warmup_steps: 'int' = 3, clone_outputs: 'bool' = True) -> 'CodecCUDAGraphRunner'	Capture one fixed-shape codec call with graph-aware VAE randomness.	yes
CODEC_CATALOG	constant	voicehub.components.audio.codecs.catalog	constant	Immutable catalog of the registered shared audio codecs.	yes
codec_is_stochastic_vae	callable	voicehub.components.audio.codecs.base	(codec: 'Any', *, view: 'AudioCodecComponentView | None' = None) -> 'bool'	Conservatively identify codecs whose encode/forward path samples a VAE.	yes
codec_target_is_stochastic	callable	voicehub.components.audio.codecs.base	(codec: 'Any', target: 'str') -> 'bool'	Resolve stochasticity for one callable codec boundary.	yes
CodecCatalogEntry	class	voicehub.components.audio.codecs.catalog	(codec_id: 'str', family: 'str', variant: 'str', owners: 'tuple[CodecOwnerBinding, ...]', representation: 'CodecRepresentation', integration: 'CodecIntegration', stages: 'CodecStageManifest', stochastic_vae: 'bool', separable_autoencoder: 'bool', implementation_paths: 'tuple[str, ...]', primitives: 'CodecPrimitiveManifest', optimization: 'CodecOptimizationManifest', aliases: 'tuple[str, ...]' = (), gaps: 'tuple[str, ...]' = ()) -> None	One canonical codec family/variant and all active TTS owners.	yes
CodecCodeBatch	type alias	voicehub.components.audio.codecs.base	type alias	Dense or ragged codec-token batch accepted by shared codec helpers.	yes
CodecCompileComponent	enum	voicehub.optimization.codecs	(value)	Structural codec boundary selected for compilation.	yes
CodecCompilePolicy	enum	voicehub.optimization.codecs	(value)	Whether compilation may fall back, must work, or is disabled.	yes
CodecCUDAGraphCaptureError	exception	voicehub.optimization.codecs	inherited constructor	A codec call cannot be captured with fixed-shape CUDA Graphs.	yes
CodecCUDAGraphRunner	class	voicehub.optimization.codecs	(graph: 'Any', static_args: 'tuple[Any, ...]', static_kwargs: 'Mapping[str, Any]', static_output: 'Any', *, target_name: 'str', clone_outputs: 'bool', stochastic_target: 'bool') -> 'None'	Fixed-shape CUDA-graph replay with owned static input buffers.	yes
CodecIntegration	enum	voicehub.components.audio.codecs.catalog	(value)	Where the codec graph sits relative to its owning TTS architecture.	yes
CodecKernelBackend	enum	voicehub.kernels.codecs	(value)	Implementation families understood by codec-only kernel policies.	yes
CodecKernelPass	class	voicehub.optimization.codec_accelerators	(*, backend: 'CodecKernelBackend | str' = <CodecKernelBackend.AUTO: 'auto'>) -> 'None'	Configure only modules exposing the codec-specific selector protocol.	yes
CodecOptimizationCompatibilityError	exception	voicehub.optimization.codecs	inherited constructor	A requested codec optimization has no safe structural target.	yes
CodecOptimizationConfig	class	voicehub.optimization.codecs	(policy: 'CodecOptimizationPolicy | str' = <CodecOptimizationPolicy.EXACT: 'exact'>, kernel_backend: 'CodecKernelBackend | str' = <CodecKernelBackend.AUTO: 'auto'>, compile: 'CodecCompilePolicy | str | bool' = <CodecCompilePolicy.AUTO: 'auto'>, compile_components: 'tuple[CodecCompileComponent | str, ...] | CodecCompileComponent | str' = (<CodecCompileComponent.AUTO: 'auto'>,), compile_config: 'TorchCompileConfig | Mapping[str, Any] | None' = None) -> None	Serializable codec optimization settings.	yes
CodecOptimizationPlan	class	voicehub.optimization.codecs	(config: 'CodecOptimizationConfig', context: 'OptimizationContext', passes: 'tuple[OptimizationPass, ...]', compile_targets: 'tuple[OptimizationCompileTarget, ...]', decisions: 'tuple[CodecOptimizationDecision, ...]', target_type: 'str', _target_identity: 'int') -> None	Resolved pass order and immutable structural decisions for one codec.	yes
CodecOptimizationPolicy	enum	voicehub.optimization.codecs	(value)	Numerical/semantic fidelity allowed by one optimization plan.	yes
CodecOptimizationResult	class	voicehub.optimization.codecs	(plan: 'CodecOptimizationPlan', application: 'OptimizationResult') -> None	Applied codec plan with reversible access to the eager model.	yes
CodecRepresentation	enum	voicehub.components.audio.codecs.catalog	(value)	The tensor representation crossing the TTS/codec boundary.	yes
CodecStageAvailability	enum	voicehub.components.audio.codecs.catalog	(value)	How one encoder, quantizer, or decoder stage is implemented.	yes
coerce_codec_codes	callable	voicehub.components.audio.codecs.base	(value: 'CodecCodeBatch | Tensor | Iterable[Tensor]') -> 'CodecCodeBatch'	Normalize common code containers without changing their tensors.	yes
DenseCodecCodes	class	voicehub.components.audio.codecs.base	(values: 'Tensor', lengths: 'Tensor | None' = None) -> None	A dense [batch, codebook, frame] discrete-code batch.	yes
diffusion_cache_request	callable	voicehub.optimization.diffusion_cache	(model: 'Any') -> 'Iterator[Any]'	Isolate every cache target for one public generation request.	yes
diffusion_cache_summary	callable	voicehub.optimization.diffusion_cache	(model: 'Any', *, details: 'bool' = False) -> 'dict[str, dict[str, Any]]'	Return cache telemetry for every architecture-owned target in a model.	yes
DiffusionArchitectureKind	enum	voicehub.optimization.diffusion	(value)	The diffusion or flow formulation used by an active TTS graph.	yes
DiffusionCacheConfig	class	voicehub.optimization.diffusion_cache	(method: 'DiffusionCacheMethod | str' = <DiffusionCacheMethod.DBCACHE: 'dbcache'>, front_blocks: 'int' = 1, back_blocks: 'int' = 0, residual_diff_threshold: 'float' = 0.08, warmup_steps: 'int' = 2, warmup_interval: 'int' = 1, max_cached_steps: 'int' = -1, max_consecutive_cached_steps: 'int' = 3, max_accumulated_relative_error: 'float | None' = None, predictor: 'DiffusionCachePredictor | str' = <DiffusionCachePredictor.REUSE: 'reuse'>, taylor_order: 'int' = 1, compute_step_mask: 'tuple[bool, ...]' = (), compute_step_policy: 'DiffusionCacheStepPolicy | str' = <DiffusionCacheStepPolicy.DYNAMIC: 'dynamic'>, num_inference_steps: 'int | None' = None, force_refresh_step_hint: 'int | None' = None, force_refresh_step_policy: 'DiffusionCacheRefreshPolicy | str' = <DiffusionCacheRefreshPolicy.ONCE: 'once'>, probe_downsample_factor: 'int' = 1, metrics_history_size: 'int' = 256, synchronize_distributed: 'bool' = True, epsilon: 'float' = 1e-06) -> None	Serializable Cache-DiT-style settings for a VoiceHub DiT block list.	yes
DiffusionCacheMethod	enum	voicehub.optimization.diffusion_cache	(value)	Architecture-owned block-cache layouts.	yes
DiffusionCachePass	class	voicehub.optimization.diffusion_cache	(config: 'DiffusionCacheConfig | Mapping[str, Any] | None' = None) -> 'None'	Reversibly enable approximate block caching on declared DiT modules.	yes
DiffusionCachePolicy	enum	voicehub.optimization.diffusion_cache	(value)	Whether approximate diffusion caching is disabled, optional, or required.	yes
DiffusionCachePredictor	enum	voicehub.optimization.diffusion_cache	(value)	Approximation used for a skipped middle-block residual.	yes
DiffusionCacheRefreshPolicy	enum	voicehub.optimization.diffusion_cache	(value)	Whether a configured cache-refresh hint runs once or periodically.	yes
DiffusionCacheStepPolicy	enum	voicehub.optimization.diffusion_cache	(value)	How an explicit step-computation mask controls cache decisions.	yes
DiffusionGuidanceStrategy	enum	voicehub.optimization.diffusion_sampling	(value)	Classifier-free-guidance evaluation policy.	yes
DiffusionModelOptimizationSupport	class	voicehub.optimization.diffusion	(model_type: 'str', architecture: 'str', kind: 'DiffusionArchitectureKind', operations: 'tuple[DiffusionOperation, ...]', training: 'bool', distributed_training: 'bool', optimization_passes: 'tuple[str, ...]', sampling_techniques: 'tuple[str, ...]') -> None	One public TTS model's declared diffusion optimization surface.	yes
DiffusionOperation	enum	voicehub.optimization.diffusion	(value)	Optimization-relevant operations present in an active sampler.	yes
DiffusionPredictionCacheMethod	enum	voicehub.optimization.diffusion_sampling	(value)	Whole-model prediction cache selected at the sampler boundary.	yes
DiffusionSamplingConfig	class	voicehub.optimization.diffusion_sampling	(target_steps: 'int | None' = None, schedule: 'DiffusionScheduleStrategy | str' = <DiffusionScheduleStrategy.NATIVE: 'native'>, solver: 'DiffusionSolverStrategy | str' = <DiffusionSolverStrategy.NATIVE: 'native'>, stork_stages: 'int' = 9, guidance: 'DiffusionGuidanceStrategy | str' = <DiffusionGuidanceStrategy.NATIVE: 'native'>, guidance_start: 'float' = 0.0, guidance_end: 'float' = 1.0, adaptive_guidance_threshold: 'float' = 0.01, adaptive_guidance_warmup_steps: 'int' = 4, adaptive_guidance_patience: 'int' = 2, prediction_cache: 'DiffusionPredictionCacheMethod | str | bool' = <DiffusionPredictionCacheMethod.DISABLED: 'disabled'>, cache_interval: 'int' = 2, cache_warmup_steps: 'int' = 2, cache_max_consecutive_steps: 'int' = 2, cache_rel_l1_threshold: 'float' = 0.08, cache_error_budget: 'float' = 0.2, teacache_coefficients: 'tuple[float, ...]' = (), smoothcache_compute_step_mask: 'tuple[bool, ...]' = (), taylor_order: 'int' = 1, epsilon: 'float' = 1e-06) -> None	Serializable sampler-level acceleration configuration.	yes
DiffusionSamplingController	class	voicehub.optimization.diffusion_sampling	(config: 'DiffusionSamplingConfig')	Request-scoped schedule, guidance, and prediction-cache controller.	yes
DiffusionSamplingPass	class	voicehub.optimization.diffusion_sampling	(config: 'DiffusionSamplingConfig | Mapping[str, Any] | None' = None) -> 'None'	Reversibly enable sampler-level diffusion acceleration.	yes
DiffusionSamplingPolicy	enum	voicehub.optimization.diffusion_sampling	(value)	Whether sampler-level acceleration is disabled, optional, or required.	yes
DiffusionScheduleStrategy	enum	voicehub.optimization.diffusion_sampling	(value)	How a native schedule is rebuilt when target_steps is smaller.	yes
DiffusionSolverStrategy	enum	voicehub.optimization.diffusion_sampling	(value)	Sampler integration rule for a supplied velocity prediction.	yes
DiffusionStepContext	class	voicehub.optimization.diffusion_sampling	(index: 'int', total_steps: 'int', timestep: 'Tensor | float', next_timestep: 'Tensor | float', lane: 'str' = 'default', solver: 'str' = 'euler', stage: 'str' = 'main', outer_step: 'int | None' = None) -> None	Identity and solver state for one model evaluation.	yes
discover_codec_compile_targets	callable	voicehub.optimization.codecs	(codec: 'Any', *, mode: 'OptimizationMode | str' = <OptimizationMode.INFERENCE: 'inference'>, components: 'CodecCompileComponent | str | Iterable[CodecCompileComponent | str]' = <CodecCompileComponent.AUTO: 'auto'>) -> 'tuple[OptimizationCompileTarget, ...]'	Discover callable codec boundaries without requiring model changes.	yes
get_codec_entries_for_model	callable	voicehub.components.audio.codecs.catalog	(model_type: 'str') -> 'tuple[CodecCatalogEntry, ...]'	Return the single codec-family entry owned by an active LLM-TTS model.	yes
get_codec_entry	callable	voicehub.components.audio.codecs.catalog	(codec_id: 'str') -> 'CodecCatalogEntry'	Return one codec entry by canonical ID, family, alias, or owner model type.	yes
get_diffusion_model_optimization_support	callable	voicehub.optimization.diffusion	(model_type: 'str') -> 'DiffusionModelOptimizationSupport'	Return one registered diffusion/flow TTS model or fail explicitly.	yes
get_tts_optimization_config	callable	voicehub.optimization.tts	(target: 'str | Any', **overrides: 'Any') -> 'TTSOptimizationConfig'	Return a validated universal config for one registered TTS target.	no
get_tts_optimization_support	callable	voicehub.optimization.tts	(target: 'str | Any') -> 'TTSOptimizationSupport'	Return the statically declared optimization surface for one target.	no
get_vits_model_optimization_support	callable	voicehub.optimization.vits	(model_type: 'str') -> 'VITSModelOptimizationSupport'	Return one registered VITS model's support or fail explicitly.	no
list_codec_entries	callable	voicehub.components.audio.codecs.catalog	(*, model_type: 'str | None' = None, representation: 'CodecRepresentation | str | None' = None, integration: 'CodecIntegration | str | None' = None, has_gap: 'bool | None' = None) -> 'tuple[CodecCatalogEntry, ...]'	List entries in stable order with optional graph-free filters.	yes
list_diffusion_model_optimization_support	callable	voicehub.optimization.diffusion	() -> 'tuple[DiffusionModelOptimizationSupport, ...]'	List active public diffusion/flow TTS models by architecture traits.	yes
list_tts_optimization_support	callable	voicehub.optimization.tts	() -> 'tuple[TTSOptimizationSupport, ...]'	List optimization capabilities for every registered TTS model.	no
list_vits_model_optimization_support	callable	voicehub.optimization.vits	() -> 'tuple[VITSModelOptimizationSupport, ...]'	List registered models marked by architecture traits, not name checks.	no
OptimizationCompileTarget	class	voicehub.optimization.protocols	(label: 'str', owner: 'Any', attribute: 'str', component: 'str | None' = None) -> None	One method boundary that synthesis or training actually invokes.	no
OptimizationCompileTargetProvider	class	voicehub.optimization.protocols	(args, *kwargs)	Contract for runtimes whose executed boundary is not plain forward.	no
OptimizationModuleRoot	class	voicehub.optimization.protocols	(label: 'str', module: 'Any') -> None	One runtime-owned module tree inspected by selector passes.	no
OptimizationModuleRootProvider	class	voicehub.optimization.protocols	(args, *kwargs)	Contract for runtimes that expose more than one module tree.	no
OptimizationRuntimeProtocol	class	voicehub.optimization.protocols	(args, *kwargs)	Complete checkpoint-aware contract for composite speech runtimes.	no
optimize_codec	callable	voicehub.optimization.codecs	(codec: 'Any', config: 'CodecOptimizationConfig | Mapping[str, Any] | None' = None, *, mode: 'OptimizationMode | str' = <OptimizationMode.INFERENCE: 'inference'>, context: 'OptimizationContext | None' = None) -> 'CodecOptimizationResult'	Resolve and apply one reversible codec optimization plan.	yes
RaggedCodecCodes	class	voicehub.components.audio.codecs.base	(levels: 'tuple[Tensor, ...]', lengths: 'tuple[Tensor | None, ...] | None' = None, strides: 'tuple[int, ...] | None' = None) -> None	A hierarchical or multirate discrete-code batch.	yes
reset_diffusion_cache_metrics	callable	voicehub.optimization.diffusion_cache	(model: 'Any') -> 'int'	Reset telemetry on every cache target and return the target count.	yes
resolve_codec_optimization	callable	voicehub.optimization.codecs	(codec: 'Any', config: 'CodecOptimizationConfig | Mapping[str, Any] | None' = None, *, mode: 'OptimizationMode | str' = <OptimizationMode.INFERENCE: 'inference'>, context: 'OptimizationContext | None' = None) -> 'CodecOptimizationPlan'	Resolve a side-effect-free, structural codec optimization plan.	yes
resolve_tts_optimization	callable	voicehub.optimization.tts	(target: 'str | Any', config: 'TTSOptimizationConfig | Mapping[str, Any] | None' = None, *, mode: 'OptimizationMode | str' = <OptimizationMode.INFERENCE: 'inference'>, context: 'OptimizationContext | None' = None, registry: 'OptimizationPassRegistry | None' = None) -> 'TTSOptimizationPlan'	Resolve one safe, ordered policy for any registered TTS model.	no
separate_audio_codec	callable	voicehub.components.audio.codecs.base	(codec: 'Any', *, encoder: 'str | Any | None' = None, bottleneck: 'str | Any | None' = None, decoder: 'str | Any | None' = None) -> 'AudioCodecComponentView'	Create a non-owning component view of an existing codec graph.	yes
STORK2FlowSolver	class	voicehub.optimization.diffusion_solvers	(config: 'STORKFlowConfig | None' = None)	Stateful STORK-2/Taylor-1 integrator for one request and solver lane.	yes
STORKFlowConfig	class	voicehub.optimization.diffusion_solvers	(stages: 'int' = 9, accumulator_dtype: 'str' = 'float32') -> None	STORK-2 settings for a direct deterministic velocity field.	yes
TTSAttentionImplementation	enum	voicehub.optimization.tts	(value)	Attention policies understood by the built-in TTS resolver.	no
TTSCompilePolicy	enum	voicehub.optimization.tts	(value)	Whether compilation may fall back, must work, or is disabled.	no
TTSKernelBackend	enum	voicehub.optimization.tts	(value)	Custom-kernel policy for architecture-owned fused operations.	no
TTSOptimizationCompatibilityError	exception	voicehub.optimization.tts	inherited constructor	A requested implementation is not valid for a model or context.	no
TTSOptimizationConfig	class	voicehub.optimization.tts	(attn_implementation: 'TTSAttentionImplementation | str' = <TTSAttentionImplementation.AUTO: 'auto'>, kernel_backend: 'TTSKernelBackend | str' = <TTSKernelBackend.AUTO: 'auto'>, compile: 'TTSCompilePolicy | str | bool' = <TTSCompilePolicy.AUTO: 'auto'>, compile_config: 'TorchCompileConfig | Mapping[str, Any] | None' = None, diffusion_cache: 'DiffusionCachePolicy | str | bool' = 'disabled', diffusion_cache_config: 'DiffusionCacheConfig | Mapping[str, Any] | None' = None, diffusion_sampling: 'DiffusionSamplingPolicy | str | bool' = 'disabled', diffusion_sampling_config: 'DiffusionSamplingConfig | Mapping[str, Any] | None' = None, optimization_passes: 'tuple[str, ...]' = ()) -> None	Serializable, Transformers-style optimization settings for TTS.	no
TTSOptimizationDecision	class	voicehub.optimization.tts	(feature: 'str', requested: 'str', selected: 'str', implementation_pass: 'str | None', reason: 'str') -> None	One requested policy choice and its statically resolved outcome.	no
TTSOptimizationPlan	class	voicehub.optimization.tts	(config: 'TTSOptimizationConfig', context: 'OptimizationContext', support: 'TTSOptimizationSupport', passes: 'tuple[OptimizationPass, ...]', decisions: 'tuple[TTSOptimizationDecision, ...]') -> None	Resolved, ordered pass plan plus all fallback decisions.	no
TTSOptimizationResult	class	voicehub.optimization.tts	(plan: 'TTSOptimizationPlan', model: 'Any', application: 'OptimizationResult | None' = None) -> None	Universal result for both transformed and native-fallback plans.	no
TTSOptimizationSupport	class	voicehub.optimization.tts	(model_type: 'str | None', architecture: 'str | None', attention_implementations: 'tuple[str, ...]', kernel_backends: 'tuple[str, ...]', compile: 'bool', diffusion_cache: 'bool', diffusion_sampling: 'bool', diffusion_sampling_techniques: 'tuple[str, ...]', optimization_kinds: 'tuple[str, ...]') -> None	Architecture-level implementations available to one TTS model.	no
VITSArchitectureKind	enum	voicehub.optimization.vits	(value)	How a registered public model incorporates the VITS architecture.	no
VITSModelOptimizationSupport	class	voicehub.optimization.vits	(model_type: 'str', architecture: 'str', kind: 'VITSArchitectureKind', training: 'bool', distributed_training: 'bool', optimization_passes: 'tuple[str, ...]', kernel_operations: 'tuple[str, ...]') -> None	One public model's family role and shared optimization contract.	no

Policies, errors, and utilities

Export	Kind	Canonical module	Signature	Summary	Lazy
LLMBackendCompatibilityError	exception	voicehub.errors	inherited constructor	Raised when an engine cannot preserve an architecture's semantics.	no
LLMBackendError	exception	voicehub.errors	inherited constructor	Base exception for external language-model serving failures.	no
LLMBackendRequestError	exception	voicehub.errors	inherited constructor	Raised when an external serving request fails or is malformed.	no
ModelLicenseSpec	class	voicehub.policies.licensing	(model_type: 'str', license_id: 'str', commercial_use: 'bool | None', upstream: 'str', notice: 'str') -> None	License conditions attached to source or checkpoint artifacts.	no
OptionalDependencyError	exception	voicehub.errors	inherited constructor	Raised when the selected backend has not been installed.	no
SourceLicenseError	exception	voicehub.errors	inherited constructor	Raised when upstream source cannot legally be redistributed.	no
UnknownModelError	exception	voicehub.errors	inherited constructor	Raised when a model key is not registered.	no
VoiceHubError	exception	voicehub.errors	inherited constructor	Base exception for VoiceHub-specific failures.	no