Перейти к содержанию

Models

VoiceHub model wrappers expose one pretrained lifecycle across text to speech, automatic speech recognition, and voice activity detection. Task-specific classes keep their input and output semantics explicit while sharing lazy loading, saving, training validation, and runtime-state transitions.

Create an unloaded model from configuration when you want to inspect or adjust the wrapper before allocating checkpoint-backed state:

from voicehub import AutoConfig, AutoModelForTextToSpeech

config = AutoConfig.for_model(
    "parlertts",
    name_or_path="parler-tts/parler-tts-mini-v1",
)
model = AutoModelForTextToSpeech.from_config(config)
assert not model.is_loaded

Use the task-specific auto classes for normal discovery and construction. See the Auto Classes catalog for every registered model and the full API reference for training, optimization, serving, and extension contracts.

PreTrainedSpeechModel

PreTrainedSpeechModel is the public marker shared by every pretrained speech wrapper. It deliberately does not implement a token-embedding or language-model utility mixin. Speech models may own waveform processors, acoustic encoders, codecs, vocoders, and streaming state, so those operations remain on the relevant task base or model integration.

The task bases expose the same core lifecycle:

Member Contract
from_pretrained(source, **kwargs) Restore configuration and portable VoiceHub metadata from a local path or Hub repository without loading weights unless lazy_load=False
load() Allocate the checkpoint-backed runtime and enter inference mode once
load_for_training() Validate the configured training path and restore a differentiable runtime
validate_training_support() Check the exact backend and checkpoint before model allocation
save_pretrained(directory, include_native_export=True) Save portable metadata and, when supported, a namespaced native artifact
is_loaded Report whether checkpoint-backed runtime state exists

Loading and configuration resolution are lazy at the package boundary. Inspecting the public classes or registry does not construct a model.

Task-specific pretrained models

The source implementations are split by input contract:

Base class Main input Primary call Normalized output
PreTrainedTTSModel Text plus optional conditioning generate() TTSOutput
PreTrainedASRModel Audio or an audio path transcribe() ASROutput
PreTrainedVADModel Audio or an audio path detect() VADOutput

PreTrainedAudioModel.forward() provides the common audio request path. ASR and VAD keep separate task classes so a wrapper cannot silently return the wrong output type. PreTrainedTTSModel.forward() similarly validates that the model integration returned TTSOutput.

Use AutoModel, AutoModelForTextToSpeech, AutoModelForSpeechRecognition, or AutoModelForVoiceActivityDetection instead of instantiating an abstract pretrained base directly.

Model outputs

The public output dataclasses live in voicehub/modeling_outputs.py. They provide named fields, integer and string indexing, keys(), to_dict(), and a compact tuple view without making a provider-specific object public.

Output Required fields Optional evidence
TTSOutput audio, sample_rate Output path and provider metadata
ASROutput text, ordered segments Language, duration, confidence, words, speakers, and metadata
VADOutput Ordered, non-overlapping speech segments Duration, sample rate, frame probabilities, and metadata 
from voicehub import TTSOutput

output = TTSOutput(audio=[0.0, 0.1, 0.0], sample_rate=24_000)
audio, sample_rate = output
assert output["sample_rate"] == sample_rate

Training uses SpeechTrainingOutput and its backward-compatible TTSTrainingOutput specialization. Their first populated value is loss, matching the shared Trainer contract while allowing speech-specific fields.

Loading, saving, and sharing

from_pretrained() accepts a VoiceHub artifact, a direct checkpoint when the task factory receives model_type, or a Hub repository. save_pretrained() writes only the portable files supported by the wrapper:

Artifact Purpose
config.json Serializable architecture and checkpoint identity
generation_config.json or task inference config Saved request defaults
processor_config.json Text or audio preprocessing metadata
model_state.safetensors Optional portable trained state
native/ Optional provider-native export owned by the integration

VoiceHub does not expose a public push_to_hub() method. Saving locally and uploading with the Hub client are separate operations until VoiceHub can guarantee one registry-wide upload contract for configuration, processors, portable state, native artifacts, provenance, and license files. This is an explicit support boundary, not an implied sharing pass.

For exact artifact and resume behavior, see Save, load, and resume boundaries.