`asr_vibevoice` model guide¶

Overview¶

asr_vibevoice is a VoiceHub automatic speech recognition integration. This page is generated from the model registry and its executable data and training contracts, so the documented support stays aligned with code. Open the asr_vibevoice Colab notebook.

Quickstart¶

python -m pip install voicehub

Install VoiceHub and the provider extra shown above.
Choose a checkpoint that matches this integration.
Place a supported recording at speech.wav.
Transcribe it and inspect both the full text and timed segments.

from voicehub import AutoModelForSpeechRecognition

model = AutoModelForSpeechRecognition.from_pretrained(
    'microsoft/VibeVoice-ASR-HF',
    model_type='asr_vibevoice',
    device="cuda",
    lazy_load=True,
)
output = model.transcribe("speech.wav")
print(output.text)
for segment in output.segments:
    print(segment.start, segment.end, segment.text)

Use only authorized recordings for reference voice, transcription, detection, or evaluation. The example selects a concrete device; verify checkpoint-specific hardware needs and pin an immutable revision before production use.

Supported tasks and capabilities¶

Property	Value
Task	Automatic speech recognition
Architecture	`vibevoice-asr`
Runtime	`VoiceHub-native`
Capabilities	`automatic-speech-recognition`, `multilingual`, `speaker-attribution`, `timestamps`, `hotwords`, `long-form`, `safetensors`, `fine-tuning`, `voicehub-native`, `native-runtime`
Reusable components	—

Data contract¶

Property	Value
Readiness	`integrated-raw`
Data architecture	`prompted-multimodal`
Sample rate	24,000 Hz
Contract getter	`get_asr_dataset_spec('asr_vibevoice')`

Variant	Required fields	One of	Boundary	Other rules
`segmented-audio`	`audio`, `segments`	—	Source	forbidden: text, transcription, transcript
`serialized-audio`	`audio`	text / transcription / transcript	Source	at most one: text / transcription / transcript; forbidden: segments
`vibevoice-model-ready`	`input_ids`, `attention_mask`, `input_values`, `padding_mask`, `labels`	—	Prepared	—

VibeVoice structured long-form ASR targets and multimodal prompt inputs. Follow the shared data workflow for manifest loading, audio validation, leakage-safe splits, and model-owned preprocessing.

Checkpoints, provenance, and license¶

Property	Value
Default checkpoint	`microsoft/VibeVoice-ASR-HF`
Checkpoint status	Registry default; pin an immutable revision for production and reproducible evidence
Implementation	`voicehub.models.asr_vibevoice.modeling_asr_vibevoice.VibeVoiceForSpeechRecognition`
Configuration	`voicehub.models.asr_vibevoice.configuration_asr_vibevoice.VibeVoiceASRConfig`
Source provenance	`voicehub/architectures/vibevoice/source/SOURCE.json`
License	Checkpoint-specific

No VoiceHub-specific license override is registered. Verify the checkpoint and upstream source terms before use.

The default checkpoint identifies the expected family, not every compatible variant. Confirm the selected checkpoint's revision, access terms, provenance, and license before downloading or redistributing it.

Optimization and training support¶

All public optimizations enter this model through the shared BaseSpeechModel lifecycle. Use available_optimization_passes() to discover the public pass registry, then apply, inspect, serialize, or restore a plan through the common model API. Application remains fail-closed when the active runtime or hardware cannot satisfy a pass.

Training contract¶

Property	Value
Support	`native`
Family	`speech-sequence-to-sequence`
Recipe	`single-phase`
Default phase	`speech_recognition`
Training checkpoint	`microsoft/VibeVoice-ASR-HF`
Native training graph	`yes`

Phase	Kind	Components	Required inputs	Loss keys
`speech_recognition`	objective	`model.model.multi_modal_projector`, `model.model.language_model`, `model.lm_head`	`input_ids`, `attention_mask`, `input_values`, `padding_mask`, `labels`	`loss`

The integration accepts its declared source or prepared contract directly. Call model.validate_training_support() before constructing a trainer. Follow the shared training workflow for a one-step smoke test, validation, checkpoint resume, optimization, and portable export.

Public API¶

Purpose	Public object
Discover	`get_model_spec('asr_vibevoice')`
Load and run	`AutoModelForSpeechRecognition`
Configure	`VibeVoiceASRConfig`
Model implementation	`VibeVoiceForSpeechRecognition`
Normalized output	`ASROutput`
Training contract	`get_training_spec('asr_vibevoice')`
Optimization lifecycle	`available_optimization_passes`, `apply_optimization_plan`, `optimization_manifest`, `restore_optimization_plan`

asr_vibevoice model guide¶