Instructions to use microsoft/VibeVoice-1.5B with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Transformers
How to use microsoft/VibeVoice-1.5B with Transformers:
# Use a pipeline as a high-level helper from transformers import pipeline pipe = pipeline("text-to-speech", model="microsoft/VibeVoice-1.5B")# Load model directly from transformers import AutoProcessor, AutoModelForTextToWaveform processor = AutoProcessor.from_pretrained("microsoft/VibeVoice-1.5B") model = AutoModelForTextToWaveform.from_pretrained("microsoft/VibeVoice-1.5B", device_map="auto") - Notebooks
- Google Colab
- Kaggle
Publish an official Transformers-native VibeVoice-1.5B-HF checkpoint
Proposal
Please publish an official Transformers-native TTS checkpoint as microsoft/VibeVoice-1.5B-HF, following the existing microsoft/VibeVoice-ASR-HF precedent.
The current microsoft/VibeVoice-1.5B@c00898d257e6b46004e3e2866a47534085fb685a should remain the immutable legacy/source-provenance repository. A separate -HF repository avoids changing its weight namespace and breaking existing consumers.
Reproducible conversion
A focused release-tooling patch is ready for microsoft/VibeVoice. It pins:
- official weights:
microsoft/VibeVoice-1.5B@c00898d257e6b46004e3e2866a47534085fb685a - canonical Transformers converter/native implementation:
huggingface/transformers@640a08a597034221ca1c4fc0c129cf0118179225 - tokenizer input:
Qwen/Qwen2.5-1.5B@8faed761d45a263340a0528343f099c05c9a4323
The tool downloads only the official Microsoft checkpoint and pinned tokenizer, invokes the canonical Transformers converter, emits the native config/tokenizer/processor/chat-template/generation assets, and writes a provenance manifest. It fails closed unless all 1,204 source tensors map one-to-one to 1,204 native tensors with identical shapes and dtypes. It deliberately has no upload option; publication remains a Microsoft owner action.
The patch is complete and independently reviewed. Its GitHub fork/PR handoff is temporarily blocked on Microsoft organization SAML authorization for the contributor token; I will link the upstream PR here once submitted from an authorized machine.
Acceptance
After owner publication, this should work without remote code or any sidecar repository:
from transformers import AutoModelForTextToWaveform, AutoProcessor
model_id = "microsoft/VibeVoice-1.5B-HF"
processor = AutoProcessor.from_pretrained(model_id, trust_remote_code=False)
model = AutoModelForTextToWaveform.from_pretrained(
model_id,
dtype="auto",
trust_remote_code=False,
)
The original model card can then link to the official -HF repository while preserving the original checkpoint and revision history.