Microsoft VibeVoice Hits 40,000 GitHub Stars as ASR-7B Joins HuggingFace Transformers
Microsoft’s VibeVoice repository reached 40,330 GitHub stars this week as its speech recognition model shipped in the HuggingFace Transformers library, completing a sequence of open-source releases that started with TTS in August 2025 and now covers ASR and real-time streaming synthesis.
The project is landing mainstream attention on the same day the Transformers integration has reached typical developer discovery channels — the point at which “research model you read about” becomes “model you can use in three lines of code.”
What’s in the family
VibeVoice ships three models with different roles:
VibeVoice-ASR-7B — long-form speech recognition. Handles up to 60 minutes of audio in a single forward pass. Output is structured: speaker (who), timestamp (when), and transcript (what). Supports customized hotwords for domain-specific vocabulary. Natively multilingual across 50+ languages. Available via Hugging Face Playground and now installable through transformers.
VibeVoice-Realtime-0.5B — streaming text-to-speech. 0.5B parameters, deployment-friendly size. First-token latency of approximately 300 milliseconds. Handles streaming text input and generates up to 10 minutes of continuous speech. Available on Colab.
VibeVoice-TTS-1.5B — the original TTS release from August 2025, an ICLR 2026 Oral. Accepted for multi-speaker long-form synthesis up to 90 minutes with four distinct speakers. Microsoft removed the model weights from the repository in September 2025 after discovering the tool was being used for voice cloning at scale inconsistent with the project’s stated research intent. The Colab playground remains disabled.
Architecture
The core innovation is a pair of continuous speech tokenizers operating at 7.5 Hz — acoustic and semantic. The ultra-low frame rate compresses audio sequences enough to feed them into a standard LLM for contextual reasoning, while a diffusion head on the output reconstructs high-fidelity acoustic detail. The result is a model that understands dialogue flow at the language level and generates speech quality at the signal level, without a multi-stage pipeline.
Traction
The 40K star count places VibeVoice in the same visibility tier as Whisper clones and open-source TTS projects that have built real developer ecosystems. The community has already moved: Vibing, an AI-powered input method for macOS and Windows, is now built on VibeVoice-ASR and available as an installer download.
The ICLR 2026 Oral for the TTS paper is the research credibility signal. Most open-source voice projects do not clear that bar.
What is missing
Microsoft has not re-released the TTS weights. The stated reason was misuse. The effect is that the generation side of the system — the part that produces voice output at scale — is not publicly available. VibeVoice-Realtime-0.5B fills part of that gap but is a different and smaller model. The 1.5B TTS Oral-accepted model is absent from the ecosystem.
For developers building ASR pipelines, voice transcription with diarization, or multilingual recognition into agentic systems, VibeVoice-ASR-7B is the relevant model and it is now accessible via standard Transformers tooling.