Verification conclusion first
The repository is real and extremely hot: official Microsoft, 54,542 stars, MIT license, still updated on 09-03. The noted "14GB VRAM" checks out arithmetically: a 7B model in bf16 is exactly 14GB (7B × 2 bytes), so the package in circulation is most likely from the ASR family. But the most valuable thing to read in this repo is not the deployment table — it's the News timeline: a family history of "release, code withdrawal, pivot," and the TTS line has never been restored.
Family history: 11 days from release to withdrawal
On 2025-08-25, Microsoft open-sourced VibeVoice-TTS: 90-minute long-form audio and 4-speaker multi-voice synthesis, later an ICLR 2026 Oral. Eleven days later, on 2025-09-05, Microsoft removed the TTS code from the repository, stating it was removed because the tool was being used "in ways inconsistent with its stated intent" — the generated speech was too realistic, and misuse risk outweighed the value of an open demonstration. What was withdrawn was the generation side, not the whole project:
- December 2025: Realtime-0.5B returns (the smallest TTS-side model, streaming text input, 9 languages plus 11 English style voices)
- March 2026: ASR enters Azure AI Foundry Labs
- July 2026: ASR-BitNet ships an edge CPU engine
- September 3, 2026: ASR-Streaming releases — live transcription while speaking, labeled with "who said what when," hotword support, 10 languages
- ASR-7B (the package in circulation) runs directly: 14GB VRAM is the bf16 full model, and quantization can compress further
- ASR-Streaming and Realtime-0.5B are official and active; Colab suffices to try them
- TTS only continues via community forks — the official version stopped at the 2025-09-05 withdrawal statement. If a package is a TTS fork, its lineage is community-maintained and not backed by Microsoft's official code
The model table's Quick Try column for TTS-1.5B still reads Disabled.
Withdrawing the code 11 days after release is not a stain on the assertion-strength ladder; it is the publisher closing the verification-bandwidth gate themselves: deepfake risk was too real, so the generation side was pulled first. The pivot runs on the same foundation — the 7.5Hz tokenizer unchanged — only shifting from generating acoustic detail to understanding it. Doing understanding first when generation is risky is a structure, not a product-line blurb.
7.5Hz: the interface position on the time axis
VibeVoice's core is a continuous speech tokenizer (dual acoustic + semantic paths) compressed to 7.5Hz, plus next-token diffusion: an LLM autoregressively walks the semantic layer while a diffusion head unfolds acoustic details in parallel. The paper root is Microsoft's LatentLM (arXiv 2412.08635).
This is a third solution in the speech lineage. Where the Luna-TTS comparison contrasted AR flattening the RVQ grid (losing structure) with Block Diffusion (preserving structure) — comparing the *decoding order* interface — VibeVoice changes dimension: it compares frame rate, not order. 7.5Hz means each second of speech occupies only 7.5 tokens; 90 minutes of audio is roughly forty thousand tokens, fitting inside LLM context. At 75Hz with an RVQ grid it would be four hundred thousand — long audio is simply out. The ASR capability of "60 minutes in a single pass, 64K tokens, structured who/when/what output" is all bought by the 7.5Hz interface position. Every step down in frame rate recomputes the context economics of long audio — frame rate is the quantization granularity of the time axis, structurally isomorphic to bit-width on the spatial axis.
The BitNet line: cutting the GPU entirely
ASR-BitNet (2026-07) uses heterogeneous I8_S+I2_S quantization to shrink the model from 4.62GB to 1.58GB, running real-time (RTF < 1) on 3 CPU threads with no GPU. Compare with the edge-side lineage (Thor-U, TensorFold, Strata) still fighting over how to save 8GB of VRAM: on the speech side, the GPU has been cut out entirely. Deployment collapse reaches its extreme in the speech modality — harsher than text, because speech tokens are already 10× fewer.
Practical deployment notes
Two falsifiable predictions
1. Within 12 months, TTS returns to the official repository in a controlled form (watermarking or audit logging required up front) — or it stays permanently in forks as the textbook case of "release-withdrawal." 2. ASR-Streaming's real-time who-said-what structured output becomes the standard interface for agent speech frontends — the ground truth for meeting scenarios, with the same positioning logic as an MCP for the physical world.
It ends where the package begins: before downloading, ask which line it carries — ASR is Microsoft's present, the TTS fork is Microsoft's past, and both share the same August 2025 starting point.
---
*Verification chain: GitHub API (stars / push times / no release records), full README (all six News timeline entries, the model table's Disabled marking, the 7.5Hz wording), arXiv 2412.08635 (Multimodal Latent Language Modeling with Next-Token Diffusion), arXiv 2609.02812 (VibeVoice-ASR-Streaming Technical Report). ICLR 2026 Oral is per the README's own claim; the 14GB VRAM figure is a bf16 arithmetic inference (7 × 2 = 14GB). The circulated package was not tested; resource lineage follows the README model table.*