VibeVoice

well duh, they updated the news section https://github.com/microsoft/VibeVoice/commit/e73d1e17c3754f... which is microsoft for "we removed two dead links". AI innovation knows no limits!

I think we should stop calling this type of models open source. They are indeed "open weight." The training code is proprietary and never revealed. https://github.com/microsoft/VibeVoice/issues/102

According to this issue[0] the 1.5B model needs 6GB of VRAM. Meanwhile it looks like NeuTTS is designed to be able to run on CPU, which is nice for older/lower-spec hardware. 0: https://github.com/microsoft/VibeVoice/issues/26#issuecommen...

Checks notes, RIP VibeVoice Large 7B https://arxiv.org/pdf/2508.19205 https://github.com/microsoft/VibeVoice Nice to have forks & downloadable models now 'innit

Did they delete the repo? It's 404 for me now: https://github.com/microsoft/VibeVoice

Looks like the repo went private https://github.com/microsoft/VibeVoice I was trying to get this working on strix halo.

They mention that in the FAQ here: https://github.com/microsoft/VibeVoice/tree/main?tab=readme-... > In fact, we intentionally decided not to denoise our training data because we think it's an interesting feature for BGM to show up at just the right moment. You can think of it as a little easter egg we left for you. It's not a bug, it's a feature! Okaaaaay

This is a technically impressive release, and its architecture reveals a lot about why it achieves such high quality in dialogue generation. It's not just another TTS model; it's a well-thought-out system. Diving into its design, VibeVoice employs a sophisticated two-stage, cascaded generation process, which is key to its performance: Semantic Generation (The "What"): It first uses a powerful LLM backbone (a Qwen2 variant) to convert the input text into semantic tokens. This is where the model's deep contextual understanding comes from. It's not just reading words; it's interpreting the structure of a conversation, which is why it handles complex, multi-speaker scripts (Speaker 0: ..., Speaker 1: ...) so effectively and maintains long-form coherence. Acoustic Generation (The "How"): The semantic tokens are then passed to a diffusion-based