# Microsoft’s VibeVoice release put long-form multi-speaker TTS into an open research framework, then exposed its misuse boundary Microsoft’s VibeVoice project records an August 25, 2025 open-source TTS release for long-form multi-speaker speech, followed by code removal after misuse reports in September. Canonical: https://rangoon.ai/insights/archive-microsoft-vibevoice-open-source-tts-august-2025/ Author: Rangoon Editorial (https://rangoon.ai/insights/editorial/) Published: 2026-10-24 Updated: 2026-10-24 Event date: 2025-08-25 Tags: [Microsoft](https://rangoon.ai/insights/tags/microsoft/), [text to speech](https://rangoon.ai/insights/tags/text-to-speech/), [AI audio](https://rangoon.ai/insights/tags/ai-audio/), [open source](https://rangoon.ai/insights/tags/open-source/) Source note: Historical reporting: Microsoft’s VibeVoice repository records an August 25, 2025 open-source release of VibeVoice-TTS, described as supporting up to 90 minutes of speech and up to four speakers. The repository later says that Microsoft removed VibeVoice-TTS code on September 5 after discovering use inconsistent with the stated research intent. Microsoft Research’s later paper page describes the continuous 7.5 Hz speech tokenizers and next-token diffusion approach; it is later technical context, not a substitute for an August release record. Neither source establishes that users could safely clone a real person, that all source code or weights remained available, or that model outputs were accurate recordings. Image: Synthetic-dialogue provenance. Original Rangoon explanatory diagram: Source boundary: dated primary release reporting; original editorial analysis. No vendor interface, photograph, measurement result, or endorsement implied. ## Key takeaways - The project record treats August 25 as an open-source release of long-form, multi-speaker TTS research artifacts. - The September removal notice is part of the same product history and prevents a simple claim that the release stayed fully available. - Long-form synthetic dialogue needs disclosure, consent, provenance and review because naturalness can increase impersonation and misinformation risk. ## The official project history dates the open-source TTS release to August 25 [source 1](https://github.com/microsoft/VibeVoice) Microsoft’s VibeVoice repository records an August 25, 2025 open-source release of VibeVoice-TTS. The project describes it as a long-form, multi-speaker text-to-speech model able to synthesize up to 90 minutes of speech with as many as four speakers. This establishes a specific research-release event and a vendor-stated capability envelope. It does not prove that every generated long-form result retains consistent speakers, that a user can generate any requested voice, or that a deployment has a right to imitate a person. A length limit and speaker count are model descriptions, not quality guarantees. A real application still needs to establish who supplied a script, what voice inputs were permitted and how a listener can tell synthesized audio from a direct recording. ## Later technical material explains the research architecture without rewriting release-day access [source 1](https://github.com/microsoft/VibeVoice) [source 2](https://www.microsoft.com/en-us/research/publication/vibevoice-expressive-podcast-generation-with-next-token-diffusion/) Microsoft Research’s VibeVoice paper page, listed for ICLR 2026, describes continuous acoustic and semantic speech tokenizers running at 7.5 Hz and a next-token diffusion design. It also describes long-form multi-speaker conversational synthesis as a research goal. This helps explain why the project emphasized long sequences and turn-taking: lower-rate representations can make lengthy audio more tractable than a high-frame-rate stream. But the later paper page cannot establish what exact code, weights, interface or license a user encountered on August 25. Research architecture and release availability are different facts. The archive preserves that distinction rather than using a later publication page to inflate the original launch. ## The later removal notice belongs in the event record, not outside it [source 1](https://github.com/microsoft/VibeVoice) The same official repository says Microsoft removed VibeVoice-TTS code on September 5, 2025 after discovering uses inconsistent with its stated research intent. That later action changes how the August event should be understood: an open-source release is not proof that every artifact remained publicly available, supported or appropriate for every use. The notice also makes a practical provenance problem visible. A link, mirror or old installation may persist after an upstream change, but it does not automatically carry the upstream project’s present status or safety judgment. Any account of the release needs both dates: the August publication and the September removal statement. ## Natural synthetic dialogue requires product-level safeguards beyond research access [source 1](https://github.com/microsoft/VibeVoice) [source 2](https://www.microsoft.com/en-us/research/publication/vibevoice-expressive-podcast-generation-with-next-token-diffusion/) VibeVoice’s project documentation warns that synthetic speech can be used for deepfakes, fraud and disinformation, and says it is intended for research and development rather than unqualified commercial or real-world use. That is a direct vendor warning, not a comprehensive policy for every jurisdiction or platform. A responsible product needs consent for voice material, provenance for scripts and outputs, disclosure that audio is synthetic, limits on identity imitation, escalation for suspected impersonation and a way to correct or remove harmful material. Evaluation should include audibility of disclosure and how listeners understand speaker attribution, not only perceived naturalness. The 2025 history supports a substantial research release and a later misuse response; it does not turn a model’s output into authentic speech or relieve an operator of these controls. ## Sources 1. [Microsoft VibeVoice repository: release history and safety notice](https://github.com/microsoft/VibeVoice) 2. [Microsoft Research: VibeVoice, expressive podcast generation with next-token diffusion](https://www.microsoft.com/en-us/research/publication/vibevoice-expressive-podcast-generation-with-next-token-diffusion/) ## Related reading - [Kyutai released Hibiki’s French-to-English voice-translation research artifacts with a narrow language boundary](https://rangoon.ai/insights/archive-kyutai-hibiki-simultaneous-translation-february-2025/) - [Kyutai presented Unmute as a modular speech wrapper for text models, while deferring its open-source release](https://rangoon.ai/insights/archive-kyutai-unmute-modular-voice-ai-may-2025/) - [ElevenLabs launched Eleven v3 as an alpha for expressive speech, with real-time use still out of scope](https://rangoon.ai/insights/archive-elevenlabs-eleven-v3-alpha-june-2025/) ## More information - [Documentation index](https://rangoon.ai/llms.txt) - [AI access and policies](https://rangoon.ai/ai/) - [Source repository](https://github.com/hypler-dev/rangoon)