VibeVoice is an open-source voice AI framework from Microsoft covering both text-to-speech and automatic speech recognition. VibeVoice-TTS generates synthetic speech up to 90 minutes long with up to 4 distinct, consistent speakers across a multi-speaker dialogue; VibeVoice-ASR transcribes up to 60 minutes of continuous audio in one pass, identifying who said what and when across 50+ languages with support for custom hotwords; VibeVoice-Streaming is a lightweight 0.5B-parameter model for real-time streaming TTS at roughly 300ms latency. It's built on continuous speech tokenizers running at 7.5 Hz combined with LLMs and diffusion-based generation. Released under the MIT license, with model weights on Hugging Face and interactive playgrounds for testing ASR and streaming.
VibeVoice is an open source voice framework from Microsoft covering both directions of speech. On the synthesis side it is built for long form, multi speaker audio, so it is aimed at content that runs for an extended stretch and involves more than one voice holding a consistent identity across a dialogue, which is where many text to speech systems drift. On the recognition side it transcribes long continuous audio in a single pass and attributes who spoke when, across a broad set of languages, with support for supplying custom terms so names and jargon transcribe correctly. A separate lightweight model targets real time streaming synthesis, for cases where audio has to start playing while text is still arriving. The project is MIT licensed with weights published on Hugging Face and playgrounds for trying the recognition and streaming pieces before committing to a deployment.
The long form multi speaker angle points at obvious uses: generated podcast style dialogue, audiobook narration with distinct characters, training material with alternating voices, and any format where a single flat narrator would not carry. On the recognition side, transcribing a long meeting or interview with speaker attribution in one pass is a real workflow saver compared to stitching chunks together. It also fits teams that need voice work to stay on their own infrastructure for confidentiality reasons, since MIT licensed weights mean no per minute vendor bill and no audio leaving the building. It is not a plug and play consumer app. Expect to install a Python stack, obtain weights, and provide GPU capacity, and expect voice cloning of real people to carry consent obligations you should take seriously.
Kokoro is the lightweight end of this category, a small text to speech model that runs comfortably on modest hardware and does its job without the multi speaker and long form ambitions. Orpheus TTS aims at natural, expressive delivery and is a sensible comparison when the priority is how one voice sounds rather than managing several. Zonos is an open weight model from Zyphra centered on voice cloning and multilingual expressiveness. VibeVoice differs from all three by spanning both synthesis and transcription in one framework and by treating long multi speaker audio as the design target rather than an edge case. For a single short clip the smaller models are quicker to stand up.
Compare with: Kokoro, Orpheus TTS, Zonos
The catalog records VibeVoice as free, and the release is MIT licensed, which is about as permissive as open weights get, including for commercial use. There is no per character or per minute charge of the kind hosted speech APIs bill. Your cost is hardware and operations: GPU capacity to generate or transcribe, storage for weights, and the engineering time to run the stack. That structure favors steady volume, where a fixed machine beats metered billing, and penalizes occasional use, where a hosted API is simpler and probably cheaper.
No notes have been added for this tool yet.
No prompting tips have been added yet.
No reviews yet. Be the first!