A VibeVoice alternative for podcasts and dialogue
VibeVoice from Microsoft made a splash because its TTS model could voice up to 90 minutes of speech with up to four speakers, a whole podcast episode in one go. On 5 September 2025 the maintainers removed the VibeVoice-TTS code from the repository after finding misuse, and the model table marks the TTS-1.5B weights as disabled. What remains is VibeVoice-Realtime-0.5B, a streaming model for about 10 minutes of speech, and the project says the models are intended for research and development only.
If you came for multi-speaker podcasts or dialogue, the free SuperClonica engine gets you there another way: each speaker is a voice, each line is voiced separately, and the lines are joined into one file. It runs offline on an ordinary processor, and its stock voices may be used commercially.
What happened to VibeVoice
The VibeVoice repository is MIT-licensed and still hosts the speech recognition models and the Realtime-0.5B TTS model, with streaming text input and first audio in about 300 ms. The long-form TTS code was removed on 2025-09-05, and the TTS-1.5B weights are listed as disabled. Check the repository for the current state before you build on it.
Hear the voices first
On the page of any voice in the catalog, paste a line of your script, 30 to 100 characters, choose the language and press Voice it. In a few seconds you know which voices to cast as host and guest, with no account.
A two-host podcast from one script
Install the engine with python -m pip install superclonica-cpu. Write the script with one speaker per line, such as A: Welcome back to the show. and B: Thanks, glad to be here. Give each speaker a voice, for example F2 for the host and M3 for the guest, voice every line and join them with a short pause; our AI voice dialogue guide has a ready Python snippet that turns such a script into one WAV file. A long episode is just more lines: there is no length cap, and changing one line means re-voicing one line.
Where VibeVoice was stronger
A long-form model reads the whole conversation at once, so turn-taking, laughter and the flow between speakers can sound more spontaneous, and VibeVoice also spoke Chinese. Line-by-line voicing gives you precise control over every line and runs on any computer, but overlapping speech and interruptions have to be written as separate short lines.
Voices you can publish
The ten stock voices, M1 to M5 and F1 to F5, may be used commercially under the OpenRAIL-M model license, and so may a voice trained from your own recording or one whose speaker agreed. Character voices in the catalog are fan-made, for fan and parody use only. Never script a fake conversation between real people that could pass for a real recording.