Open-source TTS model Scylla's Band v4 released

On November 7, 2026, at 06:26, a developer posted on the Reddit community LocalLLaMA announcing the release of Scylla's Band v4, an open-source text-to-speech model. The developer described it as a long-running effort, calling it almost a year of solo work spent learning how to build a TTS model from the ground up. The stated goal was a fast model small enough to run on a phone or other mobile devices, multilingual, and capable of doing something a little different from other small TTS models.

That differentiated feature, the developer says, is singing. Alongside speech synthesis, Scylla's Band v4 can generate singing, which the developer describes as the part they cared about most. According to the post, the speech and singing heard in the demo video are entirely produced by the model, while the instrumental parts and sound effects were not generated by it and were added afterwards.

The model has roughly 131M parameters, plus a 30M phonemizer, and supports 14 voices; the accompanying summary also notes multilingual capability. Beyond the announcement itself, only one source is available, and no independent verification or outside coverage appears in the existing material.