KittenML has released code and weights for Kitten-TTS 2, a speech synthesis model that the company describes as the world's first to use 1.58-bit quantization. Ternary weights, in which each linear weight takes one of three values, sharply cut the memory required and let the model run locally or on a self-hosted server at minimal cost.
The model runs in real time on CPU alone. It offers zero-shot voice cloning, meaning no extra training: a sample of about five seconds of someone's voice is enough to make it speak the chosen text. It supports several languages and lets users control expressiveness and emphasis.
README di KittenML/KittenTTS
According to the project documentation, the flagship Kitten-TTS 2 has 1.7 billion parameters and is a speech language model with expression control. The release follows a run of recent launches in the same space: Kyutai's Pocket TTS brought voice cloning to models of just 1.9 million parameters, while Gradium's Phonon also aims to run locally on CPU, GPU and NPU without sending voice to the cloud.
The project is available on GitHub and Hugging Face, with a community on Discord. KittenML's developer, Stellon Labs, also offers a larger version of the model and paid APIs for speech synthesis and transcription.
No comments yet. Be the first!