Kitten-TTS 2: New Open-source SOTA 1.58 bit TTS (CPU-only with voice cloning, multilingual, emotion control)

KittenML has just released code and weights of Kitten-TTS 2, which is a super-efficient frontier TTS model, and the world's first 1.58-bit TTS model. It runs in real time on just a CPU, with zero-shot voice cloning, multilingual support, and expressive / emphasis control. Code and weights are open-source. Ternary(1.58-bit) Weights means it can run locally or be self-hosted for super-cheap. Github: https://github.com/KittenML/KittenTTS Hugging Face: https://huggingface.co/KittenML/kitten-tts-2 KittenML's Discord: https://discord.com/invite/VJ86W4SURW Key features and Advantages Ternary TTS: Every linear weight in the body is one of three values (-1, 0, +1) with group size of 128 for scales. Tiny footprint and fast CPU inference. The model has 1.7B parameters. Real-time on CPU and consumer GPU's too: No GPU needed for real-time generation. Optimized ternary kernels for CPU inference provided. Expressive voice cloning: zero-shot cloning from less than 5 seconds of reference audio, and it keeps the expressivity, accent information etc. No fine-tuning needed. Multilingual TTS: Model supports voice cloning in 20+ languages. Cross-language voice cloning can be used to create accented speech too. Expressive / emphasis control: you can steer the delivery and stress specific words, not just pick a voice. Details in the Readme. Open source (hell yeah!): code and weights are open-source and free to use. It's an oss alternative to elevenlabs for gpu-poor folks. easy to self-deploy or run locally. 7. What changed from V0.8 to 2: higher quality. Completely new architecture, natural sounding speech, ternary weights, zero-shot voice cloning, multilingual, and emphasis/emotion control.