Kitten-TTS 2: Open-source 1.58-bit TTS Model with CPU Inference
KittenML has released Kitten-TTS 2, an open-source 1.7-billion parameter text-to-speech model that utilizes 1.58-bit ternary weights. It features real-time CPU inference, zero-shot voice cloning from short audio clips, multilingual support, and expressive emotion control. This model provides an ultra-efficient, open-source alternative to commercial cloud TTS providers like ElevenLabs, enabling cheap self-hosting and real-time generation on standard consumer hardware without requiring a dedicated GPU. Every linear weight in the model is restricted to three values (-1, 0, +1), significantly reducing memory footprint. It supports zero-shot voice cloning in over 20 languages using less than 5 seconds of reference audio.
## BACKGROUND
A 1.58-bit model, also known as a ternary model, restricts its neural network weights to three states (-1, 0, and +1), carrying approximately 1.58 bits of information per weight to achieve high computational efficiency. Zero-shot voice cloning is a technique that enables a neural network to mimic a specific voice instantly using a short audio sample without requiring fine-tuning or retraining.