Kitten TTS 2 Introduces Ultra-Lightweight Multi-Lingual Speech Models
Kitten ML released Kitten TTS 2, featuring a family of ultra-lightweight text-to-speech models including Mini (82M parameters), Nano (17M parameters), and Pico (7.2M parameters). These models support multiple languages, dozens of voices, and provide ONNX int8 quantization for efficient CPU execution. These models enable efficient text-to-speech inference directly on edge devices and CPUs without requiring heavy PyTorch dependencies. This significantly lowers the barrier for deploying local, private voice generation in resource-constrained environments. The models operate at a 24 kHz sample rate and come in various sizes, with the int8 ONNX versions weighing as little as 8.0 MB for Pico, 17.4 MB for Nano, and 82.7 MB for Mini. They also offer float32 PyTorch weights alongside the ONNX runtime files.
## BACKGROUND
Text-to-speech (TTS) systems convert written text into spoken audio using deep learning models. Quantization techniques like ONNX int8 reduce model size and memory bandwidth requirements by representing weights with 8-bit integers instead of 32-bit floats, making them faster to run on standard CPUs.