KittenTTS v0.8 has launched with three new text-to-speech models ranging from 15M to 80M parameters, all under 80MB in size. The open-source library now offers kitten-tts-nano, kitten-tts-micro, and kitten-tts-mini variants, with the 15M parameter model achieving state-of-the-art expressivity for its size. These models run entirely on CPU without requiring a GPU, making them ideal for on-device applications.

Built on ONNX, KittenTTS delivers 24kHz audio output through eight distinct voices including Bella, Jasper, Luna, and Bruno. The models support adjustable speech speed, text preprocessing for numbers and currencies, and work across platforms from Raspberry Pi to low-end smartphones. Previous versions struggled with expressivity in tiny models, but the new 15M parameter variant breaks that barrier while maintaining a 25MB footprint when quantized.

The release addresses a critical bottleneck in on-device AI: the lack of tiny models that actually perform well. KittenML aims to bridge the gap between cloud and on-device TTS applications, with multilingual support coming soon. Commercial support is available for teams needing integration assistance or custom voices, as the project remains in developer preview with potential API changes.