KittenTTS - The Nano TTS

Quick Overview

Kitten TTS released an upgrade featuring three new tiny, open-source text-to-speech models (80M, 40M, and 14M parameters), with the smallest quantized version being just 25MB and capable of running on minimal hardware, including directly in a browser or on edge devices without a GPU.

Key Points: Kitten TTS released an upgrade with three new models: 80 million (Mini), 40 million (Micro), and 14 million (Nano) parameters. The smallest model, kitten-tts-nano-0.8-int8 (15M parameters), is only 25MB and optimized to run on CPU/minimal hardware without requiring a GPU. The models are built on the StyleTTS 2 architecture and offer high-quality voice synthesis with fast inference, optimized for real-time deployment. The update includes 8 distinct voices (Bella, Jasper, Luna, Bruno, Rosie, Hugo, Kiki, Leo) available across the models, particularly showcased using the 80M Mini model. Benchmark tests demonstrated fast inference times, with the Nano fp32 model achieving a Real-Time Factor (RTF) of 0.30x for conversational text. The project is open-source under the Apache-2.0 license and actively maintains a GitHub repository and Discord community for user interaction and support.

Context: The video announces an upgrade to Kitten TTS, an open-source, lightweight text-to-speech (TTS) system developed by KittenML, emphasizing its focus on creating compact, high-quality models suitable for deployment on resource-constrained devices. The announcement highlights the release of three new model sizes (80M, 40M, 14M parameters) and showcases performance benchmarks and voice quality across these different sizes and quantization levels.

Detailed Analysis

Kitten TTS has been upgraded, introducing three new tiny, open-source TTS models based on the StyleTTS 2 architecture: Mini (80M parameters, 79MB), Micro (40M parameters, 40MB), and Nano (15M parameters, 56MB). Crucially, an 8-bit quantized version of the Nano model (kitten-tts-nano-0.8-int8) achieves a file size of just 25MB, allowing it to run efficiently on CPU or minimal hardware, even within a browser or on mobile devices. The models support 8 pre-made voices (Bella, Jasper, Luna, Bruno, Rosie, Hugo, Kiki, Leo), demonstrated via conversational and technical sentences, showing strong voice quality maintained even in the smallest versions. Benchmarks confirm the models' speed, with the Nano fp32 model achieving an RTF of 0.30x on conversational text, significantly faster than previous architectures. The project is committed to packing frontier intelligence into these tiny models and encourages community engagement via GitHub and Discord.