KittenML/KittenTTS

What it solves

Kitten TTS provides a lightweight, open-source text-to-speech (TTS) library that allows high-quality voice synthesis to run efficiently on CPUs without requiring a GPU. It addresses the need for fast, low-footprint audio generation for edge deployment and local environments.

How it works

Built on the ONNX runtime, the library provides several model variants (ranging from 15M to 80M parameters) to balance performance and size. It includes a built-in text preprocessing pipeline to handle numbers, currencies, and units, and outputs audio at a 24 kHz sample rate. Users can choose from eight built-in voices and adjust the speech speed via a simple API.

Who it’s for

Developers looking to integrate voice synthesis into applications that must run on limited hardware, edge devices, or systems where GPU acceleration is not available.

Highlights

• Ultra-lightweight: Models range from 25 MB to 80 MB on disk.
• CPU-optimized: Uses ONNX for efficient inference without a GPU.
• Built-in Voices: Includes 8 distinct voices (e.g., Bella, Jasper, Luna).
• Text Normalization: Integrated pipeline for expanding abbreviations, dates, and currencies.
• Flexible Deployment: Supports both CPU and CUDA backends.