KittenTTS

KittenTTS is a Python text-to-speech library for developers adding local speech generation to apps or producing spoken audio from text. It supports voice cloning without model training and doesn't require a GPU.

• Voice cloning: Generate speech in a speaker's voice from 5 to 30 seconds of reference audio; Whisper can supply the recording's transcript if you don't provide one.
• Built-in voices and languages: Choose supplied voices or generate speech in languages including English, Arabic, Chinese, Hindi, Spanish and Japanese.
• Long text and streaming: Turn longer scripts into audio with automatic sentence splitting, or start playback as completed sections become available, including with a cloned voice.
• Text normalization: Expand written numbers, currencies, dates, units and abbreviations into spoken forms automatically.

The library runs locally on Windows, macOS and Linux. The main model needs about 6 GB of RAM for CPU use, or roughly 8 GB of free memory on a CUDA GPU. Smaller models use ONNX and also run without a GPU. An optional C++ backend builds on a fork of llama.cpp and uses GGUF weights; vLLM acceleration requires Linux and an NVIDIA CUDA GPU. Model downloads don't require a Hugging Face login.

Streaming waits for a complete section before returning audio, so playback has an initial delay. Its joins can be rougher than those in a complete generated file. A separate hosted API and browser demo run on KittenML Platform.

The library code is open source under Apache 2.0. Model licenses are separate: the main model uses the Stellon Labs Community License.