KittenTTS

Project Overview

KittenTTS is an open-source text-to-speech library. The flagship model, KittenTTS 2, is a 1.7B-parameter 1-bit speech language model that clones any speaker from about 5 seconds of reference audio and supports emotion tags and emphasis control. It also keeps lightweight ONNX models from 15M to 80M parameters that run on a CPU with no GPU, synthesizing 24 kHz speech in real time.

Voice cloning and lightweight deployment in one library that speaks in real time on a CPU.

Key Features

• Voice cloning: replicate any speaker from 5-30 seconds of audio, no fine-tuning
• Built-in voices: 47 presets, including 9 non-English voices
• Multilingual: 20 languages including Chinese, English, Japanese, Korean, French, and German
• Expression control: emotion tags, event tags, and emphasis brackets
• Decoding presets: trade stability against expressiveness per request
• Long-form text: sentence-aware chunking for article-level reading
• Text preprocessing: numbers, currencies, dates, and abbreviations expanded automatically
• Lightweight models: 15M-80M ONNX models run without a GPU

Target Audience

• Developers adding speech narration to apps
• Content creators cloning specific voices
• Engineers synthesizing speech on low-end or edge devices
• Teams producing audiobooks and voiceovers

Quick Start

1. Install:

pip install kittenml

2. Generate speech:

from kittenml import KittenTTS

m = KittenTTS('KittenML/kitten-tts-2')
audio = m.generate('Hello, world', voice='Bruno')

3. To clone a voice, pass reference instead of voice with your recording:

audio = m.generate('This is a cloned voice', reference='my_voice.wav')

4. Lightweight models load through the same KittenTTS constructor with no extra install

Core Advantages

• No GPU needed: lightweight models run on CPU, and the flagship also has a CPU path
• Low cloning barrier: a few seconds of audio replicates a voice without training
• Minimal install: one package, weights download on demand, no Hugging Face login

System Requirements

Notes

• Cloning input: reference audio should be 5-30 seconds of a single speaker; mixed voices reduce similarity
• Weight choice: default weights are 947 MiB lossless; weights=emb4 halves the size at a small quality cost
• Legacy voices: the eight voices from 0.8 remain in the new model, so older code keeps working