Kokoro

Kokoro is an ultra-lightweight TTS model with only 82M parameters, supporting zero-shot voice cloning and multilingual synthesis across 9 languages including Chinese, English, French and Japanese. It runs completely offline and locally, powering use cases like local voice assistants, audiobook generation (e.g., converting EPUB to M4B in minutes via Audiblez), and personalized voice cloning. Basic synthesis starts with just three lines of Python.

Key strengths: Extremely small—real-time on CPU, as low as 40ms latency with WebGPU; 50 preset high-naturalness voices with natural listening feel; fully local deployment means zero server cost and no privacy exposure, reducing long-term expenses.

Limitations and barriers: Requires basic Python environment and command-line skills, lacking a no-code GUI, which may deter non-technical users; may underperform larger models like CosyVoice in extreme emotional expression or complex contexts; capabilities are focused on preset voices and zero-shot cloning, with less convenient custom voice training.

Deployment and cost: Completely open-source and free; model download is ~330MB, runs locally with no per-use API fees.