A 100M-parameter voice cloner that runs on your laptop CPU

What it does MOSS-TTS-Nano is a 100-million-parameter text-to-speech model that generates speech from text and a short audio prompt for voice cloning. It outputs 48 kHz stereo audio across twenty languages, from Chinese and English to Arabic and Turkish. The project is explicitly aimed at deployment scenarios where model size and hardware requirements matter more than studio-grade fidelity.

The interesting bit The team ships a fully standalone ONNX version that drops the PyTorch dependency entirely during inference, which means it can run inside a browser extension or on a MacBook Air M4 using a single CPU core. That architecture choice—an autoregressive Audio Tokenizer feeding a small LLM—trades absolute quality for the ability to stream audio in real time on modest hardware.

Key highlights

• Only 0.1B parameters; designed for CPU inference without a GPU
• Supports 20 languages with automatic chunked processing for long inputs
• ONNX runtime variant claims nearly 2× the processing efficiency of the original PyTorch version in the team’s own tests
• Companion browser extension can run the ONNX model directly inside the browser without a separate local inference service
• Native 48 kHz, 2-channel output with streaming decode

Caveats

• Dependency setup can be finicky: the README warns that WeTextProcessing and pynini often fail to install cleanly and points to a GitHub issue for platform-specific wheel workarounds
• The project describes its own quality target as “good enough for realtime products,” which suggests it is not aiming to compete with larger cloud TTS models on fidelity

Verdict Worth a look if you need embeddable, offline voice cloning for a desktop app, browser extension, or low-traffic web service. Skip it if you need broadcast-quality narration or heavy server-side batch processing where GPU clusters are already available.