OpenMOSS/MOSS-TTS-Nano

What it solves

MOSS-TTS-Nano is designed to make high-quality, multilingual speech generation accessible for real-time products. It solves the problem of high deployment costs and latency typically associated with TTS models by providing a tiny footprint (0.1B parameters) that can run efficiently on CPUs without requiring a GPU.

How it works

The project uses a pure autoregressive pipeline consisting of an Audio Tokenizer and a Large Language Model (LLM).

• MOSS-Audio-Tokenizer-Nano: A lightweight, CNN-free tokenizer (20M parameters) based on the Cat architecture (Causal Audio Tokenizer with Transformer). It compresses 48 kHz stereo audio into a 12.5 Hz token stream using RVQ with 16 codebooks.
• TTS Pipeline: The LLM generates audio tokens which are then decoded back into high-fidelity 48 kHz, 2-channel audio.
• Deployment Options: It supports standard PyTorch inference, a highly optimized ONNX CPU version (which removes PyTorch dependencies and improves efficiency by nearly 2x), and even Android ONNX Runtime integration.

Who it’s for

• Developers building real-time voice applications who need low latency.
• Users wanting to run high-quality TTS locally on consumer hardware (e.g., MacBook Air M4) using only a CPU.
• Product teams integrating speech generation into lightweight web services or mobile apps.

Highlights

• Ultra-lightweight: Only 0.1B parameters for the main model.
• CPU-Native: Capable of streaming generation on a 4-core CPU.
• Multilingual: Supports 20 languages, including Chinese, English, German, Spanish, French, Japanese, and more.
• Voice Cloning: Supports zero-shot voice cloning using a reference audio prompt.
• High Fidelity: Native 48 kHz stereo output.
• Flexible Deployment: Available via CLI, FastAPI web demo, and ONNX for standalone CPU/Android use.