Skip to main content
Back

OpenMOSS's MOSS-TTS-Nano Clones Voices in 20 Languages on a Laptop CPU

OpenMOSS releases a 100M-parameter multilingual voice cloning model that streams speech on a 4-core CPU across 20 languages under Apache 2.0.

  • OpenMOSS released MOSS-TTS-Nano-100M, a 0.1B-parameter open-source multilingual TTS model under Apache 2.0.
  • Runs streaming inference on a 4-core CPU with no GPU, targeting realtime on-device deployment.
  • Supports 20 languages including Chinese, English, German, Japanese, Arabic, and Korean.
  • Uses a CNN-free causal Transformer tokenizer compressing 48 kHz stereo to a 12.5 Hz token stream.
  • Primary workflow is voice cloning from a short reference clip, with automatic chunking for long text.
  • Ships with CLI, FastAPI demo, ONNX build, and GitHub repo.

Most open text-to-speech models still assume you have a GPU sitting around. OpenMOSS is pushing back on that assumption with MOSS-TTS-Nano-100M, a 0.1B-parameter speech model built to stream voice-cloned audio directly on a laptop CPU, in 20 languages, with a deployment story that fits in a single Python file.

MOSS-TTS-Nano is an open-source multilingual speech generation model from MOSI.AI and the OpenMOSS team. With only 0.1B parameters, it targets realtime synthesis on CPU and ships under Apache 2.0. The goal is deployability rather than maximum fidelity: something small enough to embed in an app, a demo, or an offline product without an inference server behind it.

What is actually in the box

The system is a two-part pipeline: a discrete audio tokenizer plus a small autoregressive language model that predicts audio tokens from text. Larger models in the MOSS-TTS family chase high-fidelity synthesis, while the Nano variant trades some quality for CPU-only inference and quick integration. The pure autoregressive architecture pairs the tokenizer with a language model, making it suitable for local demos, web services, and edge deployment.

The tokenizer is the interesting piece. MOSS-Audio-Tokenizer-Nano is a roughly 20-million-parameter component designed for high-fidelity audio compression, with 48 kHz stereo input and output to reduce compression loss. It squeezes 48 kHz stereo down to a 12.5 Hz token stream using RVQ with 16 codebooks, enabling reconstruction across variable bitrates from 0.125 kbps to 4 kbps. RVQ here means residual vector quantization: each frame is encoded by 16 stacked codebooks that progressively refine the reconstruction, producing a very short discrete sequence the LLM can autoregressively predict.