MOSS-TTS-Nano (0.1B)
MOSS-TTS-Nano is an open-source multilingual tiny speech generation model from MOSI.AI and the OpenMOSS team. With only 0.1B parameters, it is designed for realtime speech generation, can run directly on CPU without a GPU, and delivers native 48 kHz stereo audio.
Introduction
MOSS-TTS-Nano focuses on the part of TTS deployment that matters most in practice: small footprint, low latency, good enough quality for realtime products, and simple local setup. It uses a pure autoregressive Audio Tokenizer + LLM pipeline.
Installation
pip install transformers==4.57.5 torch torchaudio soundfile
Quick Inference (Hugging Face AutoModel)
Load the model once, then choose your preferred generation mode:
import torch
from transformers import AutoModel
# Load model directly from Hugging Face
model_id = "maanka2/MOSS-TTS-Nano-100M"
model = AutoModel.from_pretrained(model_id, trust_remote_code=True)
model.eval()
1. Zero-Shot Voice Clone
Clone any target voice using a short 3–10 second reference audio clip:
output = model.inference(
text="Welcome to MOSS-TTS-Nano. Fast, real-time speech generation on CPU.",
output_audio_path="voice_clone_output.wav",
mode="voice_clone",
prompt_audio_path="assets/audio/zh_1.wav", # Reference speaker audio (3-10s)
)
print(f"Cloned audio generated at: {output['audio_path']}")
2. Real-Time Streaming Inference
Stream audio chunks iteratively with ultra-low latency while the model is still generating (ideal for chatbots and voice assistants):
for event in model.inference_stream(
text="Streaming generation enables instantaneous playback before synthesis finishes.",
output_audio_path="stream_output.wav",
mode="voice_clone",
prompt_audio_path="assets/audio/zh_1.wav" # Reference speaker audio (3-10s)
):
if event["type"] == "audio":
audio_chunk = event["waveform"] # torch.Tensor [channels, samples]
sample_rate = event["sample_rate"]
# Stream audio_chunk to speakers, WebRTC, or WebSocket
elif event["type"] == "result":
print(f"Streaming complete. Full audio saved to: {event['audio_path']}")
MOSS-Audio-Tokenizer-Nano
Model Specs
| Feature |
Details |
| Parameters |
0.1B (~100M) |
| Audio Output |
48 kHz, Stereo (2 Channels) |
| Quantization Scheme |
16-Codebook Residual Vector Quantization (RVQ) |
| Languages Supported |
20 languages (English, Chinese, Arabic, Spanish, French, etc.) |
| Hardware Target |
Real-time on CPU (4 Cores) & lightweight GPU deployments |
Citation
@misc{openmoss2026mossttsnano,
title={MOSS-TTS-Nano},
author={OpenMOSS Team},
year={2026},
howpublished={GitHub repository},
url={https://github.com/OpenMOSS/MOSS-TTS-Nano}
}
