MOSS-TTS-Nano-100M

Overview

MOSS-TTS-Nano is a multilingual speech generation model designed for real-time applications. With only 0.1 billion parameters, it delivers a compact footprint while maintaining low latency and good audio quality. The model supports 48 kHz stereo output and runs efficiently on CPU, making it suitable for lightweight deployment scenarios.

Key Strengths

• Small footprint: 0.1B parameters enables fast inference and simple hosting.
• Low latency streaming: Supports streaming generation with fast first-audio delivery.
• CPU-friendly: Can generate speech on a 4-core CPU without a GPU.
• Long-text capable: Handles extended input with automatic chunked voice cloning.
• Multilingual: Covers 20 languages including Chinese, English, German, Spanish, French, Japanese, Korean, and more.

Supported Languages

Architecture

The model uses a pure autoregressive pipeline combining an Audio Tokenizer and a small LLM. The underlying MOSS-Audio-Tokenizer-Nano (~20M parameters) compresses 48 kHz stereo audio into a 12.5 Hz token stream using RVQ with 16 codebooks, enabling variable bitrates from 0.125 to 4 kbps.

Best For

Voice cloning, real-time speech synthesis for interactive applications, and multilingual TTS where low latency and minimal compute requirements are critical. The model is hosted as an OpenAI-compatible API on Gigarouter, allowing drop-in integration without local infrastructure.