MOSS-TTS: A New Multilingual Text-to-Speech Model

The OpenMOSS Team has released MOSS-TTS, a new open-source model for generating high-quality speech from text. The system is multilingual, capable of producing audio in Chinese, English, and Japanese, making it a versatile tool for a range of voice applications.

The model's key innovation lies in its architecture. MOSS-TTS is a non-autoregressive system that uses a technique called a "delay-pattern." This approach allows it to model the rhythm and prosody of speech more effectively than some traditional methods, which can result in more natural-sounding intonation without generating audio one step at a time.

A Two-Stage System

Like many modern text-to-speech systems, MOSS-TTS operates in two stages:

• First, a text-to-spectrogram model converts the input text into a mel-spectrogram, a visual representation of the sound's frequency spectrum.
• Second, a HiFi-GAN vocoder takes this spectrogram and synthesizes it into a final audio waveform.

The complete model, along with instructions for use, is available on the Hugging Face Hub. While the weights are openly accessible, they are released under a custom license that prohibits commercial use, a key consideration for developers looking to integrate the technology.