MisoLabs Debuts MisoTTS, an Open Voice Model

A new contender has entered the open-source speech synthesis space. Startup MisoLabs has released MisoTTS, a text-to-speech (TTS) model that applies popular architectural patterns from large language models to the challenge of generating human-like audio.

Unlike many traditional TTS systems, MisoTTS uses a decoder-only transformer architecture, a design heavily inspired by models in the Llama family. This approach treats audio generation as a sequence-to-sequence task, similar to how an LLM predicts the next word in a sentence. The goal is to produce more natural and expressive speech by leveraging the same principles that have dramatically advanced text generation.

MisoTTS at a Glance

The model was trained on a foundation of public domain audiobooks and currently supports two languages. Key features of the initial release include:

• Architecture: 24-layer decoder-only transformer.
• Languages: English and Japanese.
• License: Creative Commons BY-NC-SA 4.0 (non-commercial use).

The release of MisoTTS highlights a growing trend of cross-pollination in AI research, where successful architectures from one domain are adapted to solve problems in another. While its non-commercial license limits its use in products, it provides researchers and hobbyists a new tool for exploring the intersection of language and speech. The model and code are available now on the Hugging Face Hub.