Nari Labs releases Dia2-2B, a streaming English dialogue TTS under Apache 2.0
Dia2-2B is a 2B English dialogue TTS that can start speaking before the full text arrives; Apache 2.0, about two minutes max.
Nari Labs published Dia2-2B, a 2B-parameter English text-to-speech model for dialogue. The card says generation can start once the first few words of text are provided, which suits streaming into speech-to-speech systems. Multi-speaker scripts use [S1] and [S2] tags, audio can be conditioned on speaker prefix clips, and word timestamps are available relative to the Mimi codec's roughly 12.5 Hz frame rate. It runs through a CLI, a Python API and a Gradio app, and needs uv and CUDA 12.8 or newer drivers.
The card gives no benchmark numbers. It says the model supports only up to 2 minutes of generation in English, about 1,500 context steps, and that real streaming support in its TTS server is listed as upcoming.
The licence is Apache 2.0, with third-party components such as the Kyutai Mimi codec under their own licences. The card says the model is for research and education, and prohibits identity misuse, deceptive content and illegal use. The card gives no release date; the date used is the Hugging Face repo creation date.
Streaming is the headline. The streaming server is still marked upcoming.
Sources
Editor's first pass (AI-assisted)
- Real 4: Apache 2.0 weights with CLI, Python API and Gradio app.
- Useful 3: Two-speaker dialogue is handy, but output caps at about two minutes and English only.
- Notable 3: A successor to a popular small-team model.
- Novel 3.5: Streaming generation from the first few words is the distinguishing feature.
Comments