Dia TTS

Nari Labs trained this 1.6B-parameter speech model to turn an entire two-person script into audio in a single pass rather than stitching separate sentences together. Turns are marked with [S1] and [S2] tags, cues such as (laughs), (sighs) or (coughs) produce nonverbal sounds, and supplying a 5 to 10 second reference clip with its transcript steers the voice, which is also the way to keep a speaker consistent between runs. The package installs with pip or uv and ships a Python API, a CLI and a Gradio app, and the checkpoint also loads natively in Hugging Face Transformers through DiaForConditionalGeneration. It has only been tested on CUDA GPUs: on an RTX 4090 it generates faster than real time in bfloat16 using roughly 4.4 GB of VRAM. Output is English only. Code and weights are Apache 2.0, so commercial use is allowed, although the authors forbid impersonation and deceptive content. Nari Labs has since released Dia2, a streaming successor in 1B and 2B sizes, so new projects should compare the two. A good fit for podcast-style demos, game dialogue and voice agent prototypes.