Dia (TTS)
lastmod 2025-04-22created 2025-04-222 edits
An open-weight TTS model that generates realistic dialogue
- Dia is a 1.6B-parameter TTS model that generates high-quality dialogue speech from a text script, with emotion and tone controllable through an audio prompt
- Developed by Nari Labs; "Nari" means "lily" in native Korean
- Speakers are specified with [S1], [S2], nonverbal expressions such as (laughs) and (coughs) can also be generated, and simple voice cloning is supported
- Runs directly on HuggingFace, with browser-based testing without installation and ZeroGPU support
- Currently supports English only and requires 10GB+ VRAM; quantized models and multilingual support are planned
Dia: a dialogue-focused speech synthesis model
- Dia is a 1.6B-parameter open-weight TTS model developed by Nari Labs
- Instead of generating each speaker's voice separately like conventional TTS, it generates the whole dialogue at once
- Demo: Hugging Face Space
- Code: GitHub repository
Key features
Dialogue speech generation
- Speakers can be specified with [S1], [S2] in the text
- Nonverbal sounds such as (laughs), (coughs) can be inserted as text
- Emotion, tone, and voice style can be specified with an audio prompt
Voice cloning
- Providing sample audio together with its transcript as text enables voice cloning
- Can be tried on the Hugging Face Space by uploading audio
- See example/voice_clone.py for a detailed example
Using as a library
from dia.model import Dia
model = Dia.from_pretrained("nari-labs/Dia-1.6B")
output = model.generate(text)
See also
- Dia - An open-source diagramming program.
- Speech synthesis
- TTS