Dia
An open-source text-to-speech model that generates English dialogue locally, with voice cloning and NVIDIA GPU inference under Apache 2.0.
Dia generates spoken English conversations from a written script. It's for developers and researchers who want to produce two-speaker dialogue on their own hardware, with audio references to guide the voices and delivery.
- Two-speaker dialogue: Generate both sides of a conversation in a single pass using a script that identifies each speaker.
- Voice cloning and delivery: Use an audio prompt to reproduce a voice and influence the emotion and tone of the generated speech.
- Nonverbal sounds: Include laughter, coughing, sighs and other vocal sounds, though these can produce unexpected output.
Dia is open source under Apache 2.0, with downloadable pretrained weights on Hugging Face. It works through Hugging Face Transformers and includes a local Gradio interface and command-line support. A separate Hugging Face browser demo processes uploaded voice references on that hosted service.
Local inference has been tested on NVIDIA GPUs with PyTorch and CUDA. RTX 4090 benchmarks used about 4.4 GB of VRAM at reduced precision and 7.9 GB at full precision. English is the only supported generation language. Very short scripts can sound unnatural, while long scripts can produce speech that's unnaturally fast.
