#4949th of 51 in VoiceFor voice, speech and music
dia
1.6B text-to-speech model that generates two-speaker dialogue from a transcript
- Stars
- 19.4k
- License
- Apache-2.0
- Last commit
- Nov 2025
- Language
- Python
Missed the latest GitHub check; these facts may be older.
Overview
Dia is a 1.6B parameter English text-to-speech model from Nari Labs that turns a script tagged with [S1] and [S2] into multi-speaker dialogue. It supports nonverbal tags such as (laughs) and (coughs), and voice cloning from a 5-10 second audio prompt with transcript. The repo ships inference code, a Gradio UI and a CLI, and the model also loads through Hugging Face Transformers.
Who it is for: Developers needing local English dialogue speech synthesis with voice cloning
Strengths
- Generates two-speaker dialogue in one pass using [S1]/[S2] tags
- Voice cloning from an audio prompt plus transcript
- About 4.4GB VRAM in bfloat16/float16, ~2x realtime on RTX 4090 with compile
- Apache-2.0 weights on Hugging Face, with a Transformers integration and Gradio UI
Weaknesses
- English only
- Tested on GPUs only; CPU support listed as not yet available
- No fixed voice: speakers differ per run unless seed or audio prompt is set
- Very short or long inputs sound unnatural; unlisted nonverbal tags cause artifacts
What it needs
- GPU required
- Needs PyTorch 2.0+, CUDA 12.6, Descript Audio Codec
- Models: Dia-1.6B-0626, Hugging Face Transformers
Also in Voice
See all 51| Rank | Project | Score |
|---|---|---|
| 1 | voicestudioDesktop app for voice cloning, dubbing and dictation with local speech models | 83 out of 100 |
| 2 | xiaozhi-esp32ESP32 firmware for a voice chatbot with MCP device and cloud tool control | 80 out of 100 |
| 3 | whisper.cppC/C++ port of OpenAI Whisper for local speech-to-text transcription | 78 out of 100 |
| #4 | VoiceboxLocal voice studio for cloning, TTS, dictation and agent speech | 71 out of 100 |
| #5 | buzzOffline desktop app for transcribing and translating audio with Whisper | 71 out of 100 |