Dia

Dia is a 1.6-billion-parameter open-weights text-to-speech model from Nari Labs — a two-person,

Korea-based startup founded by undergrads Toby Kim and Jay — released in April 2025. It directly generates

highly realistic dialogue from a transcript in a single pass: multiple speakers with distinct voices, emotional

tone control via audio conditioning, and non-verbal vocalizations (laughter, coughs, throat clearing) driven

by plain-text tags.

Provider-claimed quality: Nari Labs says Dia rivals NotebookLM-style podcasts and

surpasses ElevenLabs and Sesame-1B in naturalness and prosody, based on comparisons on the team's own demo

page — treat as provider-claimed, since no standardized independent TTS benchmarks exist for it. Independent

TechCrunch hands-on testing found it "worked quite well," with voice quality competitive with other tools and

"among the easiest" voice-cloning the reporter had tried.

Released under Apache 2.0 (commercial use allowed),

it runs on most modern PCs with ~10GB of VRAM. English only. A successor, Dia2, shipped in November 2025.

Independent: No standardized benchmark scores exist for Dia. TechCrunch hands-on testing (Apr 2025) found quality competitive with leading tools and voice cloning "among the easiest" the reporter had tried.

Provider-claimed (Nari Labs demo page): "rivals NotebookLM's podcasts and surpasses ElevenLabs and Sesame-1B in naturalness and prosody" — based on the team's own demo comparisons, not an independent benchmark.