Dia

Dia is a 1.6-billion-parameter text-to-speech model from Nari Labs, a two-person startup founded by Korean undergrads, released April 2025 under Apache 2.0. It directly generates ultra-realistic dialogue from a transcript in a single pass, with speaker tags ([S1], [S2]) for two-speaker conversations, emotion and tone control via audio conditioning (zero-shot voice cloning), and nonverbal vocalizations — laughter, coughs, throat-clearing — that most TTS models skip or read aloud literally. It runs on consumer GPUs (~10GB VRAM) and reached ~19K GitHub stars within months.

The honest picture: Dia only supports English, and its nonverbal tags are probabilistic — cues are requested, not guaranteed, and can produce unexpected output. It has no whisper mode, no intensity control, and no speed parameter. For long-form single-speaker narration, the smaller Kokoro-82M is often the more practical pick; Dia's strength is specifically naturalistic multi-speaker dialogue. A Dia 2 was announced in late 2025, so check which checkpoint you are pulling.