Dia

nari-labs • Updated 2025 May 30 21:19

, No training script, just TTS model

(laughs), (clears throat), (sighs), (gasps), (coughs), (singing), (sings), (mumbles), (beep), (groans), (sniffs), (claps), (screams), (inhales), (exhales), (applause), (burps), (humming), (sneezes), (chuckle), (whistles)

RoPE

RMS Normalization

Grouped-query Attention

Byte Level Tokenizer

CFG Scale

Delay Pattern

A type of

Positional Embedding used to distinguish individual channels. Audio is represented and generated as multiple channels (default 9) of code sequences rather than a single stream. For example, if the first channel generates information at time t, the second channel generates at t - delay[1]. When handling multi-channel audio codes, this defines the rules or mechanisms for how each channel references temporal information to generate the next code.

Nari Labs: Dia Examples | Notion

Comparison between Dia-1.6B (ours), ElevenLabs Studio, and Sesame CSM-1B. Plus fun examples (including audio prompt use).

https://yummy-fir-7a4.notion.site/dia