Two undergrads built an open-source speech AI rivaling ElevenLabs
Korean startup Nari Labs just released Dia, an open-source 1.6B text-to-speech model built by two undergraduates with zero outside funding. The model handles emotional tones, multi-speaker tags, and nonverbal cues like laughter and coughing, outshining commercial leaders like ElevenLabs in benchmark tests. It proves solo builders can now challenge billion-dollar labs.
Dia is a new open-source 1.6B parameter text-to-speech model from Korean startup Nari Labs that delivers remarkably expressive, podcast-quality speech without a commercial price tag.
If you have been paying for ElevenLabs or marveled at the realistic conversational banters in Google NotebookLM, Dia aims directly at that exact territory.
What it can do
Dia stands out most in handling conversational messiness and emotional nuance.
Traditional speech models often stumble over script cues like pauses, sighs, or laughter, either reading them literally or generating glitchy artifacts. Dia supports native multi-speaker tags, allowing seamless multi-person dialogue within a single generation pass.
More impressively, it natively renders complex emotional tones and nonverbal cues like natural laughter, coughing, and even screams.
In side-by-side benchmark evaluations against commercial heavyweight ElevenLabs Studio and open-source rival Sesame CSM-1B, Dia came out ahead in conversational timing, overall expressiveness, and script adherence on nonverbal audio.
How it was built
The most striking part of Dia is who built it. The model was trained by two Korean undergraduates with very little prior AI engineering experience, operating with zero outside venture funding.
Inspired by the viral audio overviews inside Google NotebookLM, the duo set out to recreate that lifelike audio dynamic. To overcome prohibitive compute costs, they applied for compute access through Google's TPU Research Cloud program.
Nari Labs founder Toby Kim noted that releasing the open weights is just the foundation. The startup is already working on a consumer app designed around social audio creation and remixing.
Why it matters
For indie developers, podcasters, and game creators, Dia is worth looking at immediately. High-fidelity expressive voice generation has long been locked behind steep token-based commercial APIs; an open-source 1.6B model competing on equal footing dramatically lowers development costs.
More broadly, Dia is proof of how radically accessible cutting-edge AI development has become. When two college undergrads with zero funding can challenge a multi-billion-dollar market leader using public cloud grants, the barrier between frontier labs and solo tinkerers has effectively collapsed.
Sources (1)
- @_doyeob_(opens in a new tab)Originalx.com