People seemed to be enjoying the fire example, so I'm also adding it to the thread here. Thanks for all the support :)))

Two undergrads. One still in the military. Zero funding. One ridiculous goal: build a TTS model that rivals NotebookLM Podcast, ElevenLabs Studio, and Sesame CSM. Somehow… we pulled it off. Here’s how 👇

No, we were not AI experts from the beginning. It all started when we fell in love with NotebookLM’s podcast feature when it was released last year.

But we wanted more. More control over the voices. More freedom in the script. We tried every TTS API on the market — none of them sounded like real human conversation.

Our biggest blocker? Compute. But Google came in clutch, giving us access to TPUs through their Research Cloud. Huge shoutout to TRC ❤️

Of course, TPUs are HARD. We had to learn everything — JAX, Flax, parallel compute, cluster orchestration, Pallas kernels and more.

DeepMind’s How to Scale and HuggingFace’s Ultra-Scale Playbook were super helpful. If you are interested in training large models, go read them now!

3 months later, we had a fully trained 1.6B model. It took longer and was harder than expected, but totally worth it. The best time to start is today!

And now, we’re turning Dia into a B2C app. Generate fun convos, remix content, and share with friends. 🔓 Join the waitlist for early access:

Try Dia now! Github: HF:

Sesame is stoned out of its parameters

Thanks for open sourcing this! It would be amazing if you could also share a technical report of your findings while training the model.

How to train it to speak other non-english language? To be specific: Bahasa Indonesia

lol you 100% used Rick and Morty samples heavily to generate Fire

Absolutely mogged 😭

😂😂🫵😁. Listening to this fire 🔥 example, Sesame CSM-1B was not getting anywhere closer to the Nari Dia-1.6B 😁

Hilarious. Congrats on the launch this looks awesome!

Congrats! Did you use synthetic data for training? How many hours? Did you use dialogs from movies and annotated them by prompting Gemini Flash?

This one so fucking cool

On which dataset did you train?