Choose a Category
Hear Say!?
Who said that?

Try Hear Say!? on Reddit Now! Who said that?

Inspiration

The internet has gone mad for Ghibli memes since OpenAI's latest image model was released earlier this week. The ultimate irony is that Hayao Miyazaki himself was deeply opposed to the use of AI to replace artists. However art is just scratching the surface, deepfake audio can now easily replicate anyones voice with only a clip of a few seconds, and with this there is a risk that we'll lose track of reality all together. We thought let's put together a quiz of famous quotes and use deepfake tech to have them said back by all of the possible options. We gathered up some of the most popular Studio Ghibli memes from the past 48 hours and used these characters to put together Hear Say!? a quiz like no other...

Hayao Miyazaki. Hayao Miyazaki, co-founder of Studio Ghibli, was disgusted by a demo of AI generated art he was presented in 2016

What it does

Redditor is presented with a quote said by one of the personalities listed and has 20 seconds to figure out who said it

On selecting a personality a deepfake of them saying the quote is played

After the audio plays, the Redditor is awarded points according to how fast they got it right, or none if they got it wrong

How we built it

We used devvit to set-up a webview which was necessary due to the use of Web Audio API to play back the audio, Web Audio API is not available in blocks. The deepfake audios were generated with E2/F5 TTS using audio recorded from YT with Blackhole and the Ghibli images were scraped from X (who needs GPT4.5 Pro when you've got internet access?).

Deepfaking George W Bush by recording YT audio of him talking and generating clips with E2/F5 Text-to-Speech

My initial idea was to vibe code the whole thing, however evidently Claude hasn't been extensively trained on Devvit documentation so kept trying to run the app locally or started leveraging libraries/frameworks that aren't compatible with Reddit. Setting-up specific system prompts with Devvit AI to fine-tune Claude's results saved a lot of time in the end.

Challenges we ran into

We were originally planning to give redditors mic access, however it appears that Devvit does not yet enable microphone permissions, so we changed the scope of the game a bit so the users own voice input wouldn't be required. We were intending to set-up scoring and leaderboards during the hackathon but got too bogged down in Devvit fundamentals, optimizing the front-end and overall usability that we didn't actually get to Devvit's playtest feature in the end.

Accomplishments that we're proud of

The quiz itself is pretty neat and timely, hopefully it challenges peoples assumptions a bit with some of the quotes, and it's super fun to hear the deepfake voices say them back. Can't wait to see what people make of it.

What we learned

Devvit 101's

How not to vibe code

What's next for Hear Say!?

Setting it up scoring and leaderboard system. Enabling anyone to create their own quiz, ideally with an agentic backend so audio retrieval and deepfake creation can be completely automated within the app itself so user can publish in seconds.