update on my ai waifu app, can use web search react to images even picture of herself

Posted by aziib@reddit | LocalLLaMA | View on Reddit | 33 comments

using qwen 3 VL for the llm and the vision (really good for recognize popular characters and even recognize their appearances)

using SerpApi for the web search

the tts is using omnivoice tts (support 600+ languages) that i make a custom api that i recently open source it, get it here: https://github.com/aziib/omnivoice-tts-api

my ai waifu project stil in work in progress, i just hope there is free web search api, SerpApi has some search limit usage per month.

[-]

martapap@reddit

It always creeps me out what some ppl reduce women/feminine figures too.

[-]

Leafytreedev@reddit

Ah, I see. So this is why ram is so expensive.

[-]

duckrollin@reddit

The stick arms and massive boobs make her look kinda hideous tbh. The voice and latency is great though.

[-]

Angryceo@reddit

its the teenage/basement dwllers dream.

[-]

duckrollin@reddit

I feel bad for her, she reminds of the dog breeds that were bred to have such a short snout that they can't breathe properly anymore.

[-]

Roth_Skyfire@reddit

Please think of the virtual characters.

[-]

Angryceo@reddit

hey now! i have two french bulldogs lol

[-]

Nelson-Bolt@reddit

Shit it's even worse with audio turned on.

[-]

Nelson-Bolt@reddit

so that means they can not call something cringe?

[-]

Don't you know? "Hidden profile", "You don't even own the game" are top final bosses among arguments on every Steam game discussion. Keep in mind setting profile to private / hidden there may also hide the icon which shows next to the username in their comment and indicates that they own the game. Also, even when the profile is not set to private, the absence of the icon just means the user doesn't own the game on Steam, but they still may own it on different platforms, so that's another weak point of the argument. 😂

[-]

Velocita84@reddit

It's cool and all but i just can't with the english voice, a real waifu should speak japanese, i think even just translating the english output with a small non-llm translation model and passing that to a capable japanese tts should be enough

[-]

aziib@reddit (OP)

she can speak japanese if i ask, it's just my choice, omnivoice is support 600+ language include japan so it shouldn't be a problem. https://x.com/i/status/2042980568204480791

[-]

Velocita84@reddit

Just having the main llm output japanese would kinda... Make it hard to understand if you don't speak japanese. I'm talking about that visual novel vibe where the voicelines are in japanese but the text in english.

[-]

aziib@reddit (OP)

it is possible to use translator by translating the generated ai text back to english so she still speak japanese but the text window is in english and user message translated to japanese before sending to ai, i've done that in silly tavern. but nah too much work.

[-]

firest3rm6@reddit

I think you should check out moeru-ai on GitHub. Or did you fork from there?

[-]

aziib@reddit (OP)

no, it is completely from scratch and with some vibecoding in antigravity.

[-]

firest3rm6@reddit

For the web search your current solution has 100 free searches per month. You could use Serper.dev which is newer, 2500 free searches per month and I think also cheaper. (If money is the bottleneck)

[-]

aziib@reddit (OP)

thanks, will add this too so i can test it more.

[-]

snzo@reddit

seek treatment

[-]

aziib@reddit (OP)

this is the treatment

[-]

aziib@reddit (OP)

thanks, will open source if it's ready, still got some bugs and still searching best speech to text that has no delay for talking with my ai.

[-]

Nyghtbynger@reddit

Top Goon

[-]

Warm-Put3482@reddit

dem