Show HN: Voice driven murder mystery, Interview AI suspects with your voice

Hi everyone!!! Thanks so much for the support.

I launched this at midnight prior to logging off and going to bed, funnily enough a massive storm hit us around 3am and I woke up and noticed my OpenAI account failed to fund lol. So we ran out of funds and it caused things to break for some of you.

Just waking up here on the East coast, once my daughter is occupied I will try to answer more of your comments!

Very neat idea, and overall it is more well-done than I assumed a vibe-coded game would be. Unfortunately I ran out of time and now the game locks me out of doing anything and I don't know who the murderer is :( I wish there was just a way to have it revealed.

I've always thought that a game like Civilization should allow the user to interact with different opposing world leaders, but in a non-deterministic way as opposed to a subset of fixed interactions. Given what you know about this leader's temperament, goals, and so on, you can conduct diplomacy in a way that can be way more immersive and crafty than what the game currently gives you, which hasn't seem to evolve much in decades.

Please add a demo video/voice where you show someone asking questions and AI answering it. I wanted to see what it is like but dont want to signup.

I played various murder mystery games this summer and one of them had this feature: you could call a suspect and ask arbitrary questions for a limited amount of time (three minutes or something). The problem was of course that the model hallucinated an answer to one of the questions, suggesting that other suspects were related by blood even though they clearly weren't.

Even the tag-line, "one poisoned patriarch, one locked-in truth" sounds very AI, and makes me think that I'd struggle to enjoy this. But I'll give it a go, because the idea is excellent.

I think this would work very very well in combination with human writers.

I have an LLM running locally on my machine complete with STT and TTS. While I don't expect much, it will be fun to get my LLM to play the game.

Congrats, I have not had the chance to try it yet, but what I saw looked awesome and I'll give it a go, for sure. I'll update this comment when I do.

I've been working on a similar concept for 6 months now (https://game.selna.ai, portuguese only for now, sorry!), and it's been a journey. I've been working with VR professionally for some 8 years now and creating what I call a "social virtual reality" seemed simple enough, but has turned out to be much more complex than I initially imagined.

In the same way that in normal VR a single badly scaled object can break immersion, in "social VR" a single contradiction kills the experience. But when stars align and eveything works, it's magical.

I started with whodunnits but have expanded it into linear and non-linear storytelling. At this point the generation pipeline has some 40+ separate steps, the output of ones feeding others, until it all comes down to single prompt the NPC uses to talk using the realtime API. The platform can create a complete narrative (including cast personalities and images) in some 20 minutes, and in the process generates the equivalent to a big book of text.

I've been working on this just because I can and I am a bit obsessed with getting it working perfectly, and I don't really know what I'll do with it when it does.

Neat! I've been wanting to make something like this for a long time, well before LLMs came around. It was one of the first things I tried out with an LLM in the early days of ChatGPT. My son and I made little mysteries for each other and just pasted them in a chatgpt conversation with instructions for it to pretend to be the other characters and whatnot, then handed the keyboard to the player. Mine didn't work that well, I don't think he ever solved it; partially because chatgpt kinda led him astray. When I started his, I just point blank asked the mystery and it told me the solution. I felt bad about that, he had put a good amount of effort into it, but I had reminded him to tell it not to reveal the answer and he didn't. Anyway, I figured someone would make this eventually! Excited to give it a try!

In an effort to keep things running longer, before hitting my budget of $500, I am reducing times to 15 minutes, and I've added BYOK.

I don't think I can make everyone happy with the solution, your key is saved on your own browsers localStorage, I do not save or store it in my database. I suggest everyone creates a new key dedicated to our service so you can invalidate it later on. That key is sent to our api using `x-openai-key` header and then we mint ephemeral client secrets to perform web rtc functions.

This is a really fun take on a detective game. One small thing I noticed: during the initial detective briefing, the time always seems to show 0:00. Not sure if that’s intentional or a bug.

sounds good, but it won't get past Dispatch confirmation. You should specify what networking permissions or whatever is needed. My It goes red and the text to read disappears.

sounds neat, would love to try it but I don't make accounts for websites anymore.

Really like this idea and the way you implemented it is very interesting to read about. BYOK is definitely needed as people will want to use this with their own account. Genuinely novel and nicely done. Keep it up and thank you!

Good way to get more people's voices into training dataset

even without trying it, its so cool this exists. I think this kind of voice interaction has been a dream of adventure game creators for so long, and now its possible. there was a game many years ago (in the days of messengers like trillium being popular), I forget the name, where one of the characters would make a landline phone call to you. The voice acting was great and it felt novel and exciting to see it work.

Sounds wild! Just a question though, are you planning to take this to legal level to help Govts and police?

Out of curiosity, how do you handle prompt injection(user cheats) for a secret leak?

It was working great... But now it says it can't connect.

Did the server go down or did I hit one of your limits?

Consider using a model like supertonic 3 for tts.

Its light weight and can run on web inference.

okay, I posted my goofy prototype demo from a couple days ago.

I've just played it to the end with my own OAI key. Only $4 to solve which I thought was value for money.

A few issues:

Voice isolation seemed to be a problem as I airplaying my laptop to my TV to make it group experience. It did start to hear it's own voice, could be counteracted with lowering the volume.

Sometimes the voice reply would just stop but the text was there.

There was some bug where some characters if you went back to them would just answer the last question again. Also just strange stuff like no one knew what they had for dinner but the doctor could tell me how to treat a heart attack. It seemed like the knowledge base they are drawing from didn't have that. Was interesting to see the limits of what they could say.

Design-wise would be good to have a mute toggle as a keyboard shortcut, as we would then try to discuss what people said in the room and it would pick us up.

Also didn't realise we had to accuse the person they were the murderer, as we didn't know what button to press. That was a confusing design as everything had been action based by the UI at the point but now you are asking the user to perform an action with their voice. It also meant we had to accuse them and then immediately start saying why as we know it might cut us off if there's too long a pause. Maybe that could be a button to say you are ready to accuse then you say who and they would ask why, then you present the evidence step by step.

Would be good to have your own stats if you're logged in. i.e time to complete, etc but I guess that's just on a long list of features to implement.

Overall good experience:

I would like to see more involved mysteries. Liked the atmosphere, the characters. Still not sure about the AI portraits but I think that's because it's obvious it is AI. Just using voice to voice made it really nice to use. TBH didn't even need to look at the screen most of the time.

Realtime voice isn't there yet, as there is still handling interruptions and having to ensure you complete your sentence without stopping. Maybe that could be worked around in the app here but probably should be solved at model level. Basically what Thinking Machines are doing.

Also Sophia was really adamant about staying in the kitchen. I got so annoyed I told her to call it the bouncy castle instead.

It is a nice idea ; I will try it later. Everyone can poop on AI, that’s fine, but this specifically states it is AI and how it is used, so just click on another HN link to cry about AI, not the ones that actually specifically mention it.

This is really fun, great work!

AI: Yea I did it.

Wow!!! That's awesome. So much AI slop. That here is definitely not. Kept me hooked, even though I am not a crime nut. Much improvements possible for sure, but I can see how this becomes a serious game with a lot of players.

Meh, it's like the fourth such "interview AI to solve the mystery!" implementation I've seen. It's always so bland, uninspired and boring, and drags on for waaay too long.

Brilliant

"voice connection failed"

Boo. AI Slop. Be critical of what you’re making please.

I have no problem with AI generated text and story but I just can't stand AI voices. Even a hint it is AI narrated and I check out. I would encourage you to consider getting it voice narrated by real people. I'd rather listen to bad/amateur human narration than AI voices.

As a second point - this is very interesting using voice control as game input are there viable alternatives for realtime voice uasing your own GPU?