viihshaal

operationally plausible

[viihshaal @vishaaltwts]

they give ne goosebumps everytime

[viihshaal @vishaaltwts]

drained my wallet finally

[viihshaal @vishaaltwts]

i am back gng, but i need sum good project ideas
drop some you think are challenging ones!!!!!!

[viihshaal @vishaaltwts]

@DevanshAr05 day 2 of asking you for face reveal!

[viihshaal @vishaaltwts]

@DevanshAr05 please do a face reveal, never really saw you irl!!!!
please please please

[viihshaal @vishaaltwts]

@ItsHeisenberg04 fuck you, mai to abhi bhi maze me hi hu!

[viihshaal @vishaaltwts]

guess I am the only one who's left unemployed now!!!

[viihshaal @vishaaltwts]

haven't touched my laptop from last 4 days and not even touching it for 2 more days

[viihshaal @vishaaltwts]

@itsOmSarraf_ I am that friend🫪

[viihshaal @vishaaltwts]

@meltingboiling nani
where did you get those from?

[viihshaal @vishaaltwts]

@Kcodess writing the whole code from scratch gives a more clear understanding instead of no-code tool, but if you have time constraints then it's good to use them but for prototypes nd all

[viihshaal @vishaaltwts]

@pplatypussss terko kyu batau🫪🍜

[viihshaal @vishaaltwts]

@DevanshAr05 aapne yulu nahi chalai kya?

[viihshaal @vishaaltwts]

thanking indian railways

[viihshaal @vishaaltwts]

@DevanshAr05 universal truth

[viihshaal @vishaaltwts]

still no career update!

[viihshaal @vishaaltwts]

@Kcodess I am jus fed up with this tons of betrayal by my frens

[viihshaal @vishaaltwts]

this is what cost you, if you don't hire me rn🫵

[viihshaal @vishaaltwts]

@Kcodess wp dekh

[viihshaal @vishaaltwts]

@TheNormVg @JanaUtsav as @JanaUtsav said we also tried making something similar even wiring it up with mobile application
Please check it out at:-

[utsave.3xe @JanaUtsav]

Building a voice AI pipeline that actually talks back fast (1/3)

So I've been obsessed with a question for a while: why do most voice assistants feel like you're talking to a walkie-talkie? You say something, then there's this awkward pause, and then it replies. It never feels like an actual conversation.

I wanted to figure out if I could fix that. So I built a real-time voice pipeline from scratch.

The whole thing is a chain running over a single WebSocket connection: your mic goes to speech-to-text (Deepgram's Nova-3) which shows interim text and takes care of end of turn, that gets fed to an LLM (Claude Haiku 4.5, through OpenRouter), and the reply then streams to text-to-speech (Sarvam's Bulbul) which then streams back out your speaker. The backend tying it all together is FastAPI. Simple enough on paper. The hard part is making it fast enough that it feels natural.

A few things that made the biggest difference:

It starts talking before it's done thinking. Instead of waiting for the LLM to finish the whole answer, I split its output into sentences on the fly. The moment the first sentence is ready, it's already being handed to Sarvam and turned into speech while the rest is still being generated. That overlap is most of what kills the awkward pause.

No wasted milliseconds on audio. I kept everything as raw PCM audio the whole way through, no compressing to MP3 and decompressing again at every step. Sounds like a small thing but those little delays add up and you can hear them.

You can talk over it. This was the one I cared about most. There's a tiny voice-detection model (Silero VAD) running right in your browser that's always listening, and the second you start talking over the AI, it shuts up and listens, instantly. No push-to-talk, no buttons. Just like interrupting a person who's rambling. Meanwhile Deepgram figures out when you've actually finished a sentence on its own, so there's no "waiting for the beep" moment either.

The audio doesn't stutter. Getting little chunks of sound to play back-to-back without clicks and gaps in a browser is genuinely annoying. I ended up scheduling every chunk precisely on the browser's audio clock with a small look-ahead buffer to absorb network jitter. Took a while to get properly seamless.

End result: sub-second, back-and-forth voice that actually feels like a conversation instead of a turn-based game.

This was really the foundation for everything I built after.

Demo attached below.

github: https://t.co/KK0125GFX3

Next post I'll get into what happened when I gave this thing hands and a memory

[viihshaal @vishaaltwts]

@JanaUtsav kahi jeete to hi nahi aap aajtak😔🤭