roxy

- ai yapper by default 👽
- building voice agents for a living 🗣️

[roxy @arjun__ai]

Been seeing double responses in @OpenAI @OpenAIDevs for gpt 5.4 api for a while after it’s release. This is bugging me in production with random repeated responses. Why is it still unresolved?

[roxy @arjun__ai]

@ElevenLabs @ElevenCreative Waiting for realtime avatar apis!

[Dileep Venugopalan @Dileep_Eth]

Hey @royalenfield - a thread about how your customer support handles a stranded rider on a warranty-covered motorcycle.
My Himalayan 450 (KL 52 U 2633) failed mid-ride on 9 June with ECU-TPS failure. Stuck at Teknik Motorcycles, Sarjapur Rd, Bangalore since.

[roxy @arjun__ai]

Mole cleared 110gb of storage in my system. Holy

[roxy @arjun__ai]

@aasishrajrr Don’t tell me u gonna be a dad😭

[roxy @arjun__ai]

@Elevenlabs be doing everything except improving their TTS. Comeonn u guys are lacking there. Pullout v3 out of beta… add streaming end point to v3…more practical emotional/sound tags etc can point out a bunch of niche flag bugs too lmk

[ElevenLabs @ElevenLabs]

Today we are launching ElevenMusic, a new platform to discover, remix, create, and earn from music, built on the ElevenLabs music model.

[roxy @arjun__ai]

@mishalabdula @aasishrajrr @ProductHunt 🎉🫶

[roxy @arjun__ai]

Remember when people used to run Doom everywhere? It’s that exact moment…history repeats

[Bhavy☄️ @Bhavani_00007]

Rate my setup guyys

[roxy @arjun__ai]

@aniketshah30 Fan of rice n chicken..Congrats team❤️

[roxy @arjun__ai]

@SubinSiby @sama 💆‍♂️

[roxy @arjun__ai]

@Mayank_022 Goat is back💆‍♂️

[roxy @arjun__ai]

> I was out to Shillong for five days.
> Uninstalled X cuz my storage was full.
> Reinstalled back now, sees Clawdbot, moltbot, moltbook, molthub bs everywhere.
> Uninstalls again.

btw here’s rainbow falls in Meghalaya

[roxy @arjun__ai]

Where's @StackOverflow at?

[Yash Bhardwaj @ybhrdwj]

Tailwind lays of 75% of their team. the reason is so ironic:

> their css framework became extremely popular w AI coding agents, 75m downloads/mo
> that meant nobody would visit their docs where they promoted paid offerings
> resulting in 40% drop in traffic & 80% revenue loss

[roxy @arjun__ai]

This might be the most interesting thing I've seen in Huddle global meet this year

[roxy @arjun__ai]

@elder_plinius You are absolutely NOT right !

[roxy @arjun__ai]

Sad to say that @instagram was never built for photographers or for posting photos. The quality drop in posts are just insane.

[roxy @arjun__ai]

So neuroplasticity is the thing left for us to solve in llms.

[Mayank Pratap Singh @Mayank_022]

Before jumping into 𝐕𝐞𝐜𝐭𝐨𝐫 𝐐𝐮𝐚𝐧𝐭𝐢𝐳𝐞𝐝 𝐕𝐚𝐫𝐢𝐚𝐭𝐢𝐨𝐧𝐚𝐥 𝐀𝐮𝐭𝐨𝐞𝐧𝐜𝐨𝐝𝐞𝐫𝐬,

let me explain why I implemented this from scratch.

For the past few months, I’ve been deep into audio deep learning:
understanding how text-to-speech, speech-to-text, and audio-to-audio models work.

Along the way, I covered mel spectrograms, Fourier transforms, signal processing basics, and even infra pieces like WebSockets and WebRTC.

A few weeks ago, I implemented a 𝐓𝐓𝐒 𝐦𝐨𝐝𝐞𝐥 𝐟𝐫𝐨𝐦 𝐬𝐜𝐫𝐚𝐭𝐜𝐡. I initially coded NVIDIA’s Tacotron2 architecture

Code
https://t.co/O29ldXlD6X

but instead of training it, I moved on to a 𝐓𝐫𝐚𝐧𝐬𝐟𝐨𝐫𝐦𝐞𝐫-𝐛𝐚𝐬𝐞𝐝 𝐓𝐓𝐒 𝐦𝐨𝐝𝐞𝐥, which I trained for ~24 hours on an A100. [will share the code stuff in the future]

After that, I started building a speech-to-text model.

While working on the STT architecture, one block kept appearing repeatedly:

𝐑𝐞𝐬𝐢𝐝𝐮𝐚𝐥 𝐕𝐞𝐜𝐭𝐨𝐫 𝐐𝐮𝐚𝐧𝐭𝐢𝐳𝐞𝐫𝐬

To understand RVQ, I needed VQ-VAE.
To understand VQ-VAE, I needed Variational Autoencoders.
And to understand VAEs properly, I first implemented a vanilla autoencoder.

So I ended up implementing all three:
Autoencoder → Variational Autoencoder → Vector Quantized VAE

𝐀𝐮𝐭𝐨𝐞𝐧𝐜𝐨𝐝𝐞𝐫𝐬 compress data into a latent space and reconstruct it back, while 𝐕𝐀𝐄𝐬  add a probabilistic structure by learning a distribution and regularizing it with a prior, enabling generation but often causing blur and instability.

AE Code
https://t.co/Fvua50p5oH

VAE Code
https://t.co/yzF2RPHKWD

𝐕𝐐‑𝐕𝐀𝐄 changes this by making the latent space discrete. The encoder maps inputs to the nearest vector in a learned codebook, turning each latent into a token from a fixed vocabulary.

VQ-VAE Code
https://t.co/kKaENveqHg

Why this matters:
• Speech and audio are inherently discrete
• Quantization yields efficient compression
• Discrete tokens integrate seamlessly with Transformers

Unlike VAEs, VQ‑VAEs use a deterministic encoder, no KL loss, and learn the token prior separately. They train via reconstruction, codebook, and commitment losses with straight‑through gradients , making them fundamental to speech synthesis, recognition, and modern audio tokenizers.

I’m going to implement RVQ next and share a detailed post about building TTS and STT models from scratch, voice cloning, and fine-tuning audio models.

If you’re into deep learning or computer science, 𝐟𝐨𝐥𝐥𝐨𝐰 𝐦𝐞 for insightful content

— and don’t forget to 𝐫𝐞𝐩𝐨𝐬𝐭 this!

[roxy @arjun__ai]

Had this deep convo on male character development, how increasingly impressionable someone becomes with loses/heartbreaks. That feeling of losing someone who knows 99.9% of you and the ego that fires up to prove that everything they knew about you is wrong by evolving yourself.

[roxy @arjun__ai]

@dsa @livekit As a voice agent dev, playing around with @livekit agents is therapeutic.