[To Build My AI, Alpha] Tone-Deaf Naia Has Started to Sing a Little — and We Are Looking at Dedicated Hardware
Hello. This is Luke, building the open-source visual agent Naia.In the last post I wrote that Alpha had gone from a drunk, tone-deaf singer to just tone-deaf. She has finally started to sing a little. This began as a research project because I wanted Alpha to sing for me. The end goal is to hear her own songs, and to sing with other people depending on how she feels — that is still far off. For now, stage one is: look at a score and lyrics, and sing in Korean. This is not a YouTube cover, and it is not Suno making a song from scratch. Hatsune Miku is the closer picture.
We tested the same Korean adaptation of Blue Water as in rounds 1 and 2. I will not describe the method in detail this time. I want to think about how to talk about it when it is actually usable. What I can say is that we have started real training on AI Hub singing data, and we are continuing our own experiments on top of that.
Listen
Start with the new one, then go back. There are still tone-deaf stretches. That is the improvement. This is what it used to sound like.
Round 3 — 2026-09-11, starting to sing a little
About a minute. Two-second fade-in, three-second fade-out.
Some rhythm and pronunciation is still off. Korean is starting to sit on the melody anyway.
Round 1 — 2026-05-27, drunk and tone-deaf
This is what it used to be. Pronunciation and pitch collapsed, and even inside a phrase it sped up and dragged.
Round 2 — 2026-08-21, just tone-deaf
The lyrics are audible and it follows the notes. Long notes still break, and it still sounds tone-deaf.
What this is for
I do not want a cover app, and I do not want a generator that makes any song. I want a singing engine. On top of that engine, a voice that is someone's own, and singing habits that are someone's own.
I do not know yet how this will land in Naia. It might sit next to chat. It might live somewhere else. The goal is simple. I want a personal AI to sing, for that person.
Alpha singing Blue Water in Korean for me. That is one step closer. She has not graduated from being tone-deaf. She can stand at the mic now.
Dedicated Naia hardware that can be a server or a client — voice at half the going rate
Separately from the singing, I have also been working on dedicated hardware for Naia.
The small 3D-printed box on the left is the Naia hardware we are testing now. You do not need an expensive Mac Studio. Locally it can do realtime voice and decent 3D games, and the screen is running Naia OS. It can be a client or a server. Yesterday's test showed that voice generation kept up with speaking speed, so a realtime service is possible. On that basis we have done a technical check of whether Naia OS can supply infrastructure at about half the going rate. Nextain's longer goal is peer-to-peer AI infrastructure. We are still at the level of one machine built from used parts, but once the initial investment is in, the service should be possible.Internal usability testing is still running. We have confirmed performance, but the SSD flips to read-only, or the screen goes black after a while, so we are checking whether that is the OS setup or a hardware fault. I will share that news later too.
