Well, my project has been cancelled, and left at the moment where I wrote only description. Because I got some really not good news, I had to turn my focus on math.

Not everything is bad though, I got into internship in a Startup company and actively exploring new things in this field.

Over the past week, I had been working on a script that cuts audio based on the speaker. The goal is to build a dataset and automate the process of data mining, since manually processing large amounts of data would take too long. To be honest I never thought it's so hard to collect data from scratch.

The first problem I faced was library incompatibility. One library conflicted with another’s version and so on. And Python 3.12 was not supported by several audio libraries. As a result, I had to dig through the documentation to find compatible versions.

Then I found out that the processed audio has shitty output quality. After unsuccessfully trying to improve it, I realized it's better use a ready model, a working solution that already exists.

Also it changed my view on vibe coding. It works well as long as you understand what’s happening. AI is genuinely useful for generating code that would otherwise take hours to write or when you’re working outside your main field. However, for larger projects or when it's time for deployment, the code should ultimately be written by a professional.

So, while doing the job from last post, I tried several common models, and want to share the ones that showed best results.

For separating background and voice I used Demucs, the last version. And installed with Python 3.10, with newer versions it's not working. Also there is an API that improves voice better than Demucs, but it's not free.

And for defining the speakers, pyannote community-1 version(3.1 works worse compared to community) did decent job even though sometimes it considers coughs or laughter as a separate sound.

Learning is a skill itself, we have to learn how to learn

Today I finished what I've been doing for 3 weeks. Finally, it's a huge relief.

I wrote previously about what I had done. Next, I needed to write the transcription of audios and define what emotion the speaker uses while speaking.

Whisper small uz v1 does great job with transcripting uzbek speech (so far the most accurate model for uzbek speech).

Overall, there wasn't much problems. The main issue was the speed of processing. Even though my GPU worked on 100%, it took 20 hours to finish. And it's not all the audio yet 🙏

This is the best roadmaps for now, on the video, the author lists all necessary instruments and explains why they are needed.

I am not posting lately. But soon I will post my program that anyone can use, tho I doubt it's useful for ordinary people

To be honest, the idea of channel where I post what I do was wrong. In my opinion, I am not specialized enough to do it yet.

From now on, I decided I will post stuff I find useful and sometimes projects I did.

Maybe after couple of years I will be able to guide people.

SO MY PROJECT IS READY

EVERYONE USE IT http://localhost:8000

Okay this time it's real
audioprocessorbysardor.app

I don't think ordinary people find it useful because what it basically does is making audio dataset and pushing to server.

The main problem I faced was with my server's memory, as I am using it for free, and I had to cut some of libraries installation

Я говорю Claude: «Ты сеньор веб-разработчик с 20 годами опыта»:

The end of unemployment

Working with a great view