Hey Hacker News,

I built FlowTunes, an app that plays music designed for focus. I'm scratching my own itch, but figure folks here might appreciate it, too!

It has some unique features I wanted over similar apps: - It's free forever, sans ads. - Effectively endless music (more will be added faster than you can listen). - Mix in custom background soundscape, such as birds or streams. - Ultra-minimalist interface.

The music (3000+ tracks across 10 channels, to date) was generated with Suno AI, which was an interesting experience, worth sharing.

First of all, I have mixed feelings about AI-generated music, and I'd love to hear your thoughts! To me, it feels more appropriative than text and images. Probably because music is an art-form that really moves me. Somehow, though, I feel more ok about instrumental-only music, as used here. I'm not saying that's rational, as it's clearly valuable art without the words and voices. But it feels... less bad...? Your thoughts?

On the technical side, I'm probably amongst the heaviest creators of AI-gen music (about 20K tracks or so) and I'm happy to share some learnings:

1. Suno still seems to have the highest and most consistent output quality.

2. Overall, the AI just isn't that great yet, regardless of which tool you use... Like most things AI, it seems amazing until you actually look at its output critically. It's almost there for instrumental music, but not quite. I had to build a lot of tooling behind the scenes to help me manually curate a lot of junk quickly, and what's left is still, admittedly, not good enough.

3. It's hard to create full tracks. In theory, Suno generates up to 2-minutes initially, with 1-minute extensions, but this is super hit-and-miss in practice (even on recently released v3.5 beta). Tracks frequently have large gaps in the middle, where the song ends and a new one starts up that ignores your prompt (feels like hallucination in text gen), amongst other issues. I generated waveforms for all tracks so that I could visualize this, and used ffmpeg to automatically detect and throw out the worst offending tracks.

4. Quality degrades significantly over time. The music generally gets more tinny and washed-out as the track goes on. I originally wanted long tracks (10+ minutes), but gave up on that as it's almost impossible to get good tracks that long. It's even hard to get a 2-3 minute track that doesn't degrade. I tried to curate this out, but you'll definitely hear this in the app. The AI just isn't good enough yet...

Overall, I have to admit that the end-result isn't up to the standard I was hoping for here. I've scratched my itch, but if there is sufficient interest, I'll pull up my sleeves and work to improve the music. To take it to the next level, I'd hire experienced music producers, and build more tooling to help them create and curate great music for focus across interesting genres. I don't think it can be done well without skilled humans in the loop.

Quick note on business model: This is super cheap to run, and I own the rights to all the music, so I intend to just keep it out there for free. If, by some fluke, it gets some interest, I'm keen to take this further with a lot more, and better, music. I'd fund that work by introducing a freemium model, with some new channels/features behind a reasonable subscription. I had success with a very generous forever-free offering at my previous startup (Mealime), and I'm allergic to ads.

So... please share your thoughts, and especially critique! What would make it better for you? How do you feel about the music being AI-generated?

Links: - iPhone app (best experience): https://apps.apple.com/us/app/flowtunes-music-for-focus/id65... - Web app: https://www.flowtunes.app