you are viewing a single comment's thread
view the rest of the comments
[–] 2 points 1 month ago* (2 children)

The Xeon server would be a good bet. Your other machine would be potentially bottleneck for memory (though it meets min spec if the server isn’t doing anything else). There’s a NOAVX docker deployment available, would be slower but should work fine. Just be sure to disable anything associated with lyric detection, it’s an absolute performance nightmare.

I ran it on a Ryzen 5500u mini-PC with 32 GB RAM with the standard deployment with AVX2 support and scaled up to three worker threads. For a collection of 53k tracks it was processing about 100 per hour that way with lyrics/whisper translation enabled, but once I turned that off it was doing 1300-1400 tracks per hour.

——

Edit - the 6600T would work too. I found with lyrics disabled, each worker only used between 500MB and 2 GB of RAM. Long as the server isn’t under load while scanning I think that would work, and would be faster for having AVX2 support.

  • source
  • parent
  • hideshow 4 child comments
  • [–] 2 points 1 month ago*

    That's cool to know, however the Xeon runs FreeBSD, so I would need to create a VM if it doesn't work in Linuxulator, (FreeBSD's Linux compatibility layer, works sorta like wine does)

    I have 120k tracks. I like music.

    I should do some research this weekend I reckon.

    Great, now I have one more thing I gotta do this weekend. Thanks a lot. Lmao.

  • source
  • parent
  • [–] 1 point 3 weeks ago (1 child)

    So.

    I did a thing.

    I have audiomuse-ai running its main, complete docker compose script, with all containers, on my 8GB Raspi5, and worker-only containers running on:

    • An 8GB bhyve VM on my FreeBSD box
    • An E2-6110 AMD pre-ryzen APU with 16GB of ddr3
    • A Ryzen 5800x w 32GB RAM

    They've been running about a week, and I'm a little over a third of the way through

    Once the initial analysis is complete, I'll stop all worker containers and leave it all just running fully on the pi5.

    I also created a worker-only addon for the 6600T machine, but as it is already running HAOS and Jellyfin, I was getting a lot of OOM-related failures when it was running.

    But I also have 32G of used, eBay bought, ddr4 SODIMMs.coming for it.

    Bonus: Most of my homelab is in this. The only things missing are my Sophos running OPNsense, and the raspi5. Oh, and my actual desktop machine.

  • source
  • parent
  • hideshow 2 child comments
  • [–] 2 points 3 weeks ago* (1 child)

    Yo, that’s awesome!

    Pro tip for you, ASR (whisper - lyric detection/transcription) can be kind of bad, but if you have some spare resources, it takes very little to host a local LRCLIB database and clone lrclib.net (they have a GitHub page). This massively speed up lyric analysis for me using the API against a local site instead of getting 429s against lrclib.net or relying on ASR.

    Lyrics are the biggest longest part of the scans. My whole collection was like 3+ weeks with lyrics stuff on, but only 2 days with just MusiCNN and CLAP.

  • source
  • parent
  • hideshow 2 child comments