MusicNN considerations for big libraries

[AverageHoarder]

Before I activate this feature for my library, I have a few questions:

1. Does it share the risk of slowing the web-ui down to a crawl like the tag-based recommendation engine? And if so, would the performance of Symfonium accessing lms via the api be negatively impacted or only the web-ui?
2. Would it make sense to move the docker container from my NAS cpu to my tower cpu and then move it back after the extraction? My tower has a 10G direct connection to the NAS.

[epoupon]

Hello!
By default, half the number of logical CPUs are used by the scanner. API operations are usually requested sequentially by clients (except for covers) so it should be usable/fine. Same for the UI.

I would say you can just try a scan. After a minute you will have an clear idea of your extraction throughput and if it is acceptable or not.

Edit: sorry misread your first question. Recommendations should be very fast, but I have not tested on a very large database. Feedbacks are welcome.
There is no impact on OpenSubsonic API except for the audioSimilarity endpoints and getSimilarSongs which is also plugged on it.
By the way, I also significantly optimized the tag based engine so it should scale better than before.
I will perform tests on a fake 1M db to see how it goes.

[simonbun]

Never mind, that would remove them from lms, which I don't want. There's no way to only exclude files from being analyzed with MusicNN, is there?

Not that I'm aware of.

[AverageHoarder]

I've gone over all "error" files. The preliminary results are:
69 mp3 files that play fine but have genuine errors (multiple streams, broken streams etc.)
9 flac files that are free of errors but low volume or weird audio (ambient/noise/whispering)
The rest are mostly silent tracks (minus a few broken files I have fixed or am trying to fix).

@epoupon are you interested in my "error" files as test data? Since it's mostly silence it's only 1 gig for 991 tracks.
I sorted them into:

broken_mp3s
false_positive
mostly_silent_or_short_tracks

folders. In the mostly_silent folder are a few subfolders I had to use to avoid name collisions.

[epoupon]

Sure!

[AverageHoarder]

Are recommendations good enough?

I've played around with the sonic analysis modes in Symfonium (Transition Maestro, Echo Match and Steady Vibes). The recommendations range from quite good to comically bad. Perhaps it's my collection (I have a ton of blues and jazz) but it often slips in a blues or jazz song even when coming from a pop or folk song. However it's great fun trying to figure out what on earth made the model put songs close together that don't appear to have much in common.

I also have never used a similar tool before so I don't have a point of reference. If I had an Nvidia GPU I'd probably try AudioMuse-AI on my tower and see if it produces similar results for given songs.

Here's how long the recommendations take to load after a scan:

2026-06-13 07:30:44.799859+00:002026-06-13T07:30:44.799Z 140567040473856 [info] [RECOMMENDATION] [audio-similarity] loading...
2026-06-13 07:31:00.463282+00:002026-06-13T07:31:00.463Z 140567040473856 [info] [RECOMMENDATION] [audio-similarity] computing reduced vectors... Reducing from 200 to 60 dimensions
2026-06-13 07:31:30.140621+00:002026-06-13T07:31:30.140Z 140567040473856 [info] [RECOMMENDATION] [audio-similarity] computed reduced vectors: 1055713 tracks, 88700 releases, 24949 artists
2026-06-13 07:31:30.140707+00:002026-06-13T07:31:30.140Z 140567040473856 [info] [RECOMMENDATION] [audio-similarity] computing track distance threshold using 500 samples...
2026-06-13 07:32:06.895267+00:002026-06-13T07:32:06.895Z 140567040473856 [info] [RECOMMENDATION] [audio-similarity] track distance threshold = 0.179912
2026-06-13 07:32:06.903641+00:002026-06-13T07:32:06.903Z 140567040473856 [info] [RECOMMENDATION] [audio-similarity] computing release distance threshold using 200 samples on 1000 candidates...
2026-06-13 07:32:07.392625+00:002026-06-13T07:32:07.392Z 140567040473856 [info] [RECOMMENDATION] [audio-similarity] release distance threshold = 0.335965
2026-06-13 07:32:07.394812+00:002026-06-13T07:32:07.394Z 140567040473856 [info] [RECOMMENDATION] [audio-similarity] computing artist distance threshold using 200 samples on 1000 candidates...
2026-06-13 07:32:13.576309+00:002026-06-13T07:32:13.576Z 140567040473856 [info] [RECOMMENDATION] [audio-similarity] artist distance threshold = 0.332505
2026-06-13 07:32:13.576420+00:002026-06-13T07:32:13.576Z 140567040473856 [info] [RECOMMENDATION] [audio-similarity] loading complete!

[epoupon]

Thanks for the report!
I don't know which endpoints symfonium triggers for these algorithms. The track path endpoint is really weaker than the similarity track endpoint, I know I have to work on it to make it better.
LMS also applies dynamic penalties in order to bring more diversity, to avoid consecutive same release or artist, to favor distance from previous track... That may explain why sometimes you could get off recommendations.
There is also the lms-recommendation tool to perform manual similarity or path request on tracks, it is hepful to spot oddities (but you have to pay for the 1m30s warmup for you big collection).