Just another update: junk filter auto-pilot is dead 
TL;DR: The relatively high computational cost of the junk filters, plus the fact that the algorithm can’t see track boundaries or album metadata through Roon, have killed the junk filter auto-pilot concept. The recommendation system, however, is still workable.
Here’s the version for those who like to read…
One of the founding principles of the auto-pilot feature was “never do harm”. In other words, the algorithm should never alter the listener’s experience for the worse (for more than a second or two, anyway… hey, that’s still faster than you are!). It turns out, the concept itself violates that principle for two reasons.
- These filters are not free to run in terms of processing power. They could easily push a machine on the edge into the red, leading to stuttering via buffer underruns, with no action on your part… just sitting, enjoying the music, and then it stops, restarts, stops, etc.
- Content is too widely variable to guarantee that no harm would be done. I resisted this idea for a long time, telling myself “if my brain can do it, an algorithm can do it.” In this case, that’s false for a few reasons.
Here’s the recording that landed that last idea for me (wonderful performances, by the way)…
Here are a couple of spectrogram shots:
Whatever tf is going on here, the background alternates between flat-ish and having a hump centered around 50 kHz plus a spur around ~76-78 kHz.
No auto-piloted algorithm could work with this without violating “never do harm”. In fact, the only reason my brain can figure this out is because it has access to a fact that the algorithm doesn’t: this content all comes from a single recording session.
The core problem here is that Roon exposes f@#$all metadata to HQPlayer… no track boundaries, album identifiers, etc. In other words, to the algorithm, the next frame may well be different content altogether. It has no way of knowing until the source format changes. (Even if we had this metadata, e.g. those using HQPlayer Client rather than Roon, algorithms “smart” enough to figure this one out are well outside of what I’d scoped for this concept.)
So we end up in a situation in which the algorithm, going frame-by-frame, is constantly switching on and off to treat that spur. It can’t see itself going back and forth ten times per track because it has no clue what a track is. It can’t see the damage that it’s doing via inconsistency alone.
All that being said, the recommendation system is salvageable because it puts a human brain capable of seeing the whole picture between the algorithm and the changing of the setting.
Recall that one of HQPTuner’s goals is to let the interface do the teaching. I don’t expect users to arrive with an innate understand the junk filters or the hi-res issues they treat, nor do I expect them to trawl the internet researching examples (or worse, asking AI for advice… I have tested this, it is really bad). HQPTuner should then, at the very least, guide users towards the right settings, and explain why when it can be certain so that users can learn to make educated choices when it can’t.
Whew, that was a lot, geez. Thanks for coming to my TED talk, am I right?