Hacker News Reader

How well it works for identifying music is going to be an issue of how well you can construct the feature vector which is not covered. If you map every song in to the same small subset of space, LSH will struggle. Map them to well separated places and it will shine.

View on HN

This is wrong, or misleading at best.

It's a known difficult open problem to create a good audio representation (feature vector). Good, in the case of LSH, would be that representation distance correlates well with human perception.

For something like Shazam, you want this representation to be invariant to minor transformations of the audio. That's the interesting problem.

How is that not aligned with what I’m saying? You add some detail, but that doesn’t make me wrong.

Reply on news.ycombinator.com