On Friday, September 03, 2021, at 8:04 AM, magnuswootton81 wrote: > b) do a knn match to a database, if it isnt near enough to any of them, you > store it as a new class. (so its unsupervised learning, at least this part > of it.) Oh man... you just got me an idea how to do something similar. I choose a time resolution (to get segments) and a frequency resolution (to delibaretly ignore details). Then I go through the audio from left to right and build classes the same way you do, but even simpler, because all I use is a pixel-wise sum of differences.
Add to that detection of some common transformation operations (like in in my last recognizer - frequency shift/gain/time warp) and you follow up with a "detail" step where you discover the exact nature of the differences between similar segments. ------------------------------------------ Artificial General Intelligence List: AGI Permalink: https://agi.topicbox.com/groups/agi/T89dc799e249e2d0b-M5db7acd32c089ee95ac650b2 Delivery options: https://agi.topicbox.com/groups/agi/subscription
