Skip to content

Commit

Permalink
grammar correction in time learn page
Browse files Browse the repository at this point in the history
  • Loading branch information
MattS6464 committed Aug 28, 2023
1 parent d0aba86 commit 41f251a
Showing 1 changed file with 1 addition and 1 deletion.
2 changes: 1 addition & 1 deletion src/routes/(content)/learn/time/+page.svx
Original file line number Diff line number Diff line change
Expand Up @@ -68,7 +68,7 @@ These simple models of sound are simple, but are valid representations of how di

This question emerges in several different places, especially in our own tutorials and ultimately we think it is important to recognise that machine listening is context specific. For example, the [classification with a neural network](/learn/classification-neural-network) tutorial deals with time in a straightforward and effective way. A representation of different timbres is constructed by capturing and storing several examples of very brief moments in time using the [MFCC descriptor](/reference/mfcc). This analysis procedure is designed with somewhat static sounds in mind, where the variation of those MFCC values will be relatively stable and won't evolve too rapidly. The neural network that learns from this data is also quite clever. In theory, with enough examples and sufficient training the network can be good at differentiating timbres with just these instantaneous snapshots of sound data provided to it. For example though, if you wanted to explicitly have the same neural network classify sounds based on the evolution of timbre, you would probably want to explore a different method of capturing the _temporal_ features of those sounds, rather than just looking at singular spectral frames of analysis.

As another example, the [2D Corpus Explorer tutorial](/learn/2d-corpus-explorer) operates on a set of basic assumptions about how time might be encoded into data. Like the classification tutorial, it uses MFCC data alongside [statistical derivatives](/reference/bufstats#derivatives) to try and encode the evolving spectral structures of short sound samples. This approach is actually quite prone to producing noisy and nonsensical data so we slap another algorithm on after that called [UMAP](/reference/umap), and hope that it can unfurl some structure from that mess of data its given. The idea is that after UMAP has analysed and compressed the data it was given, a new set of fewer numbers will emerge that have some kind of statistical relevance and were able to introspect on the time varying aspects of the sounds encoded in those statistical derivatives. This approach, and the classification tutorial's approach are of course entirely validated by their success in musical practice. They are by no means _perfect_ approaches to encoding time and we are not looking for exact answers here, rather, they are functional patches based on a context-specific musical application.
As another example, the [2D Corpus Explorer tutorial](/learn/2d-corpus-explorer) operates on a set of basic assumptions about how time might be encoded into data. Like the classification tutorial, it uses MFCC data alongside [statistical derivatives](/reference/bufstats#derivatives) to try and encode the evolving spectral structures of short sound samples. This approach is actually quite prone to producing noisy and nonsensical data so we slap another algorithm on after that called [UMAP](/reference/umap), and hope that it can unfurl some structure from that mess of data it's given. The idea is that after UMAP has analysed and compressed the data it was given, a new set of fewer numbers will emerge that have some kind of statistical relevance and were able to introspect on the time varying aspects of the sounds encoded in those statistical derivatives. This approach, and the classification tutorial's approach are of course entirely validated by their success in musical practice. They are by no means _perfect_ approaches to encoding time and we are not looking for exact answers here, rather, they are functional patches based on a context-specific musical application.

This is really just the start of thinking about _time_ and we hope that you delve into it more through a series of [Learn Articles](/learn) that we think might provoke you to consider how you represent time through data, and how the evolution of sound can be captured and encoded into workflows with the FluCoMa toolkit.

Expand Down

0 comments on commit 41f251a

Please sign in to comment.