[HN Gopher] Automatically classifying the content of sound files...
___________________________________________________________________
Automatically classifying the content of sound files using ML
Author : msephton
Score : 43 points
Date : 2023-08-13 17:11 UTC (1 days ago)
(HTM) web link (blog.gingerbeardman.com)
(TXT) w3m dump (blog.gingerbeardman.com)
| mnky9800n wrote:
| This was a bit disappointing because I was hoping this would be
| an unsupervised learning example. This is not the author's fault.
| However there is a large need for unsupervised acoustic event
| detection and classification. Current methods like lta/sta, match
| filter, etc. Are not good enough anymore. I will be happy to see
| this kind of technology grow or be pointed in the right
| direction.
| jononor wrote:
| Can you provide some more details on your use case? In general
| a powerful approach is to use a neural network to extract
| vector embeddings and then do simple models on top. The neural
| network could for example be a pretrained audio classifier
| (like YAMNet mentioned in article), or a model trained with
| self-supervised learning on representative (unlabeled) data. A
| small model for unsupervised event detection / segmentation
| could be agglomerative clustering. A data-efficient supervised
| approach could be K Nearest Neighbors.
| mnky9800n wrote:
| As an example, I have 9 months of hydrophone cable data that
| has a large number of sounds that occur across a large
| variety of time lengths and frequencies. It would be nice to
| have a method that highlights certain signals given specific
| meta data. Tbh, what I think I want is a pipe dream.
| gopher_space wrote:
| I think some of the folks working with cetaceans might be
| able to point you towards interesting areas. I'm aware of
| work on processing petabytes of highest-res hydrophone data
| and I'm pretty sure they're tagging things that look like
| new clicktrains in some manner. Does this sound
| appropriate?
|
| As an aside, I've been kicking around the thought of what a
| software resonator might look like and where the analogy
| would work. I feel like there's a _really_ efficient filter
| you could pull out of the concept.
| jononor wrote:
| If you are willing to label 100 instances of each event
| type, then the approaches suggested above should work. And
| half of the data should be used to quantify how well it
| works.
| akomtu wrote:
| What kind of metadata?
| msephton wrote:
| Sorry, can you explain what you were hoping for? Do you want to
| be able to classify sounds in a live audio feed? That should be
| possible relatively easy, by recording a feed and then every so
| often cutting off the most recent chunk for processing. If I've
| misunderstood, please explain further.
| ladberg wrote:
| I think they were hoping that the off-the-shelf model the
| author used was a novel technique trained on unlabeled data,
| not a dataset labeled by humans.
___________________________________________________________________
(page generated 2023-08-14 23:01 UTC)