If I had 1000 audio files where three people are independently saying an animal at the same time, there can be 9 independent labels of animals. What features should I select from the audio file, and how should I normalize them to build the most accurate SVM to classify the audio files? For reference, I am doing this project in Python,so specific code examples and libraries would be very helpful. Thanks! -Joe
1
-
$\begingroup$ Why would you use an SVM for this (in 2023)? Sounds a bit like a school project. $\endgroup$– hendrikCommented Dec 4, 2023 at 6:35
Add a comment
|
1 Answer
$\begingroup$
$\endgroup$
there can be 9 independent labels of animals.
One assumes the three speakers are making uniform draws
from {opossum, anteater, roadrunner} or some similar
urn.
In this case counting syllables (3) would be uninformative,
but for a different vocabulary we may see different results.
Simple techniques can isolate those three animals from one another.
For a less constrained input vocabulary, 50 ms or 20 ms MFCC windows would be a pretty typical feature to train on.
Find a pre-trained model to tweak, or train from scratch.
