Re: OT programming challenge:

On Sat, 9 Mar 2024 18:14:32 -0500, Chris Ahlstrom wrote:

rbowman wrote this copyrighted missive and expects royalties:
 

https://arstechnica.com/information-technology/2024/03/matrix-

multiplication-breakthrough-could-lead-to-faster-more-efficient-ai-models/

>
Now here's a real challenge for the mathematicians in cola...  Can you
beat an ω of 2.371552?

 Yeah, I can beat matrix experts easily... not.

https://en.wikipedia.org/wiki/Convolutional_neural_network
Gems like that are what make minute improvements important. I have a general understanding of what goes on in a CNN. Lucky for code monkeys like me people who really know what they're doing package it up in Torch, TensorFlow, and so forth. Image processing has some uses that aren't intuitive. For wake words like 'alexa' you digitize the audio signal, run STFTs over a sliding windows, maybe throw in mel filters to emphasize human speech characteristics, and produce a spectrogram. Image classification is a well trodden path so you train your model with spectrograms of many different people saying 'alexa', boil it down until you can do inferences on the local device and you're good to go. Once the device wakes up speech processing gets shipped off to a server. One goal is to be able to do more locally and avoid the expense and security problems.
Fortunately you can change the wake word from 'alexa'. I asked one of our QA people who goes by 'Alex' and she said she changed it to 'echo'. The other choices are 'amazon' or 'computer'.
Again lucky for me there are nicely packaged libraries for all this good stuff too.
https://github.com/mborgerding/kissfft
KISS-FFT would be a sacrilege to PhysfitFreak.
"It is wonderful that free, highly optimized libraries like FFT_BRANDX exist. But such libraries carry a huge burden of complexity necessary to extract every last bit of performance.
Sometimes simpler is better, even if it's not better."
Microprocessors don't have a lot of SRAM/RAM and many don't have a FPU. If simple gets the job done processing audio signals it's definitely good enough.