Path: ...!2.eu.feeder.erje.net!feeder.erje.net!fu-berlin.de!uni-berlin.de!individual.net!not-for-mail
From: rbowman <bowman@montana.com>
Newsgroups: comp.os.linux.advocacy
Subject: Re: OT programming challenge:
Date: 10 Mar 2024 03:20:45 GMT
Lines: 52
Message-ID: <l54n8cF9619U1@mid.individual.net>
References: <65e9cad3$0$4689$882e4bbb@reader.netnews.com>
	<l4ufsbF9kdnU4@mid.individual.net> <usf2m9$1n0jh$2@dont-email.me>
	<l518d6FmepqU1@mid.individual.net> <usgfe4$20gfq$2@dont-email.me>
	<l5272pFqgibU12@mid.individual.net> <usi3d7$2dh6i$2@dont-email.me>
	<l53msoF4bvgU1@mid.individual.net> <usiqgo$2ijlr$1@dont-email.me>
Mime-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit
X-Trace: individual.net Iq0kX5sSJEAFX/F3JbNrmAr22Pu1YBRACnA6KWlDn/vX4yCjWm
Cancel-Lock: sha1:jrdivQckGTPcOZOLqaUI/oJMLUE= sha256:jv9Xcb41OUP7JbZvLKtMck5/jjaYx81bmwVubLA6j3Y=
User-Agent: Pan/0.149 (Bellevue; 4c157ba)
Bytes: 3165

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.