My Whisper AI trial vs. Oticon More

Volusiano · November 28, 2021, 11:02pm

happymach:

Volusiano:

Whisper decides to skin the cat the hard way, while the others used an easier way to skin the cat. I gave several examples, but I guess you don’t want to hear it or understand it. The other spent their resources and development cycle to train their DNN up front then condensed it into a formula-like result then implement this result much more easily into the HAs. Whisper meanwhile chooses to do it the hard way and that’s why they need the kind of power and resources that require a brain to implement and execute.

I don’t think this analysis of the differences between the two systems is correct. From my own reading of the Whisper documentation, I’ve taken away that both brands have DNNs that are trained “up front.” What is different from about Whisper’s design is that the resulting signal-processing model is updated with every firmware cycle (every 3 months), perhaps on the basis of user-recorded data, whereas Oticon’s is more or less set in silicon and unchangeable. In addition, Whisper’s signal processing model itself appears to involve more complex (i.e more CPU intensive) real-time treatment of sound, involving more “points” of fit.

I’ve already given a detailed explanation of my interpretation of how Whisper implements their DNN vs how Oticon implements their DNN. I don’t care to repeat it here, but I’ll give a link below so anyone interested can click on it to read up on it again. I’m not going to argue the details again here, we can agree to disagree on whether my analysis is correct or not. Other readers will be the judge of what they read.

But I’ll make a few observations below:

It’s common sense to deduce that a young, small start-up like Whisper probably doesn’t have the time and resources to fully and thoroughly train their DNN with enough data using a development platform offline that can be as big as to anyone’s heart content (a la Oticon More 12 million sound scenes worth of training that probably took years to collect and process). So they take a different approach and make the brain their “online” development platform (vs the Oticon offline development platform), invest to build a library worth of a few thousands hours of unique sounds (based on what I read from their whitepaper), but not collect tens of millions of sound scenes up front like Oticon did, then continue to collect data from users via the brain storage to continue to train their DNN.

So while Oticon did all the HUGE amount of training up front so they can afford to have a finalized version set in silicon, Whisper doesn’t have the resources to take this approach, so they choose a “continuous” approach of starting out with less training data to begin with, but keep on collecting more and more users data over time to perfect their DNN. So Oticon more or less have their DNN learn as much up front then “graduate” into the silicon, while Whisper has no resources to do this so they develop a system of “learning as they go along” then eventually graduate after so many years later with eventually enough users data collected.

So while Whisper spins all this stuff about the brain being a positive thing of being able to make major breakthrough improvement updates every 3 months and continually being able to collect real life user data; to me it seems more like it’s a beta version that is released prematurely and therefore needing update every 3 months to fix bugs and introduce more lacking basic features, instead of introducing major breakthrough every few months. And the user data collection bit seems more to me like a DNN not fully and thoroughly trained enough up front, so the only way to make up for their lack of resources to train and graduate up front is to continually collect data and train as you go.