Very funny that it's become about microphones again. We did...

Very funny that it's become about microphones again. We did microphones to absolute death three years ago on the RAP forum. I believe it was @Southoz who identified this as a potentially serious problem. The company did address the issue but was somewhat dismissive, saying that although there were many different manufacturers they all used a very limited number of chipsets and components because the mobile form factor was very specific and demanding.

So suddenly it has now become an issue?! Perhaps a lot more component manufacturers have entered the market over the years.

But if the different microphones have different recording characteristics – distorting certain frequency ranges, chopping peaks, filtering non-vocal frequencies to highlight speech – then the signatures extracted from the raw sound could certainly be affected, which could in turn affect the predictions made by the algorithm. The outcome would be similar if the microphone driver code alters the raw samples before the application gets the signal. There are lots of variables at play before ResappDx gets its digital hands on the cough sound.

The amount of variability in play here would most certainly require a process for each sound configuration in the field:

1. Play the test data through the device's microphone and check the agreement levels of the algorithm's predictions.
2. If the agreement levels are outside an acceptable tolerance, retrain the algorithm using sounds recorded through the device microphone.

And of course if some hardware/driver combinations are outside the acceptable tolerance then that means the product isn't getting the same agreement as indicated in the certification submission. Bleagh!

To me that's a pretty compelling argument for supplying the product on a limited set of hardware with known characteristics. What luck there is such a thing!