The integrating decoder currently works like this: It looks for dashes by integrating the signal multiplied by a dash-like folding function over a time interval that is slightly longer than one dash, thereby sensing the edges as well as the bulk of the dash. This procedure is repeated at a great number of points along the entire telegram, of the order every 10 ms. Since the length of an individual dash varies, this is then redone for a range of possible dash lengths.

The integrals with the highest scores point out the likely dashes. Once a dash has been located, many nearby candidates can be dropped as they would overlap.

The same procedure is then used to find the likely dots, using a shorter integration interval and a dot-shaped folding function.

Some dash and dot candidates will overlap, especially when a dash is a bit weak in the middle. A choice is then made, by comparing the strength of the dash candidate to the strengths of the two candidate dots. It turns out that the strength ratio limit is a quite sensitive parameter – either many “t” will come out as ”i”, or the other way around, if the limit value is not right. I tweaked the limit by testing against a set of SAQ recordings of varying quality.

I was happy to find that the integrating scheme seems to do better than the previous tone/silence edge-finding schemes. However, my impression is that the ear+brain is unbeatable when the signal to noise ratio is poor. And the brain does it with no floating-point arithmetic whatsoever!