Filters with less than 100Hz bandwidth do improve the SNR within the critical band.
Filters with more than 100Hz bandwidth do not improve the SNR within the critical band.
And as I said, this is only true for signals that are only masked by pure noise (i.e. band-filtered white noise).
I've been a CW op for 32 years now. So maybe a newbie compared to some. But I'm going to repeat my opinions, backed up by my experience over the decades.
I've tried using narrow filters (meaning less than 300Hz). They induce a certain mental state where pretty soon every static crash sounds like a dit or a dah. This is not an immediate thing, but sets in after operating for a while.
Why does this happen? Hit a 100 Hz wide filter with a static crash that is 50 ms long. A static crash is by definition a wide bandwidth signal with a lot of energy at a broad range of frequencies. Thanks to the filter, only the portion of the static crash in that 100 Hz window gets through. If there's energy in that window it comes through, and when it comes through it's nearly indistinguishable from a real dit that's 50 ms long, because the window is so narrow.
You know how long a dit is at 24 WPM? 50 ms.
You know what the sideband bandwidth of a dit at 24 WPM is? For typical keying rise and fall times of commercial rigs (2.5 ms) the sidebands go out +/- 100 Hz for a an actual bandwidth of 200Hz; due to poor keying shape of typical commercial rigs the actual occupied bandwidth of a dit at 24 WPM will be more than 400 Hz usually. That means that a 100 Hz wide filter is actually throwing away the "sharpness" of the beginning and end of the dit - features that my ears and brain might otherwise use to distinguish a real dit from a static crash, but which are simply removed by using a ridiculously narrow (100 Hz) filter.
With a much wider filter, assuming no interfering carriers, the same static crash conditions do not sound like a dit. Because I hear the static crash frequencies outside the same frequency of the morse dit, and I immediately know it's a static crash, not a dit. I'm using the BESP (Brain Ear Signal Processor) in my head to make that determination. And the fact that energy outside a ridiculously narrow bandwidth is available to do that.
The psycho-acoustic effect of using a narrow bandwidth filter for extended listening periods is even more nefarious. Switching to a narrow filter for a little while, because of interfering carriers nearby, is one thing. But I find that listening to narrow bandwidths for extended periods of time induces a strange kind of audio psychosis, where the restricted bandwidth just drives me nuts.
Do I ever use the narrow bandpass filters? Yeah, but I do not prefer them. I abhor them. They are there to get rid of nearby interfering carriers that cannot be removed with a simple notch. (Incidentally, narrow bandwidth notches, I do not find nearly so objectionable, especially when used with a nice wide filter. Yeah, the keyclicks from the interference comes through, and it's annoying to hear how awful so many CW transmitters sound especially during a DX contest, but the long term acoustic psychosis in me is not triggered by them.)
IME some filters are more appropriate for narrow use than others. "Brick wall" filters are the worst because of poor time domain response. Gaussian filters (which have gently sloping sides) do a lot better. I have homebrewed Gaussian crystal filters and Gaussian audio filters and find them far preferable to brick wall filters. What is a shame, is that for decades now the ham radio community has treated filter shape factor as something to optimize in the direction of brick wall, and away from gently sloping sides. This is a huge disservice to folks who buy commercial rigs built for specsmanship rather than actual usability. If you can get a nice 500 Hz Gaussian "roofing filter" for you fancy new rig, you are in for a world of joy when it comes to CW operating. But AFAIK I can't buy them, so I build them.
I know, I'm bucking the trend. Everyone else says that you need narrower, narrower, narrower filters, and they need to be more like a brick wall. I'm saying the opposite.