Cool idea, though I was surprised by the level of fidelity loss in the fountain example. I would’ve expected that to be a good case scenario for noise cancellation so maybe it just needs some more time to iterate and improve on its level of “false positive” removal.
“Comma-la” unfortunately doesn’t help much for people without US accents lol (though of course people in the US are who the question and answer are most relevant to). On first reading – without the accent or something close to it – it implies “kom-uh-luh”, whereas with the accent it implies something more like “kah-muh-luh”, just based on how people pronounce “comma” differently.