NeuralMorse – reinventing Morse code with neural networks

"A typical elementary student in counties where Chinese characters are used (mainly China and Japan) learns 1,000 to 2,000 characters by the time they reach the sixth grade."

I'm guessing that the author is trying to say that it is easy enough for a 12 year old to accomplish.

But the children don't learn those 1000-2000 characters when they are 12. If they started learning characters from 5 years old (kindergarten), then it takes a few hours a day over 7 years, during the time of life one is typically a language sponge, to pick up NeuralMorse.

Also, if I were to do this, I would make the distance between each of the notes a different number of steps. This would help facilitate remembering words based on their "song" while making it impossible to accidentally transpose. For example, abab should not be the same intervals as cdcd.

The author mentioned that the number of notes was constrained because too many notes would require perfect pitch. I'd say the same goes when there is not enough variation of intervals. In the case the shared ab and cd interval is missing out on an opportunity for clarity.

When all the intervals are different, perfect pitch isn't necessary to determine whether an incoming message is starting on ab or dc. You just need to hear any other note to figure it out.

A useful side effect is that the actual starting pitch is then made irrelevant, making it less prone to doppler effects... just keep the intervals intact.