Why were these specific frequencies chosen other than needing to fit into the voice band
Erm, in the end it's all about fitting in the voice band of 300..3400 Hz.
For one, 1 kHz and 2 kHz are almost equally spaced within the voice band, having the maximum distance between each and its boundaries, giving optimal placement and separation.
Next, 200 Hz between the signals each side can produce(*1) is small enough to not interfere (much) with the other side, while at the same time far enough apart to allow reliable discrimination with simple circuitry.
Last one (well, for here) is to keep harmonics down or far out. Far out is the best way, as filters tend to eliminate best signals that are way outside the passing window. I.e. as further out, as more energy is needed to get a signal thru. While this does not matter much on subscriber lines, it does on long distance (exchange to exchange) transmissions, as here single voice channels get multiplexed - saving a lot of copper. A technology already pioneered in the late 1930.
At that point we must take a step back and notice that the telephone network was made for speech and handling one channel. So there can not be specific filters for modem, but only for the whole channel. The analogue fall of of filters isn't much of an issue with regular audio. It becomes one with narrow band signals like modems create. Their frequencies are as well quite defined, thus harmonics carry more power, thus being harder to be cancelled out.
So, as so often in life it's about finding a middle ground among many parameters.
*1 - To be clear, both signals of one side are never produced at the same time. Quantum states are not an issue here so a bit is either 0 or 1, Space or Mark. Thus it's either (1070 Hz or 1270 Hz) and (2025 Hz or 2225 Hz). Thus no two concurrent signals, present at the same time will ever be less than 755 Hz apart.