DLE code needs to be doubled in the audio stream.

[CABrouwers]

I believe the DLE (0x10) character in the voice data stream is treated as an escape character before an in-band command. That means that if it appears in the data it might trigger a spurious action by the modem. The solution is to double every occurrence by adding the following line before the write statement:
data = data.replace(b'\x10', b'\x10\x10')
That means that the data chunk might become bigger than 1024 bytes. So I recommend also decreasing the chunk size to 960

reference: https://en.wikipedia.org/wiki/Voice_modem_command_set