Thanks for responding quickly. Even if there is no known practical attack on longer stretches of normal speech, I wouldn’t be surprised if these are being actively worked on or already exist without us knowing, especially considering that Wire and other encrypted voice apps are getting more popular.
Now that I think about it, having the noise of the variation with network strength mixed with the variation in bitrate in response to the input signal could make an attack more difficult. But if the network is stable enough, that may not matter.
Perhaps an approach using CBR for encoding the sound and dynamically adjusting the constant bitrate with the network strength is worth considering.