I don’t know (a quick glance at the documentation doesn’t make it clear) if the codec does this automatically, but the step which immediately occurs to me is to cut the sample rate. Telephone audio, using either A-law or μ-law companded PCM, which at 8 bits per sample is far from being an efficient codec (but very robust & easy to implement), renders speech quite well at a sampling rate of 8 kHz. The fewer samples you have to encode per second, the more bits per sample you can allocate at any rate of bits per second.