For speech (with a single, centred narrator, and no sound effects etc), I use lame -V 9 -mm --resample 22.05

That's lowest-rate VBR (implies -q 2 - slow, high-quality encoding), merging left+right to mono, and half the normal CD sample rate (implies low-pass filter at about 10kHz).

That gets it down to about 30-40 kb/s and I can't hear any ill effects (in the car, at least).

As your ears are different to mine, I suggest you encode one track at three or four different settings, using what you read in this thread as a base, and compare them to decide what's acceptable to you.