Vocoder
[see page 14, Voice Coder] - Pass input speech through a filter bank to get a series of bands, then re-synthesize the signal after transmission by passing back through the filter bank.
Analyse in certain frequency bands and only transmit the output of those bands. Basically group frequencies together (or calculate them) into bands and then store these discrete groups. You need only as many bits as you have bands.
This approach is decent at storing and transmitting low-bit speech, but it isn't good at distinguishing whose speaking. You can make out the words, but not the voice of the person behind them.