It varies, but usually I compress my diaphragm to force air from my lungs past a set of vibrating membranes which produce a base tone, then modulate that tone using the various muscles in my mouth to produce a set of distinct acoustic patterns which my conversational partner can decode. If that's working well I'll also modulate the frequency and volume of the tone itself to encode additional semantic information, but that's not strictly necessary for communication to take place.
