Voice of Fire is a sound sculpture with 13 flames that act as speakers. As people speak into the microphone, their voice comes out on the thirteen flames at different delays. As they are speaking, their voice is recorded, and when no one is speaking, the system plays back random bits of people speaking, different soundbites on different flames, gradually fading away to silence.
Tim Black and his team of mad scientists built the Voice of Fire for BurningMan 2003.

Here's how it works: ionize a propane flame so that it conducts electricity; place a 400 volt current across the flame; and modulate the current with an audio source. The current will cause the plasma gases in the flame to expand and contract at audio rate, moving air, and making sound.
More details here.

My contribution to the project was to write the audio software. I used Max/MSP running on a powerbook G4, with a MOTU 828 audio interface that would take input from the microphone, and produce 13 outputs, one for each flame. The core of the software is a sound snippet database system with a silence detector and random-interval recording system that would record soundbytes from 15 to 45 seconds long, and place them in a filesystem organized by time. A random playback system would select a file to play, and send it to its assigned flame. When a snippet finished, the system would delay a random amount, then play the next one. The random delays would increase if no one was speaking into the microphone, so that when someone spoke, there would be a burst of activity that would gradually fade to very sparse, only occasional sounds coming out. For each sound snippet, the burner would burst into flame (if it wasn't already going), wait a bit, then speak.