Thanks, Sander. I'm on Win 7 -- but I suspect that is a clue. Perhaps Mathematica is just calling the OS speak function, so it never actually owns the waveform data.

By way of explanation, I was using Mathematica to build a test CD for my stereo. (A bit of fun.) For later processing, I have a track of combined discrete frequencies (chosen in the midpoints of a 6-way crossover ) and also one of white noise. For these I plan on playing the CD and capturing the output with a Yeti microphone, and then processing the result in Mathematica, for example with Periodogram. But for just listening use, I wanted to play the test tones one at a time, each preceded by a voice announcement giving its frequency. So I need to embed a sound primitive which is the voice announcement.

Best, David

On Mac OS X you can call the function say (which is what Mathematica does I guess) and give it some more arguments to save it to a file, in the terminal:

 say -o hi.aac 'Hello, World'