We describe a method and an implementation for producing a highly compressed representation of speech, in the order of 40 bps. This compression method uses a speech recognition engine to analyze the speech signal at the morphological level, i.e. the words. The words are then coded using a word-level text compression mechanism. After decompression, the speech message is recovered using Text-To-Speech. We report experimental results of our implementation. In particular, we observed that the human listeners were able to recover from errors introduced by the speech recognition engine, and that the human perceptual errors were highly dependent on the content of the messages, especially regarding the familiarity with the topic.

Appears in Proc. Symposium on Signal Processing for Communications, IEEE Globecom, San Francisco, CA, December 2003.

Copyright (c) 2003 by IEEE. All rights reserved.

Download paper.