In this paper, we describe a phonetic vocoder based on syllable-units which represents speech waves by extremely low rate (100 bits/s) using a speech recognition tequnique. We take syllables into consideration as the unit of recognition / synthesis. Speech waves are transformed into a sequence of frames, each of which consists of LPC cepstrum, PARCOR coefficients, pitch and power. After the 0(n)DP matching with reference patterns, the input speech is transformed into a sequence of Japanese syllables. The information of recognized syllable contains the category of syllables, duration, power and pitch, and is represented by 16 bits. Using this vocoder, speech can be represented by only 100 bits/sec.