A system is described which performs time-alignment of continuous speech with phonetic transcription. The approach combines several techniques popular in A.S.R. (Dynamic Programming, Clustering) together with the explicit use of speech specific knowledge. The system is speaker independent, fully automatic and is able to cope with phonological variations like elision or assimilation of phonemes and insertion of pause or noise-like segments. It has been tested on several speakers and has proven to be well suited for the direct estimation of parameters required by a statistically-based recognition algorithm, working on a speaker-dependent mode.