Speech segmentation and recognition using syntactic methods on the direct signal

The acoustical parameters are progressively extracted from the sampled signal up to the phonems by use of syntactic methods. The rewriting rules, which are context-sensitive, are controled by the parameters that are evaluated on the signal itself. This lends the syntactic analysis to a main semantic point. These techniques are applied in time domain to the speech signal. Some preprocessing is done in order to code the elapsed time between the extrema of the signal and the extremum aplitudes. Particularly the voiced aeras are analysed pitch synchronous by means of a "formants extractor" which makes use of the extrema. The prosody of the signal is computed. The phonetic recognition results are given for various classes of french phonems. The are archieved on a mini-computer operating in real time.