A new analog-digital filtering method for real-time speech recognition

Using the low-pass filtering properties of a particular new type of analog-to-digital converter, we have developed a numerical filtering method which permits us to obtain, by programming, a large scale of passband filters. The procedure involved is very simple, fast, and consequently very well adapted to real-time treatment. It consists of sampling the analog signal at different rates and then linearly combining the samples thus obtained. These filters have a number of very interesting characteristics: for example, the frequency response curves resemble very closely the response of the basilar membrane. This combination of analog-digital coding and passband filtering seems especially well adapted to the treatment of the speech signal; we are presently applying the method to the problem of real-time speech recognition.