Technology

The central innovation of the ADnota product family is the voice search engine. It consists, much like other engines of its kind, of four major parts: feature extraction, end point detection, spectral distortion measurement and time alignment.

The feature extraction uses the classical algorithms. They have been adapted to the special needs of mobile devices.

adnota invested a large part of its development efforts in the search for optimal distortion measurement functions. Most of the popular functions have been evaluated against a huge, specially recorded keyword-database. The most promising functions and parameters were subsequently used as the starting point for an extensive optimisation process. Optimisation took place on a cluster of Linux PCs. The algorithm used was "evolutionary strategy" (Evolutionsstrategie), a method, developed during the sixties and seventies of the last century by I. Rechenberg at the Technical University of Berlin. It simulates natural evolution. Its outcome was a set of weights for the speech parameters and distortion functions, adequate for all the languages in our database.

The search results are ordered in a list, best match first. Refining of the search is done by sorting the lists resulting from different searches into a new list. The sorting algorithms, developed by adnota, bear some resemblance with fuzzy logic operations.