A speech noise whitening filter for system identification

An echo path modelling device is presented that utilises two parallel least mean square (LMS) digital adaptive filters (DAFs) Simulations show that statistical properties of speech cause LMS system identification filters to effectively ´stall´ at incorrect states invalidating speech circuit characterisations. The proposed configuration employs a noise whitening filter to periodically decorrelate segments of speech and as a result significantly improves the performance of the modelling filter. The primary application of the device is for in-service non-intrusive measurement devices (INMDs) on telephony channels that require a fast and accurate characterisation of the echo path to estimate call quality degradations.<>