Forensic Automatic Speaker Recognition with Degraded and Enhanced Speech

In an exploratory study a set of signal degradations typically found in forensic speaker recognition tasks were investigated for their effects on the performance of an automatic forensic speaker recognition (SR) system. Enhancement algorithms tailored to the types and degrees of the degradations were applied. Comparing equal-error rates (EERs) for voice comparisons with ten male speakers based on the original, degraded, and enhanced voice signals the performance of the SR system was most affected by pop music in both single-channel and 2-channel recordings and noise inside a fast moving car-and the effects of exactly these types of additive noise could be reduced most by speech enhancement. Road traffic and restaurant noise did not affect the system's performance significantly (EERs <2%) and could not be reduced further. The same applies to severe amplitude clipping and room reverberation, which produced equal error rates between zero and half a percent.