Voice Analysis and Noise Filtering


Voice comparison, voice analysis and filtering of background noise

Method description

As a consultant for several authorities (courts, public prosecutor's offices, police agencies for organized crime) in different states of Germany for many years, problems often occurred concerning the error-free evaluation of recorded speech signals. A cause for this problematic evaluation of speech transmissions was dominating background noise, sometimes covering the language completely. Such situations often occur during a speech transmission from a moving vehicle or if the transmission is overlaid with music and other sources of dynamic noise.
In addition the question was often asked, whether two human voices were identical or not.



This motivated me to develop two computer-aided methods and/or computer programs for an improved evaluation of speech transmissions and for voice analysis and voice comparison. One method makes it possible to filter background noises and thus resulting in much better understanding of speech signals. The other method allows proof of identity or difference of two voices. (voice analysis or voice comparison)



The basic principle for both methods is the theory that human speech can be interpreted as a one-dimensional structure, thus making it possible to apply new scientific findings and methods from real structural research, signal analysis and spectral analysis respectively. In both methods the speech signal or speech structure is interpreted as a real structure. In the frequency domain a real structure consists of periodic and diffuse parts. The disturbance through background noise in a speech recording is caused by an overlapping of noise components. Through several fourier deconvolution operations these noise components can be removed from the real structure (noisy speech), without distorting the speech itself.



The interpretation of recorded speech as a structure or signal also allows the structure of the human voice to be defined and characterized, for the digitized voice signal or the voice structure respectively contains all characteristics of the human voice. Through physical-mathematical methods, these characteristics can be extracted to calculate a voice profile (voiceprint). In this way, voice comparison can be accomplished. Prerequisite for this is a representative duration (approx. 60-80 seconds) of undisturbed speech recording.

In the following, the methods are briefly described:

- Noise filtering and reduction during speech transmissions (telephone, cellphone/mobile phone and other methods of transmissions of speech)

- Voice analysis (voice comparison report)

I hope I've succeeded in presenting you a short overview of possibilities in application and I am confident, that my services will be of interest to you. I'll gladly be available for further information or questions.

Prof. Dr. Sameh Rahman