Philipp Schwesig, "Development of a Realtime System for Gender Adaptive Formant Extraction", 2008.

The purpose of this diploma thesis is the extension of the existing formant extraction system to the parameter of the gender. The formants describing the resonance frequencies of the vocal tract are slightly different for male and female speakers. This was the essential point leading to the attempt of improving the robustness of the existing formant extraction system by the gender information. Previous to this diploma thesis it has been shown that the fundamental frequency provides reliable information about the speakers sex which can be used to adaptively control the formant extraction. The voicing probability is one of the supporting features of the gender computation and therefore an existing voicing algorithm has been integrated into the real time environment during this diploma thesis. The detection of the voicing works mainly on two cues. The first one is the energy distribution of the signal and the second one the harmonicity [11]. The main feature for calculating the gender probability is the mean fundamental frequency. The process of extending the formant extraction system is divided into the implementation, evaluation and optimization of the gender detection algorithm and its dependencies. This thesis firstly introduces the topic of speech detection and production, followed by an overview of the technologies used and concludes with the implementation and evaluation of the described algorithms. A performance test is done by integrating the components into a real time environment that enables a live online demonstration. The results revealed that, indeed, reliable gender detection based on the fundamental frequency and the voicing of the speaker is possible. An algorithm introduced by Mustafa and Bruce[19] has been used for the evaluation of our test results and it could be shown that our algorithm works significantly better on the sample level as well as on the mean of all samples of the signal. How much the robustness of the feature extraction has been increased by the gender detection is discussed at the end and shown by the results of the online demonstration system.