An analysis of the influence of acoustical adverse conditions on speaker gender identification

Speaker gender as a biometric feature plays an important role in numerous voice-based services. In this work we perform an accuracy analysis of a gender recognition system in different acoustical environments (indoor and outdoor auditory scenes). At the evaluation stage, each sentence has been mixed with several types of background noise using various signal-to-noise ratio levels. Then a voiced parts of speech have been extracted and parametrized using features based on filter banks and vocal-tract properties. The obtained feature trajectories have been non-linearly smoothed in order to minimize the influence of adverse conditions on the spoken sentences. The observed accuracy is acceptable for voice-based tasks where the gender information can improve their performance.