Gender Recognition Using Neural Networks and Asr Techniques

The paper presents the simple technique of speaker gender recognition that uses MFCC features typically applied in automatic speech recognition. Artificial neural network is used as a classifier. The speech signal is first divided into 20 ms frames. For each frame, Mel-Frequency Cepstral Coefficients are extracted and the created feature vector is provided into a neural network classifier, which individually classifies each frame as male or female sample. Finally, the whole utterance is classified by selecting the class, for which the sum of corresponding neural network outputs is greater. The advantage of the method is that it can be easily combined with speech recognition, because both processes (gender recognition and speech recognition) are based on the same features. This way, no additional logic and no extra computational power is needed to extract features necessary for gender recognition. The method was experimentally evaluated using speech samples in English and in Polish. The comparison with other methods described in literature based on other feature extraction methods shows the superiority of the proposed approach, especially in cases where the recognition is carried out in noisy environment or using poor audio equipment.