Efficacy of Artificial Intelligence for gender classification in speech signals

The classification or recognition of human voices are required for different applications such as speech emotion recognition, medicals, communications software, and security. However, gender classification of speech signals is complex aspect of speech recognition systems; thereby requiring robust signal processing strategies. This paper investigates the efficacy of Artificial Intelligence (AI) for gender classification in speech signals. Two different AI-based gender voice classifiers namely K-Nearest Neighbor (K-NN) and Long Short-Term Memory (LSTM) were developed. First, speech signals were recorded from different male and female speakers at a sampling rate of 48 kHz. Each of the raw speech signals was filtered and the useful portion of the signal was segmented. The Mel Frequency Cepstral Coefficient (MFCC) and Mel Spectrum (MS) features were extracted from each signal via framing, hamming window, and FFT. An equal number of observations each for the male and female classes were generated. The total 2000 observations were partitioned into 80% for training and 20% for testing. The training dataset was used to train both the K-NN and LSTM classifiers.  The results obtained from testing with the testing dataset showed that the K-NN classifier gave precision, recall, accuracy and F1-score values of 0.9852, 0.9850, 0.9925 and 0.9851, respectively, whereas the LSTM classifier gave 0.9132, 0.9091, 0.9525, and 0.9091, respectively. The classifiers achieve more than 0.95 (or 95%) classification accuracy; thereby demonstrating the efficacy of the AI strategies in distinguishing between a male voice and a female voice.