The long short-term memory based on i-vector extraction for conversational speech gender identification approach

Stress causes a speaker’s voice characteristics to be changed. Emotional stress alters a person’s speech pattern such that it is distributed non-normally along the temporal dimension. Thus, the methods for identifying the gender of a non-stressed speaker were no longer effective in recognizing the gender of a speaker in stressful conditions. To address this issue, a new gender identification framework is proposed. We leveraged i-vector for capturing gender information on each speech segment. Then the long short-term memory dynamically handled all speech temporal context features and learned the long-term dependency from the input. We evaluated the effectiveness, in terms of accuracy and the number of iterations to saturate, of the proposed method by comparing it with the baseline methods in their respective abilities to identify the speaker’s gender from conversations with different durations. By learning the gender information encoded in long-term dependencies, our proposed method outperforms the baseline methods and is able to correctly identify the speaker’s gender in all conversation types.