Support vector machine based gender identification using voiced speech frames
Manish Gupta, Shambhu Shankar Bharti, Suneeta Agarwal · 2016
Gender of the speaker can easily be identified by his or her voice. Identification of gender is required in restricted areas where only specific gender is allowed. Gender identification may also play important role in speaker recognition as this can be used to reduce search space. In this paper gender of the speaker has been identified using support vector machine (SVM). Two features-Pitch and Mel Frequency Cepstral Coefficients (MFCCs) are used as components of feature vector. Voiced frames are extracted from speech signal using Short Term Energy (STE) and Zero Crossing Rate (ZCR). Pitch value is calculated by cepstral analysis method applied on extracted voiced frames. MFCCs are also calculated from voiced frames. Experiments have been performed on the database obtained from IIIT-Hyderabad consisting of 5000 speech signals having voice of males and females in multilingual (Hindi, Telugu, Tamil, Marathi and Malayalam). For training the proposed model, 200 male and 200 female speech signals are taken arbitrarily from the above database. Evaluation of the proposed model has been performed using other 200 speech signals (males and females both) from same database. Without separating voiced and unvoiced frames, the accuracy of gender identification claimed in the literature is 99%. Using pitch and MFCCs that are calculated from extracted voiced frames, the accuracy of gender identification has gone up to 99.5%.