Optimal selection of bitstream features for compressed-domain automatic speaker recognition

Low-complexity compressed-domain automatic speaker recognition algorithms are directly applied to the coded speech bitstream to avoid the computational burden of decoding the parameters and resynthesizing the speech waveform. The objective of this paper is to further reduce the complexity of this approach by determining the smallest set of bitstream features that has the maximum effectiveness on recognition accuracy. For this purpose, recognition accuracy is evaluated with various sets of medium-term statistical features extracted from GSM AMR compressed speech coded at 12.2 kb/s. Over a database of 14 speakers the results show that, using 20 seconds of active speech, a recognition ratio of 100% can be achieved with only nine of the 18 statistical features under analysis. This is a complexity reduction by a factor of two with respect to previous works. Moreover, the robustness of the proposed system has been assessed using test samples of different length and varying levels of frame losses, and proved to be the same of previous approaches.