This paper presents an improved affine projection algorithm based crosstalk resistant adaptive noise canceller (CRANC). The proposed CRANC consists of three adaptive filters, one of which is used for detecting speech activity and the others are employed to estimate the interference and crosstalk transfer functions, respectively. An efficient sequential speech detector is also proposed in this paper, which is completely built into the adaptive strategy and thus causes no delay in the input and output of the noise canceller. The proposed CRANC was tested and compared with the conventional CRANC in the white Gaussian noise environment. Experimental results show that the proposed CRANC achieved higher signal-to-noise ratio (SNR) improvements than the conventional CRANC.