Abstract: In the detection of transient evoked otoacoustic emissions (TEOAEs), the commonly used stimulus artifact reduction scheme is Derived NonLinear Response (DNLR). But obvious stimulus artifact will be left if only DNLR is used because the assumptions DNLR based on are not precisely met. The residual stimulus artifact inhibits the examination of high frequency cochlear responses with short latencies. Time window and wavelet based methods were applied to suppress the large residual component of stimulus artifact left after DNLR. Whereas these two methods are still not able to solve the difficulty of high frequency TEOAEs extraction, although they improve artifact reduction. To solve these problems, we propose an artifact reduction technique bidirectional filtering that avoids the leakage of artifact to the subsequent signals while filtering. Physical simulation and 106 clinical TEOAEs detection tests with bidirectional filtering showed that the proposed method is much more effective in reducing stimulus artifact than time window and wavelet based methods, as it not only expands the time region of TEOAEs uncontaminated by stimulus artifact but also detects out more TEOAEs with higher frequencies and shorter latencies. Moreover, it can be easily implemented in a real-time TEOAEs detection system.