Abstract: When the feedback path of the hearing aid is suddenly changed, the whistle easily exists. To solve this problem, one variable step normalized least mean square-notch filter (VSN-NF) algorithm is proposed. When the echo path is relatively stable, the echo signal is estimated by using the variable step normalized least mean square (VSNLMS) algorithm based on state classification and subtracted from the input signal. The normalized distance between short term average and long term average of the filter coefficients is used to classify the filter state, which is categorized into convergent state, transitional state and steady state. Different step sizes are employed in different state. When the hearing aid whistles, the VSNLMS algorithm is closed to stabilize the whistle frequency. Then, the notch filter for the whistle frequency is generated dynamically to restrain the whistle. Finally, when the whistle is suppressed, the VSNLMS algorithm is opened again. In addition, two notch filters for the different band are introduced to restrain the dual- band whistle for the echo path which is easy to whistle at the multiple frequencies. Compared with the other feedback cancellation algorithms for the hearing aid, results show that the comprehensive performance of VSN-NF is the best and the algorithm can quickly suppress the whistle. Furthermore, the quality of speech generated by the algorithm is higher and the operation time is shorter, which is suitable for the low-power consumption and small volume products such as hearing aids.