We consider multiuser interference suppression for direct-sequence spread-spectrum multiple-access (DS/CDMA) system without knowledge of other user's spreading waveforms using the minimum mean squared error (MMSE) performance criterion. The conventional MMSE detectors require training sequences for adaptive implementation. The previously proposed minimum output energy (MOE) detector using canonical representation requires no more knowledge than does the conventional single-user receiver: no more desired user's signature waveform and its timing. But this MOE detectors's blind adaptive algorithm is so noisy as to perform worse than the conventional LMS algorithm with a training sequence. And the MOE detector's MMSE solution is greater than that of the conventional MMSE detectors. In this paper we pro-pose an linear adaptive MMSE detector using hidden training sequence to combine the merits of training-based MMSE detector and MOE detector. And the MMSE detector using new canonical representation which make the convergence speed faster than the proposed detector using hidden training sequence. The simulation results are given to verify that the performance of two new proposed detectors is better than that of the previous MMSE and MOE detector.