Synchronous-CDMA techniques possess intrinsic protection against co-channel interference due to orthogonal codes employed and thus allow multiple users to operate within the same frequency band. In the presence of multipath, however, each signal is subject to frequency-selective fading and the code orthogonality condition does not hold leading to increased cross correlation. In these scenarios, multiuser detection techniques need to be employed to suppress interference and recover the message symbols. To implement these algorithms, explicit knowledge of the (nonorthogonal) signature waveforms of all users is required. Here, we propose a blind estimation scheme that provides a closed-form solution to the estimation of signature waveforms by exploiting the structure information of the data output. In particular, we show that the subspace of the data matrix contains sufficient information of the signature waveforms. Based on this observation, a MUSIC-like [9] algorithm which can accomplish blind estimation using a minimum number of data samples is derived.