In this paper, we introduce a new multibit watermarking method for speech signals in the joint time-frequency domain. The Wigner distribution for the speech signal is computed and the time-frequency cells in the Wigner domain that are robust to attacks are used for watermark embedding. The error introduced by the inversion of the watermarked distribution is quantized and used as a key for watermark extraction. The performance of the proposed watermarking algorithm under possible attacks is illustrated.