Most of the watermark (WM) decoding schemes use correlation-based methods because of their simplicity. In these methods, the WM signal embedded through a secret key is assumed as uncorrelated with the host signal. This is a hard restriction that can never be achieved and correlation between the received signal and the secret key becomes greater than zero even though the received signal is un-watermarked. Mostly a decision threshold specified semiautomatically is used at the decoding site. Since the audio watermarking is a nonlinear process that guarantees the inaudibility, there is no analytic way of determining an optimal threshold value that makes the WM decoding problem harder. This paper introduces a learning scheme followed by a nonlinear classification thus eliminates the threshold specification problem. The decoding process is modelled as a three-class classification problem and Support Vector Machines (SVMs) are used in the learning of the embedded data. The decoding and detection performances of the developed system are greater than 98% and 95%, respectively. When the Watermark-to-Signal-Ratio (WSR) is higher than -30dB, system false alarm ratios remain less than 2%. It is shown that the introduced WM decoding method is robust to additive noise and most of add/remove and filter attacks of Stirmark.