Conventional blind audio watermark (WM) decoders use matched-filtering techniques because of their simplicity. In these methods, WM decoding and WM detection are often considered as separate problems and the WM signal embedded by spreading a secret key through the spectrum of a host signal is extracted by maximizing correlation between the secret key and the received audio. Conventionally decoding is achieved by using a pre-defined decoding/detection threshold and tradeoff between the false rejection ratio and false acceptance ratio constitutes main drawback of the conventional decoders. Unlike the conventional methods, this paper introduces a pattern recognition (PR) framework to WM extraction and integrates WM decoding and detection problems into a unique classification problem that eliminates thresholding. The proposed method models statistics of watermarked and original audio signals by a Gaussian mixture model (GMM) with K components. Learning of the embedded WM data is achieved in a principal component analysis (PCA) transformed wavelet space and a maximum likelihood (ML) classifier is designed for WM decoding. Robustness of the proposed method is evaluated under compression, additive noise and Stirmark benchmark attacks. It is shown that both WM decoding and detection performances of the introduced decoder outperform the conventional decoders. (C) 2007 Elsevier GmbH. All rights reserved.