Most automatic detections of highly stereotyped baleen-whale calls such as the Antarctic blue whale calls (known as Z-calls) are usually performed by matching spectrogram-based templates. However, such approach presents weaknesses, such as its dependence on spectrogram parameters, as well as on a-priori subjective choices of a detection template and thresholds. Moreover, interferences (i.e. transient signals that are not a Z-call) in the considered frequency range (18-28 Hz), a main issue for all detection methods, strongly impact the detection performances. Here we propose a detector overcoming these limitations. Our algorithm is based on a subspace detector of the Antarctic blue whale call. Indeed, any Z-call can be modeled as a sigmoidal-frequency signal with unknown time-varying amplitude. The subspace detection strategy takes into account the presence of interferences, as well as intra and inter-annual frequency variations in the blue whale calls, which are commonly observed. Furthermore, the detection threshold adapts automatically to the ambient noise, a major improvement relative to predefined and fixed threshold of spectrogram correlation detectors. Thanks to subspace detection theory the detection performances are analytically known, which is crucial for detailed analyses such as density estimation. As a byproduct, the detector also provides other useful information such as the SNR of each detected Z-call and an estimation of the ambient noise. The proposed method has been applied on marine data recorded in the southern Indian Ocean, and the detector performances are compared with a classical spectrogram correlator (XBAT). To perform this comparison, three different contexts commonly encountered in acoustic databases have been defined, and our dataset has been divided into three subsets by an experienced human operator (EHO): 1) data containing only Z-calls at various SNR, 2) data containing Z-calls and various types of interferences, 3) and data with interferences only. In total, more than 105 hours of acoustic data were analyzed, displaying more than 2200 Z-calls (manually identified by an EHO) and different types of interferences (airguns, earthquakes, sounds with continuous components, ice-rifting sounds, fin whale calls, Australian pygmy blue whale calls, and an unidentified recurrent transient sound). The proposed algorithm has a true detection rate up to 15-20% better than a spectrogram-based correlation detector such as XBAT.