We have been interested in the analytical and experimental study of real-life bird song sources for several years. Bird sources are characterized by either a single or multiple bird vocalizations independent of each other or in response to others. The sources may be physically-stationary or exhibit movements and the signals are wide-band in frequency and often intermittent with pauses and possibly restarting with repeating previously used songs or with new songs. Thus, the detection, classification, and 2D or 3D localization of these birds pose challenging signal and array problems. Due to the fact that some birds can mimic other birds, time-domain waveform characterization may not be sufficient for determining the number of birds. Similarly, due to the intermittent nature of the vocalizations, data collected over a long period cannot be used naively. Thus, it is necessary to use short-time Fourier transform (STFT) to fully exploit the intricate natures of the time and frequency properties of these sources and displayed on a spectrogram. Various dominant spectral data over the relevant frames are used to form sample covariance matrices. Eigenvectors associated with the decompositions of these matrices for these spectral indices can be used to provide 2D/3D DOA estimations of the sources over different frames for intermittent sources. Proper cluttering of these data can be used to perform enhanced detection, classification, and localization of multiple bird sources. Two sets of collected bird data will be used to demonstrate these claims.