Passive Acoustic Monitoring (PAM) of marine mammal vocalizations has been intensively used in applications such as marine wildlife surveys. Marine mammals regularly produce sounds to echolocate, to communicate and while foraging and PAM is a complementary tool of visual-based observations for density estimation. Vocalizations of marine mammals are divided into two categories: impulse sounds and frequency modulation whistles. A common way to detect whistles consists in forming signal spectrogram and to check for frequency modulation tracks. Whistles tracks are made of energetic pixels identified by comparing each spectrogram pixel levels with an estimate of local noise. Despite simple to implement and quite effective, usual noise level estimation via signal low-pass filtering lead to false alarms in non-stationary noise conditions as in the vicinity of human activities or in costal observatories. A noise level estimation algorithm is proposed by taking into account the statistical properties of noise only samples of the spectrogram. The algorithm is based on minima statistics of a time-frequency neighborhood of the current spectrogram pixel where noise only pixels follow a $&_chi;^2$ distribution. The estimation is effectively applied on real data. Major drawbacks coming from low-pass filtering are avoided and the noise level is in closer agreement to the real noise level. Techniques are proposed to keep the real-time implementation of the algorithm feasible.