Simultaneous recordings of signals from multiple neurons with multi-channel extracellular electrodes are widely used for studying information processing in neural circuits. Extracellular recording data contains the spike events of a number of adjacent or distant neurons, and spike trains of individual neurons should be estimated using some spike sorting method. However, spike sorting task of actual recording data still have some challenges even though a variety of methods have been proposed. First, high-count channel probes have been used for recording simultaneous activities from vast area of cortical laminar structures, and it becomes increasingly important to develop a method which consider both a synchronized spikes in different layers and a signal which is detected in multiple channels. Second, robustness to noisy and/or non-stational data is also becomming important for prolonged recordings with high-count channel probs. In this work, we introduce a new spike sorting method for high-count channel probes with an improved signal detection and variational Bayes clustering for normal inverse Gaussian mixture models. The proposed method significantly improves the cost-performance by automatic consolidation of signals in multiple channels and accurate clustering with skew distribution models. In fact, the number of cleanly sorted neurons and spikes significantly increases.