In the context of Cognitive Radio(CR), dynamic spectrum access has become a key approach to improve the spectrum utilization in wireless network. In this paper, using fine-grained spectrum measurement data collected from Global System for Mobile Communications-Railway(GSM-R), a data-driven deep-learning method is proposed to model the spectrum pattern and a framework is developed for dynamic spectrum access. A deep spectrum generative model is adopted to guide the spectrum allocation. A deep network that combines recurrent series model and background feature embeddings is developed to model and predict the short-term spectrum occupancy, then a strategy is proposed for dynamic channel access. Furthermore, a frequency-hopping system is implemented with Software Defined Radio(SDR) platform and it is integrated with the proposed strategy. The throughput capacity of this system is evaluated with real-world historical spectrum data. It is shown that the proposed method and system can enhance the ability of opportunistic communication and utilize the spectrum resource efficiently. The proposed spectrum access framework and the implementation with SDR are of great generality, so that they can be easily integrated into systems with different scenarios and frequency spans.