CR (cognitive radio) technology has become an attractive field of research owing to the increased demand for spectrum resources. One of the duties of this technology is spectrum sensing which involves the opportunistic identification of vacant frequency bands for occupation by unlicensed users. Various traditional and state of art Machine-Learning algorithms have been proposed for sensing these vacant frequency bands. However, the common drawbacks of the proposed traditional techniques are degraded performance at low signal-to-noise ratios (SNR) as well as the requirement for prior information about the licensed user signal characteristics. More so, several Machine-Learning / Deep Learning techniques depend on simulated, supervised, and static (batch) spectrum datasets with synthesized features, which is not the case with real-world networks. Hence, this study aims to optimize real-time and dynamic spectrum sensing in wireless networks by establishing and evaluating a K-means-LSTM novice model (artifact) that is robust to low SNR and doesn’t require a supervised spectrum dataset. Firstly, the unsupervised spectrum dataset was collected by an RTL-SDR dongle and labelled by the K-means algorithm in MATLAB. The labelled spectrum dataset was utilized for training the LSTM algorithm. The resultant LSTM model’s performance was evaluated and compared to other commonly used spectrum detection models. Findings revealed that the proposed model established from the K-Means and LSTM algorithms yielded a Pd (detection probability) of 94%, Pfa (false-alarm probability) of 71%, and an accuracy of 97% at low SNR such as -20 dB, a performance which was superior to other models' performance. Using our proposed model, it is possible to optimize real-time spectrum sensing at low SNR without a prior supervised spectrum dataset.