Addressing the issues of poor real-time performance and insufficient predictive capability in continuous gesture recognition tasks based on surface electromyography (sEMG), a method utilizing Gaussian Mixture Model-Hidden Markov Models (GMM-HMMs) and Viterbi traceback is proposed for continuous gesture recognition. This approach leverages Gaussian Mixture Models-Hidden Markov Models (GMM-HMMs) to classify hand gestures into idle, ascending, steady, and descending states. A refined Viterbi sliding window marginalization technique is implemented to ensure prolonged connections between adjacent windows, enabling anticipatory prediction of subsequent gesture states. Moreover, a dynamic threshold model based on maximum likelihood is incorporated to accurately categorize gestures. In a study where 8 participants performed 12 continuous dual-gesture tasks, the method proposed attained an average recognition rate of 98.1% with a prediction time of 71ms, surpassing the LSTM model (94.2%, 309ms) and the GRU model (93.8%, 300ms), resulting in a considerable decrease in prediction time.