To avoid the impact of noise factors in speech recognition and improve the effectiveness of speech recognition, this paper proposes an intelligent product human-computer interaction speech recognition method based on a novel feedforward network learning algorithm. According to the dependent variable transformation relationship of nonlinear functions, a feedforward neural network speech recognition model is constructed. This model utilises the dependent variable transformation mechanism of nonlinear functions to more flexibly simulate the complex mapping relationship between speech signals and recognition results. A semi-supervised loss function is introduced into the model training, and stochastic gradient descent is used for iterative optimisation to achieve human-computer interaction speech recognition. Experiments have proven that the speech recognition accuracy of the method in this paper remains above 90%, and the speech recognition delay remains below 1 second, indicating good recognition performance, reliability, and application performance.