Research on Speech Recognition Model Optimization and Real-Time Speech Interaction System Based on Deep Learning

Voice recognition has been widely used in intelligent human-computer interaction, especially in the areas of voice assistant, intelligent house and autonomous driving. Due to the rapid development of deep learning, the precision and applicability of speech recognition have been greatly improved. However, in the complicated noise environment, different accents and high speed of speech, there are some problems in traditional speech recognition systems, such as poor recognition precision and slow response. In order to solve these problems, a novel voice recognition model is proposed, which combines CNN with self-attention. The model uses CNN to extract the local characteristics of voice signals, and then makes use of Transformer's self-attention mechanism to model the long-term dependency of voice data, so as to improve the precision and robustness of voice recognition. Experiments show that the proposed model achieves over 95% recognition accuracy over HMM-based speech data sets, which is 15% higher than the conventional one. Moreover, the proposed model has a good performance in real time speech interaction systems, and it can react rapidly and achieve high efficiency in speech recognition. Compared with the traditional methods, the proposed system is more adaptable in complicated environment, and it can deal with various kinds of background noise, speed difference and accent variation.