Speech Recognition Method Based on GraphRAG

Speech recognition based on large neural network models requires vast computing resources and storage, and this requirement severely limits their direct application in resource-constrained edge devices.To considerably reduce the size and complexity of models while retaining the core performance of the original models, this study proposed a novel graph representation augmented with knowledge distillation (GraphRAG) speech recognition method based on adaptive multilevel distillation pruning (AMDP).First, graph neural networks were used to build and optimize knowledge graph embeddings to enhance the model's semantic understanding of speech signals.Second, an adaptive multilevel distillation mechanism was introduced to build a teacher-student architecture composed of multiple models of different sizes.The large bottom-layer model served as the main teacher, and the student models at each layer sequentially learned from the teacher model of the previous layer.Third, after each round of distillation, the student model was refined and pruned to further compress the model size while maintaining high recognition accuracy.Last, detailed experimental tests were conducted on multiple benchmark speech recognition data sets to verify the effectiveness of AMDP-GraphRAG technology.Results demonstrate that, (1) AMDP-GraphRAG consumes minimal computing time and memory while ensuring a low word error rate (WER).WER with and without a graph is 8% and 10%, respectively.The memory usage of the baseline is about 500 MB, whereas the memory usage of AMDP-GraphRAG is about 350 MB, which is the lowest among all the values produce by all the compared models.(2) The AMDP-GraphRAG model can maintain or improve recognition accuracy while reducing the number of parameters.The baseline has 120 million parameters, and AMDP has only 10 million parameters.Moreover, the baseline's WER is 3.5%, and that of AMDP-GraphRAG is 2.8%.(3) Atlas information can enhance the generalization ability of the models.Models with graph information can adapt quickly and show improved recognition accuracy.This study can considerably reduce the scale of models, decrease computational complexity and resource consumption, and serve as a reference for improving the application of speech recognition technology in edge devices.