Methods for training and recognizing vietnamese speech for named entities

Abstract

This paper proposes a method for training and recognizing Vietnamese speech for raised entities, including the following steps: Step 1: Collect large datasets; Step 2: Filter the speech data collected from Step 1; filter from large speech datasets the speech segments and corresponding texts with high quality and accuracy; Step 3: Filter the text data collected from Step 1; Step 4: Generate aggregated speech data from the list of standardized texts containing the entities obtained from Step 3; Step 5: Enhance the data after the synthesis is complete and a list of synthesized audio files from Step 4 is available; Step 6: Build the training dataset based on the following two datasets: the synthesized speech dataset obtained in Step 4 and the actual speech dataset obtained in Step 2; Step 7: Train the speech recognition model on the combined dataset of synthesized data and actual data obtained from Step 6. Specifically, the method described in the invention creates a model after training with data according to the invention's method, consisting of synthesized speech data combined with real speech data, increasing accuracy by approximately 33% on the proper name test dataset and by approximately 40% on the address test dataset with the model trained with real speech data.