Research On Speech Recognition Technology In Embedded Platform

As embedded platform devices have a lower processor speed and smallstorage capacity compared with PC, such as mobile phone, tablet, wear digtialproducts and auto electronic digital products, Auto Speech Recognition (ASR)system in the cloud is usually applied in embeded platform devices, includingGoogle cloud speech recognition, iFLYTEK speech recognition and so on. Whilethe performance of ASR in embedded platform devices is concerned with thenetwork status of devices and the computing speed of ASR servers. As a result,ASR in the cloud awalys has larger network delay and less correlati on of thespecfied application in embedded platform devices.Main work of this paper is to work out a suitable hybrid ASR(HASR)system to meet the needs of application for embedded platform devices, and toacquire a high accuracy, a high speed of ASR and a good performance of accentrecogniton. There are four key questions which are the processor speed, thecapacity of ASR vocabulary, the speed of ASR and the nonstandard accent ofspeakers, and there are three solutions of ASR which are the ASIC of ASR, theSDK of ASR and the clound ASR in the ASR of embedded platform devices.Taking these four questions and three solutions of ASR into accont, thereproposed a HASR solution which is based on a ASIC local ASR system andextended by a clound ASR system in this paper.Then doing some research on local ASR system based on LD3320chip andGoogle clound ASR respectively to see their performance in embedded platformdevices. As these results shows, the average recognition accuracy is87.2%inlocal ASR based on LD3320, but its capacity of ASR vocabulary is no larger than50. While in Google ASR, the recognition accuracy depends on the accent ofspeakers and the ASR vocabulary frequence, and the recogniton speed dependson the network status in embedded platform devices. The delay of recognitionranges form0.3s to3s in WIFI network connection and ranges form4s to14s in2G network of cell-phone. Considering with the performance of local ASR andclound ASR in embedded platform devices, there poposeds a method of“Dynamic Switch of Scenarios” which can enlarge the capacity of ASR vocabulary, and a method of “Homonymic Vocabulary Mapping” which can dowell on the question of speakers’ nonstandard accent. With these two methods,build a HASR system which is based on a local ASR of LD3320and extended byGoogle clound ASR in embedded platform devices.Finally, run the program on the intelligent security robot to test thereliability of the HASR system, and evaluate its performance of ASR. The testresults of HASR has proofed that the performance of HASR is better than ASRmerely based on LD3320chip and enquires a higher accuracy of95.8%comparedwith oriental87.2%.