The Research And Application Of Keyword Spotting In Call Center Speech Quality Monitoring System

With the rapid development of the national economy and the increasingly frequent interaction between businesses and their customers, the call center as a communication bridge has become an indispensable part of business operations. Control or guarantee of speech service quality is the most important work in the call center management. At present speech service quality monitoring of call center is still in the inefficient manual monitoring mode. Single manual monitoring mode is not suitable for using in the current high growth of call center operations monitoring system. On the basis of the original evaluation metrics, this paper uses keywords spotting technology to achieve automatic detection of the machine. And based on detected speech keywords information, the paper designs a set of evaluation algorithm to evaluate speech quality ratings. According to the solution, the paper develops a speech quality monitoring system based keywords spotting.Firstly the paper introduces basic architecture and related technology module of the keywords spotting system. The keywords spotting system is mainly composed of model training and recognition. Its basic technology modules includes:speech preprocessing, feature extraction, acoustic model training, decoding and recognition. This paper describes the technical points involved in various technical modules, and highlights acoustic model training based on HMM. This paper utilizes the Sphinx system of SphinxTrain to complete CD clustering acoustic model training of AN4corpus.The design and implementation of keywords spotting engine is one of the main tasks of this paper. Based on the existing Sphinx-4existing recognition function module, the paper designs The data structure of filling model, i.e. syntax identifies the network, and also realizes the keywords spotting based on the network. Filling model data structure defines node as the initial node, end node, keywords node and NULL syntax node. These nodes constitute recognition network. In Sphinx, NULL is a special grammatical structure. It can automatically match any word outside the keywords table and be suitable for recursive grammatical structure. When the AN4acoustic model as keywords spotting engine model tests the performance of the engine, the experiments showed that the engine has better recognition rate and the correct rate in the case of low false index when using the acoustic model contained eight Gaussian mixture as engine model. When the WSJ acoustic model with the a large vocabulary as engine model tests the performance of the engine, the experiments showed that the engine based on the WSJ acoustic model has good recognition performance and can meet the large vocabulary recognition detection.This paper designs a set evaluation index system based the keywords generic information, and an evaluation algorithm integrated into the system to achieve accurate and automation of speech service quality monitoring. Speech service quality monitoring system based on the keyword spotting technology is developed to evaluate the service quality of the call center speech. The system includes keywords spotting engine and intelligent interactive platform. The system can be integrated into the call center systems in order to achieve co-operating.