基于无监督学习的语音签到系统

Voice Check-in System Based on Unsupervised Learning

针对语音签到系统在实际运用中识别率较低的问题,从提高对标签缺失数据的利用角度出发,提出一种利用无监督学习来提高识别率的方法。该方法基于深度置信网络隐马尔可夫混合模型(DBN-HMM),利用受限波尔茨曼机(RBM)为无监督学习提取特征参数,接着利用深度置信网络(DBN)得到对原始数据的观测概率。隐马尔可夫(HMM)据此通过前向算法求出数据的似然概率,并将概率值最大的类别作为识别结果。实验表明,使用DBN-HMM模型可以有效利用存在标签缺失的数据,提高语音签到系统的识别能力。

Aiming at the low recognition rate of speech check-in system in practical application,this paper proposed a method,from improving the utilization of tag missing data,to improve recognition rate by unsupervised learning.This method was based on Deep Belief Network mixed Hidden Markov Model(DBN-HMM),used the Restricted Boltzmann Machine(RBM)to extract the characteristic parameters for unsupervised learning,then used Deep Belief Network(DBN)to get the observation probability of raw data.Based on this,Hidden Markov Model(HMM)calculated the likelihood probability of data by forward algorithm,and took the category with the largest probability value as recognition results.Experiments showed that DBN-HMM model could effectively utilize the data with missing tags and improved the recognition ability of speech check-in system.