语音识别中基于模糊聚类分析的参数聚类

为减少语音识别中声学模型的参数量,提高参数训练的鲁棒性,基于声学决策树结构,提出利用模糊聚类分析方法对模型参数聚类,包括高斯聚类和方差共享．对大词汇量汉语连续语音识别的实验结果表明：高斯模糊聚类使高斯数减少25％时,识别率提高了0．15％．进一步做模糊方差共享,当方差减少到初始模型的24％,与同样参数量的未进行聚类的模型相比,误识率下降了3．01％,证明了模糊聚类分析在语音参数聚类中的有效性．

Parameter Clustering Based on Fuzzy Clustering Analysis in Speech Recognition

In order to decrease the parameter size of acoustic models in speech recognition and improve the robustness of parameter training, a fuzzy clustering analysis method was used for parameter clustering. Based on the structure of phonetic decision tree, the proposed parameter clustering method includes Gaussians clustering and covariance sharing. The experimental results on large vocabulary continuous Mandarin speech recognition show when the number of Gaussians is reduced by 25%, the recognition accuracy increases by 0.15%. Combined with covariance sharing, 24% covarinaces achieve 3.01% error rate reduction over the system with approximately the same parameter size. These results prove the effectiveness of fuzzy clustering in speech recognition.