论文部分内容阅读
为实现鲁棒的声韵母切分,以满足大词汇量连续语音识别系统的需求,提出一种建立损失函数,并利用浊音的“准”周期性和声母时长进行声韵母切分的方法。首先计算语音的自相关函数,接着建立代价损失函数,对计算结果采用动态规划方法检测浊音,然后根据声母时长分布规律确定声母的检测范围,最后在检测范围内对浊音段起始点前后采用听觉事件检测方法分割出声韵母。实验结果表明,采用动态规划方法相对于阈值方法提高了浊音段的检测性能,在浊音段的基础上对声韵母进行切分能够提高切分的正确率,减少噪声及汉语音变现象的影响,切分性能受声母发音方式影响较小。
In order to realize the robust segmentation of vowels and vowels to meet the demands of large vocabulary continuous speech recognition system, a method of establishing the loss function and using the “quasi” period and initials of voiced sounds for vowels segmentation . Firstly, the autocorrelation function of speech is calculated, and then the cost-loss function is established. Then, the dynamic programming method is used to detect voiced speech. Then the initial consonant detection range is determined according to the distribution of initial consonant length. Finally, auditory events Detection method segmentation vocal consonant. The experimental results show that the dynamic programming method improves the detection performance of the voiced speech segment compared with the threshold method. Segmenting the vowel vowels based on the voiced speech segment can improve the segmentation accuracy, reduce the noise and the influence of the Chinese phonetic change, The segmentation performance is less affected by the initial sounding method.