Voice sample screening method based on improved dynamic time warping algorithm

Abstract

The invention discloses a voice sample screening method based on an improved dynamic time warping algorithm. The method comprises the steps: recording a plurality of voice samples based on a same text, removing background sounds, marking vowels and consonants of the voice samples, and constructing a voice feature sequence expression of the samples; according to the annotation types of vowels and consonants in the voice short-time frames after removing of the background sounds, determining transition sounds through change information of the maximum amplitude of signals in the voice short-time frames, and annotating the transition sounds; respectively carrying out weighted calculation on a local distance and an overall distance of an improved DTW algorithm to obtain a distance between everytwo samples, and constructing a distance matrix of all the samples; and screening the voice samples according to the distance matrix. According to the method, a problem of screening the voice samplesof the same text under the conditions that the sample data volume is large and the sample quality cannot be guaranteed is solved, the screening cost is reduced, and more reliable sample data are provided for subsequent processing (such as corpus construction and deep neural network learning).

Description

technical field [0001] The invention relates to the technical field of data processing, in particular to a voice sample screening method based on an improved dynamic time warping algorithm. Background technique [0002] With the rapid development of mobile portable devices and the Internet, voice samples can be obtained from multiple channels, and their data volume is also increasing day by day. Huge speech data provides a foundation for the construction of various language corpora. At the same time, with the development of artificial intelligence, Deep Neural Network (DNN) has achieved remarkable results in the field of speech processing. In 2012, Hinton made new breakthroughs in speech recognition by using deep neural networks. Later, more network structures that conform to the characteristics of speech sequence timing appeared in the research, such as recurrent neural network (Recurrent Neural Network, RNN), long-term short-term memory recurrent neural network (Long Shor...

AI Technical Summary

Problems solved by technology

[0005] Both the construction of the corpus and the learning of the deep neural network require reasonable and correct speech samples. If the speech samples belong to the same text through manual audition, it will cause huge workload, low efficiency

Especially for languages ​​with low resources, such as various dialects of Chinese, it is difficult and costly to rely on manual screening of such speech samples

Images