Voice keyword retrieval method based on audio template

Abstract

The invention relates to a voice keyword retrieval method based on an audio template. The method comprises the following steps: to begin with, converting a voice sample template and voice to be retrieved into a probability distribution sequence; then, carrying out matching on the voice sample template and the voice to be retrieved through dynamic time warping to obtain keyword starting and ending time point in the voice to be retrieved and acoustic confidence measure scores of each appearance position; and finally, carrying out warping on the scores obtained by different voice sample templates, and carrying out ranking to obtain a retrieval result. In the retrieval process, information of a specific language is not required at all, thereby maximizing universality and transportability; and meanwhile, operation amount in the retrieval process is reduced, and keyword retrieval speed is accelerated.

Description

technical field [0001] The invention relates to the field of voice retrieval, in particular to a voice keyword retrieval method based on an audio template. Background technique [0002] The keyword retrieval task is to quickly find the location of a given keyword from large-scale and diverse speech data. The current mainstream keyword retrieval method is to convert the speech to be retrieved into text through a continuous speech recognition system with a large vocabulary. Considering the recognition accuracy of the continuous speech recognition system with a large vocabulary, the error rate of the preferred result is relatively high. Therefore, a word map containing multiple candidate information and time information is usually used, and then the text or pronunciation of the keyword to be retrieved is searched on the word map. Search and calculate the confidence level to obtain the keyword retrieval results (Shao Jian, Chinese Speech Retrieval for Large-Scale Telephone Conv...

AI Technical Summary

Problems solved by technology

[0003] However, this method requires a well-performing large-vocabulary continuous speech recognition system. Building such a system requires a large amount of annotated corpus, which significantly increases the cost of application in a new language.

Furthermore, if keywords are given in the form of speech fragments, this method requires first identifying isolated speech fragments as preferred texts, and this process usually has limited precision, further limiting the accuracy of keyword retrieval

Therefore, traditional keyword retrieval methods are usually only applicable to well-understood languages, which limits the application of keyword retrieval

Images