Universal voice awakening identification method and system under full-phoneme frame

Abstract

The invention discloses a universal voice awakening identification method and a system under a full-phoneme frame. The method comprises the following steps: firstly, training a deep neural network acoustic model, modifying a dictionary according to awakening words, constructing a decoding network based on the filler, and training a support vector machine classifier according to training samples; preprocessing the input voice, inputting processed voice characteristics into a decoding network for decoding the processed voice characteristics, calculating an acoustic score according to the deep neural network acoustic model, and obtaining decoding results; inputting the statistical magnitude of successfully recognized decoding results into the support vector machine classifier for classification, and obtaining a final recognition result. According to the method disclosed by the invention, triphonon states obtained through the extension of all atonal phonemes are subjected to modeling, andthen a universal acoustic model is obtained. During the decoding process, a decoding path is limited. Therefore, the awakening performance can be improved. Meanwhile, the later-stage processing part is combined to analyze the multi-dimensional statistics of the phoneme posterior probability and the like on each path. As a result, the hidden danger that the false alarm rate is increased is eliminated.

Description

technical field [0001] The invention relates to the field of voice wake-up, in particular to a universal voice wake-up recognition method and system under the whole phoneme framework. Background technique [0002] With the rapid development of speech recognition technology, speech recognition technology has been widely used, and the application of voice wake-up technology in smart phones and smart homes is increasingly changing human life and production methods. The traditional voice wake-up recognition system is generally constructed for a specific wake-up word (that is, the system is triggered by a user-predetermined wake-up word), and there are mainly two existing methods, one is based on dynamic time warping (dynamic time warping, The DTW) method uses the acoustic features of speech for dynamic matching. First, it is necessary to record several speeches of the same wake-up word in advance, and then dynamically match the speech of the wake-up word collected on-site with t...

AI Technical Summary

Problems solved by technology

[0003] The purpose of the present invention is to overcome the above-mentioned problems existing in the current voice wake-up recognition system, realize the rapid customization of the voice wake-up system by training a general acoustic model, and combine the limited decoding path and post-processing to overcome the problem of high cost of replacing customized words

Images