Efficient voice keyword detector training sample use method

Abstract

The invention discloses an efficient voice keyword detector training sample use method. The method comprises the following steps: training a detector for n rounds by using all samples; in the preparation stage of the kth round of training, acquiring a target score S (T) and a competitor score S (C) of a sample based on a detector Dk-1 obtained in the (k-1) th round of training, and constructing the probability that the sample participates in the kth round of training; selecting a sample set Zk used in the kth round according to the probability that the samples participate in training; and using the Zk for training in the kth round, acquiring a new detector Dk, repeating the processes of obtaining the probability that the samples participate in training, selecting the samples and training, and ending the training till the model converges or the training round is larger than a preset value. According to the method, the samples capable of providing more distinguishing information for model training are selected for training, so that the training pays more attention to the important samples, and the training efficiency is improved while the system performance is improved; the problem of class sample imbalance in keyword detection training can be relieved; and the method is simple and efficient, and has a wide application prospect.

Description

technical field [0001] The invention relates to the technical field of data processing, in particular to a method for using training samples of an efficient speech keyword detector. Background technique [0002] Voice is an important medium for information exchange. Speech keyword spotting (Keyword Spotting, KWS) refers to the detection of predefined words from a continuous speech stream. Compared with Automatic Speech Recognition (ASR), KWS only pays attention to the defined keywords, and does not pay attention to words other than keywords, which greatly reduces the difficulty of system development. Due to its fast and flexible features, KWS has a wide range of applications in the fields of audio monitoring, voice retrieval, and device control. [0003] For KWS training, the collection cost of positive samples is high, while the acquisition of negative samples is relatively easy. At the same time, the training of KWS does require a large number of rich and diverse negati...

AI Technical Summary

Problems solved by technology

However, if samples are discarded at will, samples that are "important" to training may be discarded, resulting in underfitting

Images