Keyword detection method and system based on unlabeled keyword data

Abstract

The invention discloses a keyword detection method and system based on unlabeled keyword data, and the method comprises the steps: collecting a large amount of unlabeled audio data, adding preset wake-up word audio and non-wake-up word audio in the unlabeled audio data, and forming a preprocessing audio library; classifying the audio data in the preprocessed audio library based on an unsupervised deep learning classification method; extracting features from the classified audio data to generate feature data; inputting the feature data into different types of neural network models for training to obtain a plurality of different keyword detection models; and detecting the to-be-predicted audio based on the plurality of different keyword detection models to obtain a final detection result. According to the method, comprehensive comparison classification training is carried out on the audio data, and the classification accuracy is higher.

Description

technical field [0001] The invention relates to the technical field of big data speech classification methods, in particular to a keyword detection method and system based on unmarked keyword data. Background technique [0002] Keyword wake-up technology is relatively common in daily life. For example, smart speaker devices such as Tmall Genie, when people call out the wake-up word to the device, the device can wake up and then interact with people; the realization of the keyword wake-up function requires the wake-up system The device has the characteristics of low power consumption, low false alarm rate, high accuracy rate, and low false negative rate. In order to achieve this goal, a model based on deep learning is usually used. The training of the model in the traditional method requires a large number of labeled key word data to achieve the desired effect; most of the keyword data collected through smart speakers or microphones are unlabeled data. After accumulating a la...

AI Technical Summary

Problems solved by technology

[0003] The training of the deep learning model requires sufficiently large labeled audio data, so the above-mentioned traditional data labeling method consumes a lot of manpower, material resources and time, and the quality of the audio data labeling directly affects the accuracy of the training model; Save the manpower and time cost required for labeling. At present, unsupervised big data audio classification methods are used, such as using spectral segments to classify data, that is, audio classification is performed from the perspective of spectral feature values. First, spectral feature data is extracted from audio data. According to Spectrum features find the optimal classification spectrum matrix, and finally classify the spectral data by frequency bands; although this method considers the characteristics of different frequency bands of audio to classify audio from the perspective of frequency domain, it does not make full use of the comprehensive characteristics of audio , especially time-domain features, there is a problem of low classification accuracy and easy to cause misclassification; at the same time, in the traditional method, the final classification model obtained by inputting the marked audio data set into the deep learning model for training has only one model, through a single The detection of keywords by the classification model is also prone to misclassification

Images