Modeling method as well as method and device for acquiring acoustic model

Abstract

The application provides a modeling method as well as a method and device for acquiring an acoustic model, which are applied to voice wake-up model processing, wherein the method for acquiring the acoustic model comprises the steps of: acquiring a basic model, wherein the basic model is obtained by carrying out context-independent CI modeling and context-dependent CD modeling on training data, theCD modeling is adopted for the training data in an aggregated state, and the CI modeling is adopted for the training data with monophonemes; for a given wake-up word, determining an output layer nodecorresponding to the wake-up word in output layer nodes of a CD part of the basic model; and modeling the output layer node corresponding to the wake-up word and a remaining part in the basic model to obtain the acoustic model corresponding to the wake-up word. According to the application, keywords can be rapidly customized with low cost, and the shortage of the training data can be avoided.

Description

technical field [0001] The invention relates to the field of speech recognition, in particular to a modeling method, an acoustic model acquisition method and a device. Background technique [0002] Voice wake-up is widely used in smart home and Internet of Things devices. Users can start the device by speaking a pre-customized wake-up word. [0003] There are currently two voice wake-up solutions: [0004] (1) Only standard automatic speech recognition technology is used, and wake-up words and other speech words are modeled without distinction, usually context independent (CI) modeling. [0005] The advantage of this scheme is that the same set of models can be quickly applied to different wake-up word scene requirements, and has great advantages in saving resources and rapid commercialization; but its disadvantages are also obvious: the modeling of this scheme does not highlight Wake-up words, so that during the recognition process, wake-up words have similar scores to ot...

AI Technical Summary

Problems solved by technology

[0008] Solution (2) is currently the main modeling solution based on user-defined voice wake-up, which has a higher wake-up rate and a lower false wake-up rate, and can better meet the actual needs of users; but the disadvantages of this solution are: For new wake-up words generated by different user-defined wake-up words or different wake-up word scenarios (such as different systems or products), the modeling process and data cannot be reused, and new wake-up words need to be rebuilt every time. Therefore, the cost of this scheme is high, and the required training period is long; and this scheme requires a large number of wake-up word samples as training data, and there may be a problem of insufficient training data

Images