Self-defined awakening method and device applied to intelligent voice equipment

Abstract

The invention discloses a self-defined awakening method and device applied to intelligent voice equipment. The method comprises the steps of receiving a registration text set by a user and collectingregistration voice of the user; computing a registration awakening threshold based on the registration text; computing a registration awakening score of the registration voice based on a universal awakening model, and judging that whether the registration awakening score is more than the registration awakening threshold or not; if the registration awakening score is more than or equal to the registration awakening threshold, aligning the registration voice with the registration text, and extracting a voice segment corresponding to each character in the registration text; and extracting a Gaussian posterior feature of the voice segment corresponding to each character, and generating a feature template of the whole registration voice based on the Gaussian posterior feature of each voice segment. According to the scheme provided by the method and the device, the quality of input voice can be optimized in a registration stage, so that the system can be much finer in a score computing module, and finally, a mistakenly awakening rate of similar words also can be reduced while an awakening rate is improved.

Description

technical field [0001] The invention belongs to the technical field of voice wake-up, and in particular relates to a self-defined wake-up method and device for intelligent voice equipment. Background technique [0002] In related technologies, on the one hand, the traditional registration voice wake-up scheme is as follows: first, the user inputs 3 to 5 registration voice sentences (for example, Hello Xiaochi) through the device, and then the system will use the registration voice to generate the voice feature template of the registered user. Then, when the user wakes up the device, the system will match the generated voice feature template with the registered voice feature, and calculate the score, and wake up the device when the score exceeds the set threshold. [0003] On the other hand, the custom wake-up and voiceprint recognition serial scheme is as follows: Like the registered voice scheme, the user needs to say 3 to 5 sentences of the registered voice first, and then...

AI Technical Summary

Problems solved by technology

[0006] (1) There is no module to verify the quality of the registered voice. If the registered voice has relatively large noise, or the wake-up words of the registered voice are inconsistent, the system performance will drop sharply;

[0007] (2) Due to the large differences in registered users and registration scenarios, the system cannot give an accurate threshold to balance the wake-up rate and false wake-up rate, and the model is simple, usually the performance will be relatively poor;

[0008] (3) When the phonetic words differ by only one or two words, it is easy to wake up by mistake. Since the wakeup score is calculated directly through the overall matching of phonetic features, the scores of similar words are very close

[0010] (1) Unable to register, because the customized wake-up module and the voiceprint recognition module need to know the wake-up word in advance;

[0011] (2) It takes manpower and material resources to collect a large amount of wake-up word data, train wake-up models, and voiceprint recognition models;

[0012] (3) Adding a voiceprint recognition module, usually based on a neural network model, will increase the amount of calculation and cannot be used in some low-power scenarios

Images