Wake-on-voice method and device

A wake-on voice device, comprising:

• a processor;
• and
• a memory configured to store an instruction executable by the processor;
• wherein the processor is configured to:
• obtain a voice inputted by a user,
• process data frames of the voice with a frame skipping strategy and to perform a voice activity detection on the data frames through a time-domain energy algorithm,
• extract an acoustic feature of the voice,
• perform a voice recognition on the acoustic feature according to a preset recognition network and an acoustic model,
• perform an operation corresponding to the voice if the voice is a preset wake-up word in the preset recognition network, and
• stop the voice recognition according to a preset refusal strategy, by acts of:
• judging whether a weighted sum of a voice length obtained by the voice activity detection and a number of junk words is greater than a fourth preset threshold;
• and
• determining that wake-up is failed and stop the voice recognition, if the weighted sum is greater than the fourth preset threshold,
• determine the junk words by a cluster algorithm based on edit distance, by acts of:
• connecting all syllables in parallel to form a candidate recognition network;
• performing a recognition on a preset voice training set according to the candidate recognition network and obtaining recognized results, wherein a number of the recognized results is greater than a first preset threshold;
• selecting a plurality of syllables from all the syllables to form a candidate set of syllables according to the recognized results, wherein a number of recognized results corresponding to any syllable included in the candidate set is greater than that of recognized results corresponding to each syllable not included in the candidate set;
• obtaining a first syllable from the number of recognized results, the first syllable corresponding to a largest number of recognized results from the candidate set, saving the first syllable into a junk word set and removing the first syllable from the candidate set;
• (i) computing an edit distance between a syllable last saved into the junk word set and from each remaining syllable of the candidate set respectively, (ii) saving a syllable Bk into the junk word set, the syllable Bk having a maximum edit distance between the syllable Bk and the first syllable, and (iii) removing the syllable Bk from the candidate set, and
• iteratively for each remaining syllable in the candidate set, repeating the acts of (i) computing, (ii) saving, and (iii) removing until a number of syllables in the junk word set satisfies a second preset threshold;
• and
• configuring the syllables in the junk word set as the junk words of the candidate recognition network.

A non-transitory computer-readable storage medium, having stored therein instructions that, when executed by one or more processors of a device, causes the device to perform a method, wherein the method comprises:

• obtaining a voice inputted by a user;
• processing data frames of the voice with a frame skipping strategy and performing a voice activity detection on the data frames by a time-domain energy algorithm;
• extracting an acoustic feature of the voice and performing a voice recognition on the acoustic feature according to a preset recognition network and an acoustic model;
• performing an operation corresponding to the voice if the voice is a preset wake-up word in the preset recognition network;
• and,
• stopping the voice recognition according to a preset refusal strategy via:
• judging whether a weighted sum of a voice length obtained by the voice activity detection and a number of junk words is greater than a fourth preset threshold, and
• determining that wake-up is failed and stopping the voice recognition, when the weighted sum is greater than the fourth preset threshold,
• determining the junk words by a cluster algorithm based on edit distance, comprising:
• connecting all syllables in parallel to form a candidate recognition network;
• performing a recognition on a preset voice training set according to the candidate recognition network and obtaining recognized results, wherein a number of the recognized results is greater than a first preset threshold;
• selecting a plurality of syllables from all the syllables to form a candidate set of syllables according to the recognized results, wherein a number of recognized results corresponding to any syllable included in the candidate set is greater than that of recognized results corresponding to each syllable not included in the candidate set;
• obtaining a first syllable from the number of recognized results, the first syllable corresponding to a largest number of recognized results from the candidate set, saving the first syllable into a junk word set and removing the first syllable from the candidate set;
• (i) computing an edit distance between a syllable last saved into the junk word set and from each remaining syllable of the candidate set respectively, (ii) saving a syllable Bk into the junk word set, the syllable Bk having a maximum edit distance between the syllable Bk and the first syllable, and (iii) removing the syllable Bk from the candidate set, and
• iteratively for each remaining syllable in the candidate set, repeating the acts of (i) computing, (ii) saving, and (iii) removing until a number of syllables in the junk word set satisfies a second preset threshold;
• and
• configuring the syllables in the junk word set as junk words of the candidate recognition network.