Equipment control method and device, equipment and readable storage medium

Abstract

The invention relates to an equipment control method and device, equipment and a readable storage medium, and relates to the technical field of intelligent voice. Comprising the following steps: acquiring a to-be-recognized audio which comprises voice information; identifying the to-be-identified audio based on a preset wake-up word identification model, and determining whether the to-be-identified audio contains a preset wake-up word; and if the to-be-recognized audio contains the preset wake-up word, controlling the equipment to enter a wake-up state, otherwise, controlling the equipment to keep a non-wake-up state. By using the device control method provided by the invention, the preset wake-up word and the non-preset wake-up word can be accurately distinguished, so that the false wake-up rate of the device is reduced.

Description

Technical Field

[0001] The present disclosure relates to the field of intelligent voice technology, and in particular to a device control method, apparatus, device and readable storage medium. Background Art

[0002] Voice wake-up means that the device detects a specific segment of the speaker in a continuous voice stream in real time. Voice wake-up is usually used to activate devices or turn on the system. These devices will set a specific preset wake-up word. When the user says the preset wake-up word, it means that he wants to wake up the device to issue instructions to the device.

[0003] However, current devices have a high false wake-up rate. Summary of the invention

[0004] In order to overcome the problems existing in the related art, the present disclosure provides a device control method, apparatus, device and readable storage medium.

[0005] According to a first aspect of an embodiment of the present disclosure, there is provided a device control method, including:

[0006] Acquire audio to be recognized, where the audio to be recognized contains voice information;

[0007] The audio to be recognized is recognized based on a preset wake-up word recognition model to determine whether the audio to be recognized contains a preset wake-up word; wherein the wake-up word recognition model is trained based on training samples by a machine learning method, and the training samples include a positive sample set and a negative sample set, the positive sample set includes multiple positive samples, each of the positive samples includes a positive audio and a positive label, the positive audio is the speech audio of the preset wake-up word, and the positive label includes symbols corresponding to all phonemes in the preset wake-up word; the negative sample set includes multiple negative samples, each of the negative samples includes a negative audio and a negative label, the negative audio is the speech audio that does not contain any phonemes in the preset wake-up word, and the negative label uses a preset symbol;

[0008] If the audio to be recognized contains the preset wake-up word, the device is controlled to enter a wake-up state; otherwise, the device is controlled to remain in a non-wake-up state.

[0009] Optionally, the training sample also includes a similar sample set, which includes multiple similar samples, each similar sample includes similar audio and similar labels, the similar audio refers to speech audio containing some phonemes in the preset wake-up word, and the similar label contains the preset symbol and the symbol in the similar audio that has the same phonemes as the preset wake-up word.

[0010] Optionally, the identifying the audio to be identified based on a preset wake-up word recognition model to determine whether the audio to be identified contains a preset wake-up word includes:

[0011] Segmenting the audio to be recognized to obtain multiple audio frames in time series;

[0012] Extracting acoustic features from each of the audio frames to obtain acoustic features corresponding to each audio frame;

[0013] Inputting the acoustic features corresponding to the plurality of audio frames into the wake-up word recognition model in sequence to obtain an acoustic score corresponding to each of the acoustic features; wherein the acoustic score includes a plurality of probability values, each probability value indicating the probability that the acoustic feature is a phoneme corresponding to the preset wake-up word;

[0014] Processing the audio frame according to the acoustic score of the acoustic feature and a preset decoding graph to obtain an optimal decoding path corresponding to the audio to be recognized;

[0015] If the optimal decoding path is consistent with the preset wake-up path corresponding to the preset wake-up word, it is determined that the audio to be recognized contains the preset wake-up word.

[0016] Optionally, the decoding graph includes multiple nodes, at least two of the nodes constitute any one of the decoding paths among the preset wake-up path, the non-wake-up path and the similar wake-up path, the preset wake-up path is a decoding path indicating that the device enters the wake-up state, and the non-wake-up path and the similar wake-up path are decoding paths indicating that the device enters the non-wake-up state.

[0017] Optionally, the processing according to the acoustic score of the acoustic feature of each audio frame and a preset decoding graph to obtain an optimal decoding path corresponding to each audio frame includes:

[0018] Obtaining, according to the acoustic score of each audio frame, a cost score of a decoding path corresponding to each audio frame in the decoding graph;

[0019] From the multiple decoding paths, a decoding path corresponding to the minimum cost score is selected as the optimal decoding path.

[0020] Optionally, the plurality of decoding paths include a first decoding path, and obtaining, according to the acoustic score of each audio frame, a cost score of each audio frame passing through a corresponding decoding path in the decoding graph includes:

[0021] A cost score of the first decoding path is obtained according to the acoustic score corresponding to the first decoding path and the weight of the first decoding path.

[0022] Optionally, the plurality of decoding paths include a non-first decoding path, and obtaining, according to the acoustic score of each audio frame, a cost score of each audio frame passing through a corresponding decoding path in the decoding graph includes:

[0023] The cost score of the non-first decoding path is obtained according to the cost score of the previous decoding path, the acoustic score of the non-first decoding path and the weight of the non-first decoding path.

[0024] Optionally, all the anti-labels in the anti-sample set use the same preset symbol.

[0025] Optionally, the number of phonemes that are the same between the similar audio contained in the similarity tag and the preset wake-up word is N, and N is greater than or equal to 1.

[0026] According to a second aspect of an embodiment of the present disclosure, there is provided a device control apparatus, including:

[0027] An acquisition module is configured to acquire audio to be recognized, where the audio to be recognized contains voice information;

[0028] A recognition module is configured to recognize the audio to be recognized based on a preset wake-up word recognition model, and determine whether the audio to be recognized contains a preset wake-up word; wherein the wake-up word recognition model is trained based on training samples by a machine learning method, and the training samples include a positive sample set and a negative sample set, the positive sample set includes multiple positive samples, each of the positive samples includes a positive audio and a positive label, the positive audio is a voice audio of a preset wake-up word, and the positive label includes symbols corresponding to all phonemes in the preset wake-up word; the negative sample set includes multiple negative samples, each of the negative samples includes a negative audio and a negative label, the negative audio is a voice audio that does not contain any phonemes in the preset wake-up word, and the negative label uses a preset symbol;

[0029] The control module is configured to control the device to enter a wake-up state if the audio to be recognized contains the preset wake-up word, otherwise control the device to remain in a non-wake-up state.

[0030] According to a third aspect of an embodiment of the present disclosure, a device is provided, including:

[0031] processor;

[0032] a memory for storing processor-executable instructions;

[0033] Wherein, the processor is configured to:

[0034] Execute the steps of the device control method provided in the first aspect of the embodiment of the present disclosure.

[0035] According to a fourth aspect of an embodiment of the present disclosure, a computer-readable storage medium is provided, on which computer program instructions are stored. When the program instructions are executed by a processor, the steps of the device control method provided in the first aspect of the present disclosure are implemented.

[0036] The technical solution provided by the embodiments of the present disclosure may have the following beneficial effects:

[0037] After the device can recognize the audio to be recognized, it recognizes the audio to be recognized based on the wake-up word recognition model, and determines whether the audio to be recognized contains the preset wake-up word. If it does not contain the preset wake-up word, the device is controlled to remain in a non-wake-up state. If it contains the wake-up word, the device is controlled to enter a wake-up state.

[0038] In this process, since the training samples for training the wake-up word recognition model include a positive sample set, which contains positive audio and positive labels, the trained wake-up word recognition model can accurately identify whether the audio to be recognized contains a preset wake-up word; since the training samples include a negative sample set, which contains negative audio and negative labels, the trained wake-up word recognition model can accurately identify whether the audio to be recognized contains non-preset wake-up words. In this way, the wake-up word recognition model can accurately distinguish between preset wake-up words and non-preset wake-up words, thereby reducing the false wake-up rate of the device.

[0039] It is to be understood that the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the present disclosure. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the present disclosure and, together with the description, serve to explain the principles of the present disclosure.

[0041] Figure 1 is a flow chart of a device control method according to an exemplary embodiment;

[0042] Figure 2 is a schematic diagram showing an operation of waking up a device according to an exemplary embodiment;

[0043] Figure 3 is a schematic diagram showing another operation of waking up a device according to an exemplary embodiment;

[0044] Figure 4 is a logic diagram of a device control method according to an exemplary embodiment;

[0045] Figure 5 is a schematic diagram of a wake-up word recognition model obtained through training according to an exemplary embodiment;

[0046] Figure 6is a logic schematic diagram of another device control method according to an exemplary embodiment;

[0047] Figure 7 is a schematic diagram of a decoding diagram according to an exemplary embodiment;

[0048] Figure 8 is a schematic diagram of another decoding diagram according to an exemplary embodiment;

[0049] Fig. 9 is a block diagram of a device control apparatus according to an exemplary embodiment;

[0050] Fig.10 It is a block diagram of a device according to an exemplary embodiment. DETAILED DESCRIPTION

[0051] Exemplary embodiments will be described in detail herein, examples of which are shown in the accompanying drawings. When the following description refers to the drawings, the same numbers in different drawings represent the same or similar elements unless otherwise indicated. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present disclosure. Instead, they are merely examples of devices and methods consistent with some aspects of the present disclosure as detailed in the appended claims.

[0052] It should be noted that all actions of acquiring signals, information or data in the present disclosure are carried out in compliance with the relevant data protection laws and policies of the country where the device is located and with the authorization given by the owner of the corresponding device.

[0053] Figure 1 is a flow chart of a device control method according to an exemplary embodiment. Figure 1 As shown, the device control method is applied to a device, which may be a mobile phone, a television, a speaker, a smart home, etc. The smart home may be a washing machine, an air conditioner, a microwave oven, etc., which have a voice recognition function. The device control method includes the following steps.

[0054] In step S11, audio to be recognized is obtained, where the audio to be recognized contains voice information.

[0055] The audio to be recognized refers to the voice stream emitted by the sound source in the environment where the device is located, such as the voice emitted by the user, the voice emitted by the TV or mobile phone, or the voice emitted by animals, etc. The device is equipped with an audio collection device such as a microphone to collect the audio emitted by the sound source. The audio to be recognized can also be the voice stream emitted by the device itself. For example, when the device is a TV, the audio emitted by the sound source is the voice stream played by the TV.

[0056] In step S12, the audio to be recognized is recognized based on a preset wake-up word recognition model to determine whether the audio to be recognized contains a preset wake-up word.

[0057] The wake-up word recognition model is trained based on training samples through machine learning methods. Figure 5 As shown, the training samples include a positive sample set and a negative sample set.

[0058] The positive sample set contains multiple positive samples, each of which contains positive audio and positive labels. The positive audio is the speech audio of the preset wake-up word, and the positive label contains the symbols corresponding to all phonemes in the preset wake-up word.

[0059] The preset wake-up word is a pre-set wake-up word, for example, "Xiao Ai Tongxue". "Xiao Ai Tongxue" has seven phonetic symbols: "x, iao, ai, t, ong, x, ue".

[0060] For the positive sample set A = {a1, a2, a3, ...a n}, contains n positive samples, each positive sample is represented by a n Represents. Each positive sample a n Contains: positive audio and positive label. Positive audio refers to the voice audio corresponding to the preset wake-up word. For example, if the preset wake-up word is "Xiao Ai Tongxue", then the positive audio is the voice audio of "Xiao Ai Tongxue". The positive label contains the symbols of all the phonemes in the preset wake-up word. For example, the pinyin of "Xiao Ai Tongxue" is "xiao ai tong xue", which contains seven phoneme symbols. The seven phoneme symbols are x, iao, ai, t, ong, x, ue, and the positive label is these seven phoneme symbols.

[0061] The anti-sample set contains multiple anti-samples, each of which contains anti-audio and anti-label. The anti-audio is a speech audio that does not contain any phonemes in the preset wake-up word, and the anti-label uses a preset symbol. The anti-audio does not contain any phonemes in the preset wake-up word, which can also be understood as the anti-audio containing a non-preset wake-up word, and any phoneme in the non-preset wake-up word is different from any phoneme in the preset wake-up word. For example, the non-preset wake-up word "Have you eaten?" has seven phoneme symbols "chi, f, an, l, e, m, a", and the seven phoneme symbols of "Have you eaten?" are different from the seven phoneme symbols of "Xiao Ai Tongxue" "x, iao, ai, t, ong, x, ue".

[0062] For the inverse sample set B = {b1, b2, b3, ...b m}, contains m counter samples, each counter sample uses b m Each negative sample b mIncludes: anti-audio and anti-label. Anti-audio refers to speech audio that does not contain any phonemes in the preset wake-up word. For example, "Have you eaten?" and "What day is today?" do not contain the phonemes of the preset wake-up word. Anti-labels use preset symbols, and all anti-labels in the anti-sample set use exactly the same preset symbols. For example, if the anti-audio is the speech audio of "Have you eaten?" and "What day is today?", then the anti-labels of "Have you eaten?" and "What day is today?" can be represented by the preset symbol filler. Of course, the preset symbols can also be represented by other symbols, and the present disclosure does not limit this.

[0063] The reason for uniformly representing the anti-labels with preset symbols is that for the wake-up word recognition model, all non-preset wake-up words are the same, and there is no need to distinguish and recognize different non-preset wake-up words to wake up the device, so the labels of all non-preset wake-up words can be set to preset symbols, thereby reducing the computational complexity of the wake-up word recognition model.

[0064] Since each user speaks the preset wake-up word at a different speed, the duration of each user speaking the preset wake-up word is different. Based on this, the device averages the duration of the audio to be recognized spoken by multiple different users to obtain the average duration. When the device subsequently collects audio from the sound source, it can collect audio that is longer than the average duration, thereby ensuring that the acquired audio contains the complete preset wake-up word.

[0065] In step S13, if the audio to be recognized contains the preset wake-up word, the device is controlled to enter the wake-up state, otherwise the device is controlled to remain in the non-wake-up state.

[0066] If the audio to be recognized contains a preset wake-up word, it means that the audio emitted by the sound source is used to control the device to wake up. At this time, the device can be controlled to switch from a non-awakening state to a wake-up state. The sound source emits a preset wake-up word when the sound source and the device are in the same space (for example, when the sound source is a user, the user emits the preset wake-up word toward the device), or when the sound source and the device are in different spaces (for example, when the sound source is a remote user, the remote user remotely sends a voice of the preset wake-up word to the near-end user through a mobile phone, and the near-end user plays the preset wake-up word through a mobile phone, so that the remote user can remotely control the device to wake up).

[0067] The awake state refers to a state in which the device can perform operations in response to user instructions. For example, the user outputs voice to let the speaker play music, and the speaker performs the music playback operation. This operation is performed in the awake state.

[0068] If the audio to be recognized does not contain the preset wake-up word, it means that the audio emitted by the sound source is not used to control the device to wake up. At this time, the device can be controlled to be in a non-wake-up state.

[0069] The non-wake-up state includes the standby state of the device. In the standby state, the device continuously monitors the audio emitted by the sound source in the surrounding environment with low power consumption. When the preset wake-up word is detected, the device enters the working state and responds.

[0070] Through the above technical solution, since the training samples for training the wake-up word recognition model include a positive sample set, which contains positive audio and positive labels, the trained wake-up word recognition model can accurately identify whether the audio to be recognized contains a preset wake-up word; since the training samples include a negative sample set, which contains negative audio and negative labels, the trained wake-up word recognition model can accurately identify whether the audio to be recognized contains non-preset wake-up words. In this way, the wake-up word recognition model can accurately distinguish between preset wake-up words and non-preset wake-up words, thereby reducing the false wake-up rate of the device.

[0071] The following introduces an optional embodiment involved in the above step S12, which is used to explain that in addition to using positive sample sets and negative sample sets to train the wake-up word recognition model, a similar sample set can also be used to train the wake-up word recognition model, thereby further reducing the false wake-up rate of the device.

[0072] In the related art, a user wakes up a device by saying a preset wake-up word, and a user may also wake up the device by saying a similar wake-up word, which results in a high false wake-up rate of the device and a poor user experience. For example, see Figure 2 As shown, when user A says the preset wake-up word "Xiao Ai", the device will be woken up normally and reply with the voice "I am here"; when user B says the similar wake-up word "Wang", the device may also be mistakenly woken up and reply with the voice "I am here".

[0073] In the present disclosure, when user A says the preset wake-up word "Xiao Ai", the device will be woken up normally and reply with the voice "I am here"; when the user says a non-preset wake-up word or a similar wake-up word, the device will not be woken up. For example, see Figure 3 As shown, when user A says a similar wake-up word such as "Mr. Wang", the device remains in a non-awakening state and does not respond to the user's voice; when user B says a non-preset wake-up word such as "Hahaha", the device still remains in a non-awakening state and does not respond to the user's voice.

[0074] To identify similar wake-up words, see Figure 5 As shown, the training samples also include a similar sample set, which includes multiple similar samples. Each similar sample includes similar audio and similar labels. The similar audio refers to speech audio containing some phonemes in the preset wake-up word, and the similar label includes preset symbols and symbols in the similar audio that have the same phonemes as the preset wake-up word.

[0075] For a similar sample set C = {c1, c2, c3, ...cz}, contains z similar samples, each similar sample is represented by c z Each similar sample c z Contains: similar audio and similar tags.

[0076] Similar audio refers to speech audio that contains some phonemes of the preset wake-up word. Taking the similar audio that contains four phonemes of the preset wake-up word as an example, if the preset wake-up word is "Xiao Ai classmate", if the audio is "Ni Hao classmate", "Tong Xie", "Wang classmate", "Tong Xue" and other speech audios, the four ending phonemes of these audios "t, ong, x, ue" are the same as the four ending phonemes of the preset wake-up word "Xiao Ai classmate", "t, ong, x, ue", and the audio contains four of the seven phonemes of the preset wake-up word, so these audios are similar audios.

[0077] The similarity label includes a preset symbol and N symbols in the similar audio that have the same phonemes as the preset wake-up word, where N is greater than or equal to 1. For example, for the phonemes of the similar audio "Wang Tongxue", it has the same four phonemes "t, ong, x, ue" as the preset wake-up word "Xiao Ai Tongxue", so three of the four phonemes "t, ong, x" can be replaced with the preset symbol filler, and the remaining "ue" phoneme remains unchanged, and the final similarity label is "filler ue". When selecting a symbol with the same phonemes as the preset wake-up word in the similar audio as a similarity label, you can choose one identical phoneme to remain unchanged, or you can choose multiple identical phonemes to remain unchanged. For example, you can choose to keep the phoneme ue unchanged and replace the remaining phonemes with the preset symbol filler to obtain a similar label "filler ue", or you can choose to keep the phonemes x and ue unchanged, and replace the remaining phonemes with the preset symbol filler to obtain a similar label "filler x ue", which is not limited in the comparison of this disclosure.

[0078] The reason for replacing some of the phonemes that are identical to the preset wake-up words in the similar audio with preset symbols, and keeping the other phonemes unchanged to obtain similar labels is that: if all the phonemes that are identical to the preset wake-up words in the similar audio are used as similar labels, for example, the four phonemes "t, ong, x, ue" of "Mr. Wang" are all used as similar labels, it will cause the similar labels to be too close to the seven phonemes of the positive labels "x, iao, ai, t, ong, x, ue", and there are too many overlapping phonemes between the preset wake-up words and the similar wake-up words, which affects the wake-up word recognition model from identifying the true preset wake-up words. Therefore, some of the phonemes that are identical to the preset wake-up words in the similar audio are replaced with preset symbols, and the other phonemes remain unchanged to obtain the similar label "filler ue", which is significantly different from the positive label "x, iao, ai, t, ong, x, ue" of the positive sample, so as to improve the accuracy of the wake-up word recognition model in identifying the preset wake-up words.

[0079] After the wake-up word recognition model has the function of recognizing similar wake-up words, if the audio to be recognized contains similar wake-up words, the control device remains in a non-awakening state.

[0080] In daily life, there are various audios around the device. Taking the preset wake-up word "Xiao Ai" as an example, users may mention a similar wake-up word "Tong Xie" when talking to each other, or a non-similar wake-up word "You go home immediately". Users may also point out the preset wake-up word "Xiao Ai". Therefore, after the wake-up word recognition model has the ability to distinguish between similar wake-up words, non-preset wake-up words and preset wake-up words, it can distinguish between these three types of voice audio, keep the device in a non-wake-up state when the user does not want to wake up the device, and thereby reduce the device's false wake-up rate.

[0081] Through the above technical solution, the device can control the device to be in a non-awakening state when the audio contains similar wake-up words, thereby avoiding the device being mistakenly awakened when the user says similar wake-up words, reducing the probability of the device being mistakenly awakened and improving the user experience.

[0082] The following describes a specific embodiment of step S12, which is used to explain how to determine whether the audio to be recognized contains a non-preset wake-up word, a preset wake-up word, or a similar wake-up word. Figure 4 and Figure 6 As shown, the following steps are included.

[0083] In step S21, the audio to be recognized is segmented to obtain a plurality of time-series audio frames.

[0084] Segmenting the audio to be recognized refers to segmenting the continuous audio signal of the audio to be recognized into a series of short-term audio frames, each of which contains an audio signal of a period of time, and the audio frame is usually within a time range of several milliseconds to tens of milliseconds.

[0085] The purpose of segmenting the audio to be recognized is to convert the continuous audio signal into discrete audio frames so that subsequent processing and analysis can be performed on each audio frame, and different operations such as independent sound recognition, speech synthesis and audio encoding can be performed on each audio frame.

[0086] In step S22, acoustic features are extracted from each of the audio frames to obtain acoustic features corresponding to each audio frame.

[0087] Acoustic feature extraction of audio frames refers to extracting a set of numerical representations from audio frames to identify the acoustic features of the audio signal of the audio frame. These acoustic features are usually used for tasks such as speech recognition, speech synthesis and speech classification.

[0088] In step S23, the acoustic features corresponding to the multiple audio frames are input into the wake-up word recognition model in sequence to obtain an acoustic score corresponding to each of the acoustic features; wherein the acoustic score includes multiple probability values, each probability value representing the probability that the acoustic feature is a phoneme corresponding to the preset wake-up word.

[0089] The acoustic score is used to evaluate the similarity or probability that the acoustic feature is each of the preset wake-up words. Each acoustic score corresponds to multiple probabilities, and each probability is used to indicate the probability that the acoustic feature is a phoneme in the preset wake-up word. The maximum probability among multiple probabilities can be used as the acoustic score of the acoustic feature.

[0090] For example, after the acoustic features of a certain audio frame are input into the wake-up word recognition model, if the phoneme of the audio frame is "ue", and the phonemes of the preset wake-up word "Xiao Ai" are "x, iao, ai, t, ong, x, ue", then the probabilities that the phoneme "ue" of the audio frame is these seven phonemes are 10%, 10%, 0, 0, 10%, 5%, and 99.7%, respectively. The maximum probability of 99.7% can be used as the acoustic score of the acoustic feature.

[0091] For the audio to be recognized, after the audio to be recognized F is divided into frames, X audio frames are obtained, which are represented as F = {f1, f2, f3, ... f x}, the acoustic scores of the acoustic features corresponding to the multiple audio frames calculated by the wake-up word recognition model are H = {h1, h2, h3, ... h x}, h1 represents the acoustic score of the acoustic feature of the first audio frame f1, h xIndicates the Xth audio frame f x The acoustic score of the acoustic feature.

[0092] In step S24, the acoustic score of the acoustic feature of each audio frame and a preset decoding graph are processed to obtain an optimal decoding path corresponding to the audio to be recognized.

[0093] A decoding graph is a graphical representation structure in the process of speech recognition using acoustic models and language models in the field of speech recognition technology. The decoding graph contains multiple nodes, each node represents a different time step, each time step corresponds to an audio frame, and the edges between two adjacent nodes form a decoding path, indicating the transfer from the previous audio frame to the next audio frame.

[0094] At least two of the nodes constitute any one decoding path among the preset wake-up path, the non-wake-up path and the similar wake-up path, the preset wake-up path is a decoding path indicating that the device enters the wake-up state, and the non-wake-up path and the similar wake-up path are decoding paths indicating that the device enters the non-wake-up state.

[0095] See also Figure 7 The decoding diagram shown, Figure 7 The sil in it means silence; <eps>It is empty, which means nothing is output; filler indicates all pronunciations except the preset wake-up word phonemes; kw-in indicates the start of wake-up; kw-out indicates output wake-up and wake-up is successful.

[0096] For example, with sil: <eps> / 1" as an example, it means that the input from the previous node to the next node is silence, the output is empty, and the weight of the decoding path between the previous node and the next node is 1; for example, x:kw-in / 5, it means that the input from the previous node to the next node is phoneme x, the output is the start of wake-up, and the weight of the decoding path between the previous node and the next node is 5; for example, ue:kw-out / 5, it means that the input from the previous node to the next node is factor ue, the output is wake-up success, and the weight of the decoding path between the previous node and the next node is 5.

[0097] The decoding graph proposed in the present disclosure includes 10 nodes from 0 to 9, 0 is the starting node, and 1, 8 and 9 are the ending nodes. Of course, it is not limited to these 10 nodes. When the number of phonemes of the preset wake-up word changes, the number of nodes will also change accordingly, and the present disclosure does not limit this.

[0098] Node 0 has 3 downstream nodes (node ​​1, node 2, and node 9), and there are 4 decoding paths from node 0 to its downstream nodes, namely: L0-11, L0-12, L0-2, and L0-9. Among them, L0-11 represents the first decoding path from node 0 to node 1, L0-12 is the second decoding path from node 0 to node 1, L0-2 is the decoding path from node 0 to node 2, and L0-9 is the decoding path from node 0 to node 9.

[0099] Node 1 has 1 downstream node (node ​​0), and there is 1 decoding path from node 1 to its downstream node: L1-0.

[0100] Node 2 has two downstream nodes (node ​​2 and node 3). There are two decoding paths from node 2 to its downstream nodes, namely L2-2 and L2-3, which represent the decoding path from node 2 to node 2 and the decoding path from node 2 to node 3 respectively.

[0101] Node 3 has two downstream nodes (node ​​3 and node 4). There are two decoding paths from node 3 to its downstream nodes, namely L3-3 and L3-4, which represent the decoding path from node 3 to node 3 and the decoding path from node 3 to node 4 respectively.

[0102] Node 4 has two downstream nodes (node ​​4 and node 5). There are two decoding paths from node 4 to its downstream nodes, namely L4-4 and L4-5, which represent the decoding path from node 4 to node 4 and the decoding path from node 4 to node 5 respectively.

[0103] Node 5 has two downstream nodes (node ​​5 and node 6). There are two decoding paths from node 5 to its downstream nodes, namely L5-5 and L5-6, which represent the decoding path from node 5 to node 5 and the decoding path from node 5 to node 6 respectively.

[0104] Node 6 has two downstream nodes (node ​​6 and node 7). There are two decoding paths from node 6 to its downstream nodes, namely L6-6 and L6-7, which represent the decoding path from node 6 to node 6 and the decoding path from node 6 to node 7 respectively.

[0105] Node 7 has two downstream nodes (node ​​7 and node 8). There are two decoding paths from node 7 to its downstream nodes, namely L7-7 and L7-8, which represent the decoding path from node 7 to node 7 and the decoding path from node 7 to node 8 respectively.

[0106] Node 8 has two downstream nodes (node ​​8 and node 0). There are two decoding paths from node 8 to its downstream nodes, namely L8-8 and L8-0, which represent the decoding path from node 8 to node 8 and the decoding path from node 8 to node 0 respectively.

[0107] Node 9 has two downstream nodes (node ​​9 and node 0). There are two decoding paths from node 9 to its downstream nodes, namely L9-9 and L9-0, which represent the decoding path from node 9 to node 9 and the decoding path from node 9 to node 0 respectively.

[0108] Figure 7 In the example, each node has an arrow pointing to itself. The purpose of this design is that when the user pronounces a phoneme, a phoneme has a duration. So when the phoneme appears within the duration, it will always go to the decoding path of the node with the arrow pointing to itself. For example, if the duration of phoneme x is 1ms, then within the duration of 1ms, it will always go to the decoding path of L. 22 For example, if the duration of the phoneme ong is 2ms, then within the duration of 2ms, it will consistently follow the L decoding path. 66 The decoding path.

[0109] The decoding path connecting every two nodes corresponds to a decoding rule, which includes judgment conditions and corresponding weights. The judgment conditions and weights will be explained below. Figure 7 In the decoding diagram shown, the decoding path between node 0 and node 9 is a similar wake-up path, at least one decoding path between node 0 and node 1 is a non-preset wake-up path, and at least one decoding path between node 1 and node 8 is a preset wake-up path.

[0110] The optimal decoding path corresponding to the audio to be identified can be obtained by searching from a preset decoding graph according to the acoustic features of each audio frame, including: inputting the acoustic score and the decoding graph of each audio frame into the decoder, and the decoder determines the optimal decoding path corresponding to the multiple audio frames in the decoding graph through the judgment conditions and weights of each node in the decoding graph, specifically including the following sub-steps:

[0111] Sub-step A1: according to the acoustic score of each audio frame, obtain the cost score of each audio frame passing through the corresponding decoding path in the decoding graph.

[0112] The cost score is used to evaluate the match between the audio emitted by the sound source and the preset wake-up word, or the match between the audio emitted by the sound source and the non-preset wake-up word, or the match between the audio emitted by the sound source and a similar wake-up word. The higher the acoustic score, the smaller the cost score and the higher the match. The smaller the cost score, the easier it is for the audio to pass through the decoding path; the larger the cost score, the less likely the audio is to pass through the decoding path.

[0113] The decoding graph includes multiple nodes, and two adjacent nodes among the multiple nodes constitute a decoding path. The decoding path types include the first decoding path and the non-first decoding path. The first decoding path refers to the first decoding path for calculating the first audio frame. The non-first decoding path includes the second decoding path, the third decoding path, etc. The second decoding path is located downstream of the first decoding path, the third decoding path is located downstream of the second decoding path, and so on. Since the decoding path is divided into the first decoding path and the non-first decoding path, the cost score of the decoding path will be obtained from the following two aspects.

[0114] In a first aspect, for a first decoding path among multiple decoding paths, a cost score of the first decoding path is obtained according to an acoustic score corresponding to the first decoding path and a weight of the first decoding path.

[0115] Above Figure 7 Among the multiple decoding paths, four decoding paths, L0-11, L0-12, L0-2, and L0-9, are the first decoding paths, and the remaining decoding paths are non-first decoding paths.

[0116] For calculating the cost score of the first decoding path, the cost score of the first decoding path can be obtained according to the cost score of the previous decoding path, the acoustic score corresponding to the first decoding path and the weight of the first decoding path. The calculation formula is as follows:

[0117] L 0N =-log(h1)+w 0N (1)

[0118] L 0N represents the cost score from node 0 to node N, w 0N Represents the weight of the first decoding path between node 0 and node N, where N can be any one of 11, 12, 2, and 9. h1 is the acoustic score of the acoustic feature corresponding to the first audio frame.

[0119] When N is 11, L 011 represents the cost score of the first decoding path of the two first decoding paths from node 0 to node 1, W 011 Represents the weight of the first decoding path from node 0 to node 1.

[0120] When N is 12, L 012 represents the cost score of the second decoding path of the two first decoding paths from node 0 to node 1, W 011 Represents the weight of the second decoding path from node 0 to node 1.

[0121] When N is 2, L 02 represents the cost score of the first decoding path from node 0 to node 2, W 02 Represents the weight of the first decoding path from node 0 to node 2.

[0122] When N is 9, L 09 represents the cost score of the first decoding path from node 0 to node 9, W 09 Represents the weight of the first decoding path from node 0 to node 9.

[0123] From the above formula (1), it can be seen that when calculating the cost score of the first decoding path, the weights w of multiple decoding paths from node 0 to its downstream nodes are used. 0N And the acoustic score h1 of the first acoustic feature are used to calculate the cost score of each decoding path in multiple decoding paths from node 0 to the downstream node.

[0124] In a second aspect, for a non-first decoding path among multiple decoding paths, a cost score of the non-first decoding path is obtained according to the cost score of the previous decoding path, the acoustic score of the non-first decoding path and the weight of the non-first decoding path.

[0125] For the cost score of the non-first decoding path, the cost score of the non-first decoding path can be obtained according to the cost score of the previous decoding path, the acoustic score corresponding to the non-first decoding path and the weight of the non-first decoding path. The calculation formula is as follows:

[0126] L ij =L ki -log(h x )+w ij (2)

[0127] L ij represents the cost score of the non-first decoding path, i represents the previous node of the non-first decoding path, J represents the next node of the non-first decoding path, and L ki Indicates the cost score of the previous decoding path. For example, when the non-first decoding path is the second decoding path, L ki It should be the cost score of the first decoding path; when the non-first decoding path is the third decoding path, L ki It should be the cost score of the previous non-first decoding path; h x represents the acoustic score of the xth acoustic feature. The order of the acoustic scores is the same as the order of the decoding paths. For example, when calculating the cost score of the first decoding path, the acoustic score of the first acoustic feature should be used. When calculating the cost score of the second decoding path, the acoustic score of the second acoustic feature should be used, and so on. ij is the weight of the non-first decoding path.

[0128] From the above formula (2), it can be seen that when calculating the cost score of the non-first decoding path, taking the second decoding path as an example, it is based on the acoustic score h of the second acoustic feature. x , the cost score L of the previous decoding path ki and the weight w of the second decoding path ij To obtain the cost score of the second decoding path, the cost scores of the remaining third decoding path, fourth decoding path and other non-first decoding paths are calculated according to formula (2).

[0129] Finally, after calculation by formula (1) and formula (2), the cost score of each decoding path in the decoding graph is obtained. The larger the acoustic score of the acoustic feature, the smaller the cost score of the obtained decoding path. The cost score of the previous decoding path is directly proportional to the cost score of the next decoding path, and the acoustic score of the next decoding path is inversely proportional to the cost score of the next decoding path. In this way, when the cost score of the previous decoding path is higher and the acoustic score of the next decoding path is lower, the cost score of the next decoding path will be higher than the cost score of the previous decoding path; when the cost score of the previous decoding path is lower and the acoustic score of the next decoding path is higher, the cost score of the next decoding path will be lower than the cost score of the previous decoding path.

[0130] Sub-step A2: Filter out the decoding path corresponding to the minimum cost score from the multiple decoding paths as the optimal decoding path.

[0131] The optimal decoding path with the smallest cost score among the multiple decoding paths is the decoding path with the highest matching degree with the audio to be recognized emitted by the current sound source.

[0132] Each node in the decoding graph has its own judgment conditions, which include: if the label corresponding to the audio frame emitted by the sound source is the same as the input parameter in the judgment condition of the node, the acoustic score corresponding to the decoding path between the node and the next node is also high, and the cost score corresponding to the decoding path is low; if the label corresponding to the audio frame emitted by the sound source is different from the input parameter in the judgment condition of the node, the acoustic score obtained is low, and the corresponding cost score is high.

[0133] For example, the wake-up recognition model can not only output the acoustic score of each acoustic feature, but also output the label corresponding to each audio. For example, when the audio includes the preset wake-up word "Xiao Ai", the corresponding label of the audio is "x, iao, ai, t, ong, x, ue"; when the audio includes the non-preset wake-up word "Have you eaten today", the corresponding label of the audio is "filler"; when the audio includes a similar wake-up word "Wang", the corresponding label of the audio is "filler ue". The label division process of the audio is the same as that of the above-mentioned training samples, which will not be repeated here.

[0134] When the user sends out the audio of "Wang Tongxue", the phonemes similar to the wake-up word "Wang Tongxue" are "w, ang, t, ong, x, ue", and its label is "filler ue". 012 When decoding the path, due to the filler in the filler ue label and L 012 The input parameter filler of the decoding path is the same, so L 012 This decoding path corresponds to a higher acoustic score, so L 012 The cost score of this decoding path will be lower (eg, 2).

[0135] When the label goes L 09 When decoding the path, due to the difference between ue and L in the label filler ue 09 The input parameter ue of the decoding path is the same, so L 09 The cost score of this decoding path will also be lower (eg, 1).

[0136] When the label goes L 08 When the decoding path is L, since the filler ue is consistent with the decoding path L 02 , L 23 , L 34 , L 45 , L 56 , L 67 The input parameters are different, so L 02 , L 23 , L 34 , L 45 , L 56 , L 67 The acoustic scores corresponding to these decoding paths are low. After multiple rounds of superposition, the cost score of the next decoding path is higher than the cost score of the previous decoding path, which eventually makes L 67 The cost score is higher, and although similar labels ue and L 78 The input parameter ue of the decoding path is the same, and L 78 The decoding path with a lower acoustic score, but the previous decoding path with a larger value L 67 The cost score minus the acoustic score gives L 78 The cost score of the decoding path is still large (e.g. L 67 The cost score of the cost score is 50, minus L 78 When the acoustic score of the decoding path is 0.1, the obtained L 78 The decoding path has a higher cost score of 49.9), so L 08 The cost scores of multiple decoding paths included in the decoding path are all high.

[0137] On this basis, we can select L with the lowest cost score among multiple decoding paths. 09 The decoding path is taken as the optimal decoding path.

[0138] For related technologies, see Figure 8 As shown in the figure, the decoding graph includes 9 nodes 0-8, node 0 is the starting node, and nodes 1 and 8 are the ending nodes. When the user sends the audio of "Mr. Wang", the label corresponding to "Mr. Wang" is "w, ang, t, ong, x, ue". 01 When decoding the path, since the label is the same as L 01 The input parameters of L are different, so 01 The decoding path has a higher cost score; the "t, ong, x, ue" in the label is similar to L 45 , L 56 , L 67 , L 78 The input parameters of the decoding paths are the same, so L 78 The cost score corresponding to the decoding path is low. At this time, the optimal decoding path with the lowest cost score in the entire decoding graph is L 78 , since the decoding path L 78 It is the preset wake-up path, so it may wake up the device by mistake.

[0139] Compared with the related art, the present disclosure adds an additional decoding path L 09 ( Figure 7 The part in the black box), after the wake-up word recognition model is trained by positive samples, negative samples and similar samples, and has the ability to recognize similar wake-up words, preset wake-up words and non-preset wake-up words, that is, the wake-up word recognition model will output different labels based on words of different voice types such as similar wake-up words, preset wake-up words and non-preset wake-up words; then, based on the label and acoustic score of the audio output by the wake-up word recognition model, it is determined whether the optimal decoding path corresponding to the audio emitted by the user is a similar wake-up path. When the optimal decoding path is a similar wake-up path, the device can be controlled not to wake up, thereby further reducing the probability of the device being mistakenly awakened by similar wake-up words.

[0140] For example, when a user sends an audio of "Wang Xuexue", the tag corresponding to "Wang Xuexue" is "filler ue", which is L 09 The cost score of the decoding path is compared to that of L 78 The decoding path has a lower cost score, so the decoding path with the lowest cost score in the entire decoding graph is L 09 , the decoding path L 09 It is a similar wake-up path, so it will not wake up the device.

[0141] It can be understood that the weights corresponding to similar awakening paths (e.g. Figure 7 10) is greater than the weight corresponding to the preset wake-up path (e.g. Figure 7 5) in is higher. The purpose of this design is: if the weight of the similar wake-up path is set to be the same as the weight of the preset wake-up path, then the final cost score of the similar wake-up path is close to the cost score of the preset wake-up path, which may easily cause the device to be awakened when it should not be awakened; if the weight of the similar wake-up path is set to be smaller than the weight of the preset wake-up path, when the acoustic scores of the preset wake-up path and the similar wake-up path are the same, since the weight of the similar wake-up path is set smaller than the weight of the preset wake-up path, the cost score of the similar wake-up path is also smaller than the cost score of the preset wake-up path, and then the device is less likely to be awakened, resulting in a lower wake-up rate of the device; therefore, the weight of the similar wake-up path is set higher, so that when the user says a similar wake-up word that is more in line with expectations, the acoustic score of the similar wake-up path can be lower, and then the overall cost score of the similar wake-up path can be lower, and in this case the device will not be awakened.

[0142] In step S25, if the optimal decoding path is consistent with the preset wake-up path corresponding to the preset wake-up word, it is determined that the audio to be recognized contains the preset wake-up word, and the device is controlled to be in a wake-up state.

[0143] The preset wake-up path can be as follows Figure 7 L 02 Decoding path, L 23 Decoding path, L 34 Decoding path, L 45 Decoding path, L 56 Decoding path, L 67 Decoding path, L 78 Decoding path.

[0144] For example, when a user sends out an audio of "Xiao Ai Tongxue", the preset wake-up word "Xiao Ai Tongxue" has the phonemes "x, iao, ai, t, ong, x, ue", and its label is "x, iao, ai, t, ong, x, ue". 011 Decoding path, L 012 When decoding the path, due to the labels x, iao, ai, t, ong, x, ue” and L 012 The input parameter filler of the decoding path is different, so L 012 This decoding path corresponds to a lower acoustic score, so L 012 This decoding path will have a higher cost score.

[0145] When the label goes L 09 When decoding the path, since only "ue" and L are found in the labels "x、iao、ai、t、ong、x、ue" 09 The input parameter ue of the decoding path is the same, so L 09 The cost score of this decoding path will also be relatively high.

[0146] When the label goes L 08 When the decoding path is L, since x, iao, ai, t, ong, x, ue" are respectively 02 , L 23 , L 34 , L 45 , L 56 , L 67 , L 78 The input parameters are the same, so L 02 , L 23 , L 34 , L 45 , L 56 , L 67 , L 78 The acoustic scores corresponding to these decoding paths are relatively high. After multiple rounds of superposition, for each decoding path, the cost score of the next decoding path is lower than the cost score of the previous decoding path, which eventually makes L 78 The cost score of L is the lowest among multiple decoding paths, so 78 The decoding path has a lower cost score.

[0147] On this basis, we can select the L with the lowest cost score among multiple decoding paths. 78 The decoding path is taken as the optimal decoding path, and L 78 The decoding path is the preset wake-up path, so the control device switches to the wake-up state.

[0148] like Figure 7 As shown, Figure 7 contains the third terminal node 8. When the terminal node of the decoding path with the smallest cost score is the third terminal node 8, it means that the decoding path with the smallest cost score is L 78 Decoding path, at this time, the device can be controlled to switch to the awake state.

[0149] In step S26, if the optimal decoding path is consistent with the non-wake-up path corresponding to the non-preset wake-up word, it is determined that the audio to be recognized contains the non-preset wake-up word, and the device is controlled to be in a non-wake-up state.

[0150] The non-wake-up path can be as follows Figure 7 L 011 Decoding path, L 012 Decoding path.

[0151] For example, when a user sends an audio message "Have you eaten today?", the phonemes of the non-preset wake-up word "hum hum" are "h, eng, h, eng", and its label is "filler". 011 Decoding path, L 012 When decoding the path, due to the label filler and L 012 The input parameter filler of the decoding path is the same, so L 012 This decoding path corresponds to a higher acoustic score, so L 012 This decoding path will have a lower cost score.

[0152] When the label goes L 09 When decoding the path, due to the label filler and L 09 The input parameter ue of the decoding path is different, so L 09 This decoding path will have a higher cost score.

[0153] When the label goes L 08 When the decoding path is L, since the label filler is consistent with the decoding path L 02 , L 23 , L 34 , L 45 , L 56 , L 67 , L 78 The input parameters are different, so L 02 , L 23 , L 34 , L 45 , L 56 , L 67 , L 78 The acoustic scores corresponding to these decoding paths are low. After multiple rounds of superposition, the cost score of the next decoding path is higher than the cost score of the previous decoding path, which eventually makes L 78 The cost score of L is the highest among multiple decoding paths, so 08 The cost scores of multiple decoding paths included in the decoding path are all high.

[0154] On this basis, we can select the L with the lowest cost score among multiple decoding paths. 012 The decoding path is taken as the optimal decoding path, and L 012 The decoding path is a non-awake path, so the control device remains in a non-awake state.

[0155] like Figure 7 As shown, Figure 7 contains the second terminal node 1. When the terminal node of the decoding path with the smallest cost score is the second terminal node 1, it means that the decoding path with the smallest cost score is L 011 Decoding path or L 012 Decoding path, at this time, the device can be controlled to remain in a non-awakening state.

[0156] In step S27, if the optimal decoding path is consistent with the similar wake-up path corresponding to the similar wake-up word, it is determined that the audio to be recognized contains the similar wake-up word, and the device is controlled to be in a non-wake-up state.

[0157] Similar wake-up paths can be as follows Figure 7 L 09 Decoding path.

[0158] Continuing with the example in substep A2, the optimal decoding path with the smallest cost score is L 09 If the decoding path is incorrect, it means that the audio currently sent by the user contains a similar wake-up word, and the device will be controlled to be in a non-awakening state.

[0159] See also Figure 7 As shown in the figure, the weight of the preset wake-up path is greater than that of the non-wake-up path. The purpose of this design is: Combining the above formulas (1) and (2), it can be seen that when the weight of the preset wake-up path increases, if you want to get a lower cost score, you need a higher acoustic score, and the user needs to send out more accurate audio of the preset wake-up word to get a higher acoustic score, thereby lowering the overall cost score. Therefore, setting the weight larger can ensure that the user will wake up the device only when sending out accurate audio of the preset wake-up word, thereby improving the accuracy of device wake-up.

[0160] In the related art, a variety of thresholds are set for the device in different usage scenarios to minimize the problem of false wake-ups. For example, when the device is detected to be in use, the normal threshold is used; when the device is detected to be idle, the device automatically starts a timer, and when the timer reaches the set time, the normal threshold is increased to the wake-up threshold, thereby reducing the number of times the device is awakened. When the number of times the device is awakened is reduced, the false wake-up rate is also reduced accordingly.

[0161] However, this solution can only reduce the false wake-up rate of the device when the device is in idle state, and cannot cover all false wake-up scenarios. In fact, the false wake-up rate is more easily triggered in the scenario when the device is in use. For example, when the device is a TV and the TV is in use and playing a TV series, a similar wake-up word "Mr. Wang" may be played in the TV series, causing the TV to be mistakenly awakened and reply to the user's voice "I'm here", which undoubtedly reduces the user experience.

[0162] In the present disclosure, if the audio of the sound source is recognized as a similar wake-up word, the device will not be controlled to wake up, which covers all false wake-up scenarios of the device, thereby reducing the probability of the device being falsely awakened in use and idle states.

[0163] Through the above technical solution, an additional similar wake-up path is added to the decoding diagram. When the optimal decoding path with the smallest cost score is the similar wake-up path, the device can also be controlled to be in a non-awakening state, further reducing the probability of the device being woken up by mistake.

[0164] Fig. 9 is a block diagram of a device control apparatus according to an exemplary embodiment. Fig. 9 The device control apparatus 900 includes: an acquisition module 920 , an identification module 930 and a control module 940 .

[0165] An acquisition module 920 is configured to acquire audio to be recognized, where the audio to be recognized includes voice information;

[0166] The recognition module 930 is configured to recognize the audio to be recognized based on a preset wake-up word recognition model, and determine whether the audio to be recognized contains a preset wake-up word; wherein the wake-up word recognition model is trained based on training samples by a machine learning method, and the training samples include a positive sample set and a negative sample set, the positive sample set includes multiple positive samples, each of the positive samples includes a positive audio and a positive label, the positive audio is a voice audio of a preset wake-up word, and the positive label includes symbols corresponding to all phonemes in the preset wake-up word; the negative sample set includes multiple negative samples, each of the negative samples includes a negative audio and a negative label, the negative audio is a voice audio that does not contain any phonemes in the preset wake-up word, and the negative label uses a preset symbol;

[0167] The control module 940 is configured to control the device to enter a wake-up state if the audio to be recognized contains the preset wake-up word, otherwise control the device to remain in a non-wake-up state.

[0168] Optionally, the training sample also includes a similar sample set, which includes multiple similar samples, each similar sample includes similar audio and similar labels, the similar audio refers to speech audio containing some phonemes in the preset wake-up word, and the similar label contains the preset symbol and the symbol in the similar audio that has the same phonemes as the preset wake-up word.

[0169] Optionally, the identification module 930 includes:

[0170] A segmentation submodule is configured to segment the audio to be recognized to obtain multiple audio frames in time series;

[0171] An extraction submodule is configured to extract acoustic features from each of the audio frames to obtain acoustic features corresponding to each audio frame;

[0172] An input submodule, configured to input the acoustic features corresponding to the plurality of audio frames into the wake-up word recognition model in sequence, and obtain an acoustic score corresponding to each of the acoustic features; wherein the acoustic score includes a plurality of probability values, each probability value indicating the probability that the acoustic feature is a phoneme corresponding to the preset wake-up word;

[0173] The optimal decoding path determination submodule is configured to process the acoustic features of each audio frame according to the acoustic score and the preset decoding graph to obtain the optimal decoding path corresponding to the audio to be recognized;

[0174] The recognition submodule is configured to determine that the audio to be recognized contains a preset wake-up word if the optimal decoding path is consistent with the preset wake-up path corresponding to the preset wake-up word.

[0175] Optionally, the decoding graph includes multiple nodes, at least two of the nodes constitute any one of the decoding paths among the preset wake-up path, the non-wake-up path and the similar wake-up path, the preset wake-up path is a decoding path indicating that the device enters the wake-up state, and the non-wake-up path and the similar wake-up path are decoding paths indicating that the device enters the non-wake-up state.

[0176] Optionally, the optimal decoding path determination submodule includes:

[0177] A cost score calculation submodule is configured to obtain a cost score of each audio frame passing through a corresponding decoding path in the decoding graph according to the acoustic score of each audio frame;

[0178] The screening submodule is configured to screen out a decoding path corresponding to a minimum cost score from the plurality of decoding paths as the optimal decoding path.

[0179] Optionally, the plurality of decoding paths include a first decoding path, and the cost score calculation submodule includes:

[0180] The first decoding path submodule is configured to obtain a cost score of the first decoding path according to the acoustic score corresponding to the first decoding path and the weight of the first decoding path.

[0181] Optionally, the plurality of decoding paths include a non-first decoding path, and the cost score calculation submodule includes:

[0182] The non-first decoding path submodule is configured to obtain the cost score of the non-first decoding path according to the cost score of the previous decoding path, the acoustic score of the non-first decoding path and the weight of the non-first decoding path.

[0183] Optionally, all the anti-labels in the anti-sample set use the same preset symbol.

[0184] Optionally, the number of phonemes that are the same between the similar audio contained in the similarity tag and the preset wake-up word is N, and N is greater than or equal to 1.

[0185] Regarding the device in the above embodiment, the specific manner in which each module performs operations has been described in detail in the embodiment of the method, and will not be elaborated here.

[0186] The present disclosure also provides a computer-readable storage medium having computer program instructions stored thereon, and the program instructions, when executed by a processor, implement the steps of the device control method provided by the present disclosure.

[0187] Fig.10 1 is a block diagram of a device 1000 according to an exemplary embodiment. For example, the device 1000 may be a speaker, a robot, a mobile phone, a television, a microwave oven, a computer, a digital broadcast terminal, a messaging device, a game console, a tablet device, a medical device, a fitness device, a personal digital assistant, etc.

[0188] Reference Fig.10 , device 1000 may include one or more of the following components: a processing component 1002 , a memory 1004 , a power component 1006 , a multimedia component 1008 , an audio component 1010 , an input / output interface 1012 , a sensor component 1014 , and a communication component 1016 .

[0189] The processing component 1002 generally controls the overall operation of the device 1000, such as operations associated with display, phone calls, data communications, camera operations, and recording operations. The processing component 1002 may include one or more processors 1020 to execute instructions to complete all or part of the steps of the above-mentioned device control method. In addition, the processing component 1002 may include one or more modules to facilitate the interaction between the processing component 1002 and other components. For example, the processing component 1002 may include a multimedia module to facilitate the interaction between the multimedia component 1008 and the processing component 1002.

[0190] The memory 1004 is configured to store various types of data to support the operation of the device 1000. Examples of such data include instructions for any application or method operating on the device 1000, contact data, phone book data, messages, pictures, videos, etc. The memory 1004 can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic memory, flash memory, magnetic disk or optical disk.

[0191] The power supply component 1006 provides power to the various components of the device 1000. The power supply component 1006 may include a power management system, one or more power supplies, and other components associated with generating, managing, and distributing power for the device 1000.

[0192] The multimedia component 1008 includes a screen that provides an output interface between the device 1000 and the user. In some embodiments, the screen may include a liquid crystal display (LCD) and a touch panel (TP). If the screen includes a touch panel, the screen may be implemented as a touch screen to receive input signals from the user. The touch panel includes one or more touch sensors to sense touch, slide, and gestures on the touch panel. The touch sensor may not only sense the boundaries of the touch or slide action, but also detect the duration and pressure associated with the touch or slide operation. In some embodiments, the multimedia component 1008 includes a front camera and / or a rear camera. When the device 1000 is in an operating mode, such as a shooting mode or a video mode, the front camera and / or the rear camera may receive external multimedia data. Each front camera and the rear camera may be a fixed optical lens system or have a focal length and optical zoom capability.

[0193] The audio component 1010 is configured to output and / or input audio signals. For example, the audio component 1010 includes a microphone (MIC), and when the device 1000 is in an operating mode, such as a call mode, a recording mode, and a speech recognition mode, the microphone is configured to receive an external audio signal. The received audio signal can be further stored in the memory 1004 or sent via the communication component 1016. In some embodiments, the audio component 1010 also includes a speaker for outputting audio signals.

[0194] The input / output interface 1012 provides an interface between the processing component 1002 and the peripheral interface modules, which may be keyboards, click wheels, buttons, etc. These buttons may include but are not limited to: a home button, a volume button, a start button, and a lock button.

[0195] The sensor assembly 1014 includes one or more sensors for providing various aspects of status assessment for the device 1000. For example, the sensor assembly 1014 can detect the open / closed state of the device 1000, the relative positioning of components, such as the display and keypad of the device 1000, and the sensor assembly 1014 can also detect the position change of the device 1000 or a component of the device 1000, the presence or absence of user contact with the device 1000, the orientation or acceleration / deceleration of the device 1000, and the temperature change of the device 1000. The sensor assembly 1014 may include a proximity sensor configured to detect the presence of a nearby object without any physical contact. The sensor assembly 1014 may also include an optical sensor, such as a CMOS or CCD image sensor, for use in imaging applications. In some embodiments, the sensor assembly 1014 may also include an accelerometer, a gyroscope sensor, a magnetic sensor, a pressure sensor, or a temperature sensor.

[0196] The communication component 1016 is configured to facilitate wired or wireless communication between the device 1000 and other devices. The device 1000 can access a wireless network based on a communication standard, such as WiFi, 2G or 3G, or a combination thereof. In an exemplary embodiment, the communication component 1016 receives a broadcast signal or broadcast-related information from an external broadcast management system via a broadcast channel. In an exemplary embodiment, the communication component 1016 also includes a near field communication (NFC) module to facilitate short-range communication. For example, the NFC module can be implemented based on radio frequency identification (RFID) technology, infrared data association (IrDA) technology, ultra-wideband (UWB) technology, Bluetooth (BT) technology and other technologies.

[0197] In an exemplary embodiment, the device 1000 may be implemented by one or more application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field programmable gate arrays (FPGAs), controllers, microcontrollers, microprocessors, or other electronic components to perform the above-described device control method.

[0198] In an exemplary embodiment, a non-transitory computer-readable storage medium including instructions is also provided, such as a memory 1004 including instructions, and the instructions can be executed by the processor 1020 of the device 1000 to complete the above device control method. For example, the non-transitory computer-readable storage medium can be a ROM, a random access memory (RAM), a CD-ROM, a magnetic tape, a floppy disk, an optical data storage device, etc.

[0199] Those skilled in the art will readily appreciate other embodiments of the present disclosure after considering the specification and practicing the present disclosure. The present disclosure is intended to cover any variations, uses or adaptations of the present disclosure that follow the general principles of the present disclosure and include common knowledge or customary techniques in the art that are not disclosed in the present disclosure. The description and examples are to be considered as exemplary only, and the true scope and spirit of the present disclosure are indicated by the following claims.

[0200] It should be understood that the present disclosure is not limited to the exact structures that have been described above and shown in the drawings, and that various modifications and changes may be made without departing from the scope thereof. The scope of the present disclosure is limited only by the appended claims.< / eps> < / eps>

Claims

1. A device control method, characterized in that: include: Acquire audio to be recognized, where the audio to be recognized contains voice information; The audio to be recognized is recognized based on a preset wake-up word recognition model to determine whether the audio to be recognized contains a preset wake-up word; wherein the wake-up word recognition model is trained based on training samples by a machine learning method, and the training samples include a positive sample set and a negative sample set, the positive sample set includes multiple positive samples, each of the positive samples includes a positive audio and a positive label, the positive audio is the speech audio of the preset wake-up word, and the positive label includes symbols corresponding to all phonemes in the preset wake-up word; the negative sample set includes multiple negative samples, each of the negative samples includes a negative audio and a negative label, the negative audio is the speech audio that does not contain any phonemes in the preset wake-up word, and the negative label uses a preset symbol; If the audio to be recognized contains the preset wake-up word, the device is controlled to enter a wake-up state; otherwise, the device is controlled to remain in a non-wake-up state.

2. The method according to claim 1, characterized in that The training samples also include a similar sample set, which includes multiple similar samples, each similar sample includes similar audio and similar labels, the similar audio refers to speech audio containing some phonemes in the preset wake-up word, and the similar label includes the preset symbol and the symbol in the similar audio that has the same phonemes as the preset wake-up word.

3. The method according to claim 1 or 2, characterized in that: The step of recognizing the audio to be recognized based on a preset wake-up word recognition model and determining whether the audio to be recognized contains a preset wake-up word includes: Segmenting the audio to be recognized to obtain multiple audio frames in time series; Extracting acoustic features from each of the audio frames to obtain acoustic features corresponding to each audio frame; Inputting the acoustic features corresponding to the plurality of audio frames into the wake-up word recognition model in sequence to obtain an acoustic score corresponding to each of the acoustic features; wherein the acoustic score includes a plurality of probability values, each probability value indicating the probability that the acoustic feature is a phoneme corresponding to the preset wake-up word; Processing the audio frame according to the acoustic score of the acoustic feature and a preset decoding graph to obtain an optimal decoding path corresponding to the audio to be recognized; If the optimal decoding path is consistent with the preset wake-up path corresponding to the preset wake-up word, it is determined that the audio to be recognized contains the preset wake-up word.

4. According to the method according to claim 3, the decoding graph includes multiple nodes, at least two of the nodes constitute any one of the decoding paths among the preset wake-up path, the non-wake-up path and the similar wake-up path, the preset wake-up path is a decoding path indicating that the device enters the wake-up state, and the non-wake-up path and the similar wake-up path are decoding paths indicating that the device enters the non-wake-up state.

5. The method according to claim 3, characterized in that: The step of processing according to the acoustic score of the acoustic feature of each audio frame and a preset decoding graph to obtain an optimal decoding path corresponding to each audio frame includes: Obtaining, according to the acoustic score of each audio frame, a cost score of a decoding path corresponding to each audio frame in the decoding graph; From the multiple decoding paths, a decoding path corresponding to the minimum cost score is selected as the optimal decoding path.

6. The method according to claim 5, characterized in that The multiple decoding paths include a first decoding path, and obtaining, according to the acoustic score of each audio frame, a cost score of each audio frame passing through a corresponding decoding path in the decoding graph, includes: A cost score of the first decoding path is obtained according to the acoustic score corresponding to the first decoding path and the weight of the first decoding path.

7. The method according to claim 5, characterized in that The multiple decoding paths include a non-first decoding path, and obtaining, according to the acoustic score of each audio frame, a cost score of each audio frame passing through a corresponding decoding path in the decoding graph, includes: The cost score of the non-first decoding path is obtained according to the cost score of the previous decoding path, the acoustic score of the non-first decoding path and the weight of the non-first decoding path.

8. The method according to claim 1, characterized in that: All the anti-labels in the anti-sample set use the same preset symbol.

9. The method according to claim 2, characterized in that: The number of phonemes that are identical between the similar audio contained in the similarity tag and the preset wake-up word is N, and N is greater than or equal to 1.

10. A device control device, characterized in that: include: An acquisition module is configured to acquire audio to be recognized, where the audio to be recognized contains voice information; A recognition module is configured to recognize the audio to be recognized based on a preset wake-up word recognition model, and determine whether the audio to be recognized contains a preset wake-up word; wherein the wake-up word recognition model is trained based on training samples by a machine learning method, and the training samples include a positive sample set and a negative sample set, the positive sample set includes multiple positive samples, each of the positive samples includes a positive audio and a positive label, the positive audio is a voice audio of a preset wake-up word, and the positive label includes symbols corresponding to all phonemes in the preset wake-up word; the negative sample set includes multiple negative samples, each of the negative samples includes a negative audio and a negative label, the negative audio is a voice audio that does not contain any phonemes in the preset wake-up word, and the negative label uses a preset symbol; The control module is configured to control the device to enter a wake-up state if the audio to be recognized contains the preset wake-up word, otherwise control the device to remain in a non-wake-up state.

11. A device, characterized in that: include: processor; a memory for storing processor-executable instructions; Wherein, the processor is configured to: Execute the steps of any one of the methods of claims 1 to 9.

12. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the program instructions are executed by a processor, the steps of the method described in any one of claims 1 to 9 are implemented.

Images