A voice wake-up method, a voice wake-up device, and a storage medium

Abstract

The application discloses a voice wake-up method, a voice wake-up device and a storage medium. The voice wake-up method comprises the following steps: acquiring an acoustic feature flow; extracting a first state sequence of the acoustic feature flow based on a first modeling unit; in response to a decoding result of the first state sequence by a first-level decoding graph being a preset wake-up word, acquiring a decoded acoustic feature of the decoded preset wake-up word; extracting a second state sequence of the decoded acoustic feature based on a second modeling unit; and in response to a decoding result of the second state sequence by a second-level decoding graph being the preset wake-up word, performing a wake-up operation on a voice device associated with the preset wake-up word. The voice wake-up method can perform secondary decoding confirmation on the wake-up word by constructing acoustic state sequence combinations of multiple scales, thereby reducing the occurrence of false wake-up.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of voice interaction, in particular to a voice wake-up method, a voice wake-up device and a storage medium. BACKGROUND

[0002] With the rapid development of intelligent voice interaction technology, voice wake-up technology has made great progress in recent years. Users wake up devices by saying a wake-up word in smart home devices, video conference devices or home appliances, and start human-computer voice interaction with the devices. Therefore, voice wake-up is an important part of voice interaction.

[0003] A commonly used voice wake-up method is to perform decoding search on a feature sequence obtained by feature extraction on continuous input stream data, to determine whether there is a sequence with a larger likelihood than a wake-up word feature sequence, and to wake up the device when the likelihood exceeds a set threshold. However, a single acoustic model and a single decoding graph cannot avoid the occurrence of false wake-up. SUMMARY

[0004] The present application provides a voice wake-up method, a voice wake-up device and a storage medium.

[0005] One of the technical solutions adopted by the present application is to provide a voice wake-up method, which comprises:

[0006] obtaining an acoustic feature stream;

[0007] extracting a first state sequence of the acoustic feature stream based on a first modeling unit, wherein the first state sequence is a labeled sequence corresponding to the first modeling unit;

[0008] in response to a decoding result of the first state sequence by a first-level decoding graph being a preset wake-up word, obtaining a decoded acoustic feature of the preset wake-up word;

[0009] extracting a second state sequence of the decoded acoustic feature based on a second modeling unit, wherein the second state sequence is a labeled sequence corresponding to the second modeling unit, and the second modeling unit and the first modeling unit are modeling units of different levels;

[0010] in response to a decoding result of the second state sequence by a second-level decoding graph being the preset wake-up word, performing a wake-up operation on a voice device associated with the preset wake-up word.

[0011] Wherein, the length of the decoded acoustic feature of the preset wake-up word is the length of the weighted undirected graph corresponding to the preset wake-up word. Wherein, the decoding result of the first state sequence corresponding to the first-level decoding graph further comprises a first path likelihood value, and the decoding result of the second state sequence corresponding to the second-level decoding graph further comprises a second path likelihood value.

[0012] in response to the decoding result of the second state sequence by the second decoding graph being the preset wake-up word, performing a wake-up operation on a voice device associated with the preset wake-up word, including:

[0013] in response to the decoding result of the second state sequence by the second decoding graph being the preset wake-up word, and the absolute value of the difference between the first path likelihood value and the second path likelihood value being less than a preset threshold, performing a wake-up operation on a voice device associated with the preset wake-up word.

[0014] The preset condition is that the absolute value of the difference between the first path likelihood value and the second path likelihood value is less than a preset threshold.

[0015] The modeling unit includes a word, a syllable, an initial, and a final.

[0016] Before the first state sequence of the acoustic feature stream is extracted based on the first modeling unit, the voice wake-up method further includes:

[0017] Obtaining training data and extracting acoustic features of the training data;

[0018] Based on the acoustic features, obtain an acoustic feature sequence and an acoustic annotation sequence, wherein the acoustic annotation sequence is an annotation sequence of a preset modeling unit;

[0019] Obtain the mutual information of the acoustic feature sequence and the acoustic annotation sequence;

[0020] Train the first modeling unit using the mutual information to obtain the first modeling unit.

[0021] The acoustic model is trained using the mutual information to obtain the acoustic model corresponding to the preset modeling unit, including:

[0022] Using the mutual information, a target training function is constructed;

[0023] Train the acoustic model in the direction of the target function value of the target training function becoming larger and larger until the target function value of the target training function reaches a preset threshold, and obtain the first modeling unit.

[0024] Before the first state sequence of the acoustic feature stream is extracted based on the first modeling unit, the voice wake-up method further includes:

[0025] A first decoding graph is established using a first modeling unit, a context-dependent model, a pronunciation dictionary model, and a language model.

[0026] The voice wake-up method further includes, before the extracting the acoustic features of the to-be-trained data:

[0027] The data enhancement processing includes reverberation, noise, speed disturbance, and / or volume disturbance.

[0028] Another technical solution adopted by the present application is to provide a voice wake-up device, which includes a processor and a memory connected to the processor; wherein the memory is used to store program instructions, and the processor is used to execute the program instructions to implement the voice wake-up method as described above.

[0029] Another technical solution adopted by the present application is to provide a computer storage medium, which is used to store program instructions, and the program instructions are used to implement the voice wake-up method as described above when executed by a computer.

[0030] The beneficial effects of the present application are: obtaining an acoustic feature stream; extracting a first state sequence of the acoustic feature stream based on a first modeling unit; in response to a decoding result of the first state sequence by a first-level decoding graph being a preset wake-up word, obtaining a decoded acoustic feature of the decoded preset wake-up word; extracting a second state sequence of the decoded acoustic feature based on a second modeling unit; in response to a decoding result of the second state sequence by a second-level decoding graph being the preset wake-up word, performing a wake-up operation on a voice device associated with the preset wake-up word. The voice wake-up method of the present application can perform secondary decoding confirmation on the wake-up word through the construction of acoustic state sequence combinations of multiple scales and the determination of path likelihood values, thereby reducing the occurrence of false wake-up. BRIEF DESCRIPTION OF DRAWINGS

[0031] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings needed in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0032] Figure 1 is a flowchart of an embodiment of the voice wake-up method provided by the present application;

[0033] Figure 2 is Figure 1 is a flowchart of the step before S2 in

[0034] Figure 3 is Figure 2 is a flowchart of the sub-step of S15 in

[0035] Figure 4is a flowchart of another embodiment of the voice wake-up method provided in the present application;

[0036] Figure 5 is a structural diagram of an embodiment of the decoding method provided in the present application; Figure 1

[0037] Figure 6 is a flowchart of another embodiment of the voice wake-up method provided in the present application;

[0038] Figure 7 is a structural diagram of an embodiment of the voice wake-up device provided in the present application;

[0039] Figure 8 is a structural diagram of an embodiment of the computer storage medium provided in the present application. DETAILED DESCRIPTION

[0040] The technical solutions in the embodiments of the present application will be clearly and completely described in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative work, fall within the scope of protection of the present application.

[0041] The steps in the embodiments of the present application do not have to be processed in the described order of steps, and the steps can be selectively rearranged, deleted or added according to requirements. The step description in the embodiments of the present application is only an optional order combination, and does not represent all order combinations of the embodiments of the present application. The order of steps in the embodiments cannot be considered as a limitation of the present application.

[0042] The terms "first", "second", and the like in the present application are used to distinguish different objects, rather than to describe a specific order. In addition, the terms "include" and "have" and any variations thereof are intended to cover non-exclusive inclusion. For example, a process, method, system, product or device that includes a series of steps or units is not limited to the listed steps or units, but can optionally include steps or units not listed, or can optionally include other steps or units inherent to the process, method, product or device.

[0043] A Hidden Markov Model (HMM) based on a statistical model is a commonly used speech recognition method. The HMM is an iterative training model, and the iterative training can make the established recognition model have strong adaptability to the training data set. Moreover, the HMM is a state number adjustable model, and can effectively model the dynamic characteristic speech data. ​

[0044] The application mainly designs a voice wake-up method for reducing false wake-up. Unlike the traditional method, the application starts from the perspective of acoustic state sequence, reduces false recognition of non-wake-up words by establishing acoustic state sequence combinations of multiple modeling scales.

[0045] For details, please refer to Figure 1 , Figure 1 is a flowchart of an embodiment of the voice wake-up method provided by the application.

[0046] As Figure 1 indicated, the voice wake-up method provided by the application can include the following steps:

[0047] S1, obtaining an acoustic feature stream.

[0048] In an embodiment, the voice wake-up device of the smart home device, video conference device or home appliance device obtains a voice signal, and extracts an acoustic feature stream from the voice signal.

[0049] S2, extracting a first state sequence of the acoustic feature stream based on a first modeling unit.

[0050] The first state sequence is a label sequence corresponding to the first modeling unit.

[0051] In some embodiments, the voice wake-up device extracts the first state sequence from the obtained acoustic feature stream based on the first modeling unit. According to the first modeling unit, the acoustic feature stream is forwardly inferred, and other models such as the following related models, pronunciation dictionary models and language models are used to extract features from the acoustic feature stream to obtain the first state sequence.

[0052] The first-level decoding graph is a WFST decoding graph.

[0053] S3, in response to the decoding result of the first state sequence by the first-level decoding graph being a preset wake-up word, obtaining a decoded acoustic feature of the decoded preset wake-up word.

[0054] The decoding result of the first state sequence by the first-level decoding graph further includes a first path likelihood value.

[0055] In some embodiments, when the decoding result of the first state sequence by the first decoding graph is a wake-up word, the voice wake-up device obtains the decoded acoustic feature corresponding to the preset wake-up word.

[0056] The length of the decoded acoustic feature of the preset wake-up word is the length of the weighted undirected graph corresponding to the preset wake-up word.

[0057] Optionally, the weighted undirected graph length can be a Lattice length, and the Lattice is a Directed Acyclic Graph (DAG) used to save active candidate path information in a speech recognition process. Each edge represents a candidate, for example, each edge in a syllable Lattice represents a syllable candidate, and each edge in a word Lattice represents a word candidate. Each edge can represent not only a candidate but also an adjacency relationship between candidates.

[0058] The weighted undirected graph length can be a path with the highest score in all paths in the first decoding graph, or a path that meets a preset rule in the first decoding graph.

[0059] Optionally, the preset rule can be a Viterbi algorithm.

[0060] S4, extracting a second state sequence of the decoded acoustic feature based on the second modeling unit.

[0061] The second state sequence is a label sequence corresponding to the second modeling unit, and the second modeling unit and the first modeling unit are modeling units of different levels.

[0062] In some embodiments, the voice wake-up device extracts a second state sequence from the decoded acoustic feature corresponding to the preset wake-up word based on the second modeling unit.

[0063] S5, in response to a decoding result of the second-level decoding graph on the second state sequence being the preset wake-up word, performing a wake-up operation on a voice device associated with the preset wake-up word.

[0064] The second-level decoding graph is a WFST decoding graph.

[0065] The decoding result of the second-level decoding graph on the second state sequence further includes a second path likelihood value.

[0066] In some embodiments, when the decoding result of the second decoding graph on the second state sequence is the wake-up word, the voice wake-up device performs a wake-up operation on a voice device associated with the preset wake-up word.

[0067] The voice wake-up method provided in the present application wakes up through the recognition results of two-level acoustic state sequences. When waking up, the wake-up result is obtained through the joint decision of the two-level acoustic state sequences, thereby reducing the probability of waking up the device with a non-wake-up word.

[0068] The sub-step of S5 can further include:

[0069] S51, in response to the decoding result of the second state sequence of the second-level decoding graph being the preset wake-up word, and the absolute value of the difference between the first path likelihood value and the second path likelihood value being less than a preset threshold, performing a wake-up operation on a voice device associated with the preset wake-up word.

[0070] In some embodiments, if the decoding result output by the second-level decoding graph is a wake-up word, and the following relationship is satisfied, then the decoding result output by the second-level decoding graph is a wake-up word, and the voice wake-up apparatus wakes up the voice device:

[0071] |P score2 -P score1 |<P threshold

[0072] wherein the first path likelihood value is P score1 , the second path likelihood value is P score2 , and P threshold represents a path likelihood value difference threshold.

[0073] If the decoding result output by the second-level decoding graph is not a wake-up word, the voice wake-up apparatus does not wake up the voice device, and re-acquires the acoustic feature stream continuously.

[0074] Figure 2 is Figure 1 a flowchart before the step of S2.

[0075] As Figure 2 shown, the step of S2 can include the following steps:

[0076] S11, performing data enhancement processing on the training data.

[0077] In an embodiment, the voice wake-up apparatus performs data cutting and data enhancement processing on the training data with a labeled category.

[0078] wherein the training data includes positive sample data and negative sample data.

[0079] Optionally, the data enhancement manner includes, but is not limited to, one or more of the following: adding reverberation, adding noise, speed disturbance, and volume disturbance.

[0080] The voice wake-up apparatus enhances the robustness of the first modeling unit generated by training using the training data by performing data enhancement processing on the training data.

[0081] S12, obtaining the training data and extracting acoustic features of the training data.

[0082] In an embodiment, the voice wake-up apparatus extracts features from the training data, transforms a short-time signal to a frequency domain or an inverse frequency domain for representation, and improves the representation ability of the acoustic features.

[0083] The acoustic features can be one or more of MFCC (Mel-Frequency Cepstral Coefficients), FBANK (Filter Bank), PLP (Perceptual Linear Predictive), and PCEN (Per-Channel Energy Normalization), without any limitation here.

[0084] S13, based on acoustic features, obtain acoustic feature sequences and acoustic annotation sequences.

[0085] Among them, the acoustic annotation sequence is the annotation sequence of the preset modeling unit.

[0086] In one embodiment, the acoustic feature sequence can be represented as X = {x1, x2, ..., x...} m The acoustic label sequence can be represented as L = {l1, l2, ..., l}. n}, where the acoustic annotation sequence L can be the annotation of any modeling unit.

[0087] S14, obtain the mutual information between the acoustic feature sequence to be trained and the acoustic annotation sequence of the first modeling unit.

[0088] Mutual information is used to measure the reciprocity between acoustic feature sequences and acoustic annotation sequences.

[0089] In one embodiment, the maximum mutual information criterion (MMI) is used as the mutual information between the acoustic feature sequence and the acoustic annotation.

[0090] The maximum mutual information criterion maximizes the mutual information between the observation sequence X and the word sequence L, and has the following relationship:

[0091]

[0092] In one embodiment, P(L) can be a constant, then:

[0093] MMI = log(L|X)

[0094] Understandably, in this embodiment, MMI maximizes the posterior probability of correctly labeled sentences, which is equivalent to the conditional maximum likelihood criterion.

[0095] S15, use mutual information to train the first modeling unit to obtain the acoustic model corresponding to the preset modeling unit.

[0096] For detailed steps on training the first modeling unit using mutual information, please refer to [link / reference]. Figure 3 , Figure 3is Figure 2 Flowchart of sub-steps of S15.

[0097] As Figure 3 shown, the sub-steps of S15 can include:

[0098] S151, constructing a target training function by using mutual information.

[0099] In an embodiment, the target training function based on the maximum mutual information criterion constructed by the speech wake-up device for multiple speeches can be:

[0100]

[0101] wherein, L u represents the word sequence corresponding to the u-th sentence label, is the HMM state sequence corresponding to L u , P(L) is the language model probability of the sequence L, and L represents other possible word sequences.

[0102] S152, training the first modeling unit in the direction of the target function value of the target training function becoming larger and larger until the target function value of the target training function reaches a preset threshold, to obtain the first modeling unit.

[0103] In an embodiment, the target training function F MMI constructed by the speech wake-up device is: F(L, L) = P(L|L) / P(L), wherein the numerator represents the probability of the correctly labeled word sequence, and the denominator represents the probability of all possible word sequences. The target training function needs to be trained in the direction of the maximum probability of the correctly labeled word sequence and the minimum probability of other possible word sequences, i.e., the target training function is trained in the direction of the function value becoming larger. When the target training function reaches the preset threshold, it can be considered that the target training function reaches a local optimum, the training is completed, and the first modeling unit corresponding to the preset threshold reached by the current target training function is obtained.

[0104] Before performing the S2 step, the following can also be included:

[0105] S16, establishing a first decoding graph by using the first modeling unit, the context-dependent model, the pronunciation dictionary model, and the language model.

[0106] The above embodiments are only one common case of the present application, and do not limit the technical scope of the present application in any way. Therefore, any slight modification, equivalent change, or modification made to the above content according to the essence of the present application still falls within the scope of the technical solution of the present application.

[0107] Please refer to Figure 4 , Figure 4 is a flowchart of another embodiment of the speech wake-up method provided by the present application.

[0108] As Figure 4 shown, the steps of another embodiment of the voice wake-up method provided by the present application can include:

[0109] S51, obtaining voice information.

[0110] In an embodiment, the voice wake-up device obtains sound information in the environment, extracts voice information from the sound information, and filters part of the noise.

[0111] S52, voice information feature extraction.

[0112] In an embodiment, the voice wake-up device extracts features from the obtained voice information to obtain an acoustic feature stream of the voice information.

[0113] Optionally, the acoustic features of the voice information can be any one of MFCC (Mel-Frequency Cepstral Coefficients), FBANK (Filter Bank), PLP (Perceptual Linear Predictive), or PCEN (Per-Channel Energy Normalization), which is not limited here.

[0114] This step is used to transform the short-time signal into the frequency domain or the cepstrum domain for representation, improving the representation ability of the voice information.

[0115] S53, voice multi-scale recognition.

[0116] In an embodiment, the voice wake-up device inputs the obtained acoustic feature stream into a first modeling unit to extract a state sequence corresponding to the acoustic feature stream and perform first-level decoding graph decoding; if the decoding result of the first state sequence by the first decoding graph is a preset wake-up word, the decoding acoustic feature of the decoded preset wake-up word is obtained; a second state sequence of the decoding acoustic feature is extracted based on a second modeling unit, wherein the first state sequence is a labeled sequence corresponding to the second modeling unit, and the second modeling unit and the first modeling unit are modeling units of different levels; if the decoding result of the second state sequence by the second decoding graph is the preset wake-up word, the voice device associated with the preset wake-up word is awakened, and the path likelihood value difference corresponding to the decoding results of the two-level decoding graphs is less than a preset threshold, then jump to S54, otherwise jump back to S51.

[0117] S54, voice wake-up.

[0118] In an embodiment, the voice wake-up device of the smart home device, the video conference device or the home appliance device starts the human-computer voice interaction with the user in response to the voice wake-up keyword of the user.

[0119] The voice wake-up method provided in the application can perform secondary decoding confirmation on the wake-up word through construction of acoustic state sequence combinations of multiple scales and determination of path likelihood values, so that the occurrence of false wake-up is reduced.

[0120] Please refer to Figure 5 , Figure 5 is the decoding Figure 1 structure diagram of the embodiment provided in the application.

[0121] The acoustic model decoding graph provided in the application is a WFST decoding graph.

[0122] The WFST decoding graph is a decoding network for speech recognition based on a weighted finite state transducer (WFST). The WFST decoding graph is composed of four graphs, H, C, L and G. H defines the mapping of HMM states to context-dependent factors; C defines the mapping of context-dependent phonemes to words; L represents a pronunciation dictionary, defining the mapping of factors to words; and G represents a language model, defining the mapping of words to sentences. Among them, H and C are generated by discriminative training in the training process and have been fixed; the state transducers of L and G need to be artificially specified paths.

[0123] In the WFST decoding graph provided in the application, H and C are generated by discriminative training in the training process of the decoding graph and have been fixed; the state transducers of L and G need to be artificially specified paths.

[0124] The WFST decoding graph provided in the application is divided into two parts, keyword HMM and filler HMM. The filler HMM is used to process non-keywords, such as other words or non-speech sounds such as coughing, door closing, breathing and the like. For the keyword HMM, a fine modeling method is adopted, in which each HMM can represent a syllable, a phoneme or a word, which are then combined to form a wake-up word; for the filler HMM, a coarse modeling method is adopted, in which each HMM represents a word or a scene.

[0125] In the WFST decoding graph provided in the application, the L graph is determined by combining wake-up words, silence phonemes and other scene phonemes. For example, as shown below:

[0126] <eps>0

[0127] SIL 1

[0128] <wakeword> 2

[0129] <filler1> 3

[0130] <filler2> 4

[0131] <filler3>5

[0132] …

[0133] <s>6

[0134] < / s> 7

[0135] The G graph in the WFST decoding graph proposed in the present application formulates a finite state cyclic decoding path according to the decoding process. For example, as shown below:

[0136] 0 1SIL SIL

[0137] 1 2 <wakeword> <wakeword>

[0138] 1 3 <filler1> <filler1>

[0139] 1 4 <filler2> <filler2>

[0140] 1 5 <filler3> <filler3>

[0141] 2 1 SIL SIL

[0142] 3 1 SIL SIL

[0143] 4 1 SIL SIL

[0144] 5 1 SIL SIL 0

[0146] In an embodiment, the relationship between the WFST decoding graph and H, C, L, G satisfies:

[0147] N = fact (π ε (min(det(H°C°det(L°G))))

[0148] Wherein, N represents the final WFST decoding graph, ° represents the composition of two graphs, det, min, π ε and fact are determinization, minimization, empty edge removal and simplification operations respectively. Through the above operations, a large decoding graph is compiled, and the decoding process is to seek the optimal path in the large decoding graph.

[0149] In the construction of the decoding graph proposed in the present application, considering the existence of multiple wake-up words and other types of combinations, the L graph and the G graph in the HCLG graph used for decoding are improved by additionally specifying the path to improve the recognition rate of voice wake-up.

[0150] Please refer to Figure 6 , Figure 6 is a flowchart of another embodiment of the voice wake-up method provided by the present application.

[0151] As Figure 6 shown, the voice wake-up method of another embodiment can include:

[0152] S400, acquiring an acoustic feature stream.

[0153] In an embodiment, the voice wake-up device of the smart home device, video conference device or home appliance device acquires a voice signal, and extracts an acoustic feature stream from the voice signal.

[0154] S410, inputting a first-level wake-up model.

[0155] Wherein, the first-level wake-up model corresponds to the first decoding graph described above, which is not described here.

[0156] Wherein, the first-level decoding graph decoding result further includes a first path likelihood value.

[0157] S420, whether to decode to a wake-up word.

[0158] In an embodiment, if the decoding result of the first decoding graph on the first state sequence is the preset wake-up word, the decoding acoustic feature of the decoding of the preset wake-up word is obtained, and the step S430 is jumped to; if the decoding result of the first decoding graph on the first state sequence is not the preset wake-up word, the step S400 is jumped to.

[0159] S430, obtaining the acoustic feature of the current decoding to the length of the weighted undirected graph corresponding to the wake-up word.

[0160] S440, inputting the second wake-up model.

[0161] The second wake-up model corresponds to the second decoding graph described above, and details are not repeated here.

[0162] The decoding result of the second decoding graph further includes a second path likelihood value.

[0163] S450, whether the wake-up word is decoded.

[0164] In an embodiment, if the decoding result of the second decoding graph on the second state sequence is the preset wake-up word, the step S460 is jumped to; if the obtained result is not the wake-up word, the step S400 is jumped to.

[0165] S460, whether the difference between the path likelihood values of the two models is less than a threshold value.

[0166] In an embodiment, if the following conditions are met:

[0167] |P score2 -P score1 |<P threshold

[0168] The first path likelihood value is P score1 , the second path likelihood value is P score2 , and P threshold represents the threshold value of the difference between the path likelihood values.

[0169] In an embodiment, if the difference between the path likelihood values output by the two decoding graphs is less than the threshold value of the difference between the path likelihood values, the step S470 is jumped to; if the difference between the path likelihood values output by the two decoding graphs is greater than the threshold value, the step S400 is jumped to.

[0170] S470, voice wake-up.

[0171] In an embodiment, the voice wake-up device of the smart home device, the video conference device, or the household appliance device starts human-computer voice interaction with the user in response to the voice wake-up keyword of the user.

[0172] The voice wake-up method proposed in the present application can perform secondary decoding confirmation on the wake-up word through the construction of acoustic state sequence combinations of multiple scales and the determination of path likelihood values, thereby reducing the occurrence of false wake-up.

[0173] Please continue to refer to Figure 7 , Figure 7 is a structural schematic diagram of an embodiment of the voice wake-up device provided in the present application. The voice wake-up device 500 of the embodiment of the present application comprises a processor 51 and a memory 52.

[0174] The processor 51 and the memory 52 are respectively connected with a bus. The memory 52 stores program instructions, and the processor 51 is used to execute the program instructions to realize the voice wake-up method described in the above embodiment.

[0175] In the embodiment of the present application, the processor 51 can also be referred to as a CPU (Central Processing Unit). The processor 51 can be an integrated circuit chip with signal processing capability. The processor 51 can also be a general-purpose processor, a digital signal processor (DSP), an application specific integrated circuit (ASIC), a field programmable gate array (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components. The general-purpose processor can be a microprocessor or the processor 51 can also be any conventional processor.

[0176] The voice wake-up method provided in the present application wakes up through the recognition results of the two-level wake-up word models. When waking up, the wake-up result is obtained through the joint decision of the two-level wake-up models, so as to reduce the probability of waking up the device with non-wake-up words.

[0177] The present application also provides a computer storage medium, please continue to refer to Figure 8 , Figure 8 is a structural schematic diagram of an embodiment of the computer storage medium provided in the present application. The computer storage medium 600 stores program instructions 61. The program instructions 61 are used to realize the voice wake-up method of the above embodiment when executed by a processor.

[0178] The embodiments of the present application are realized in the form of software function units and sold or used as independent products, and can be stored in a computer readable storage medium. Based on such understanding, the technical solutions of the present application, essentially or in part, or all or part of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium, and includes several instructions to cause a computer device (which can be a personal computer, a server, or a network device, etc.) or a processor to perform all or part of the steps of the methods described in the various embodiments of the present application. The foregoing storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), magnetic disk or optical disk, and various media that can store program codes.

[0179] The voice wake-up method provided in the present application wakes up through the recognition results of the two-level wake-up word models. When waking up, the wake-up result is obtained through the joint decision of the two-level wake-up models, so as to reduce the probability of waking up the device with non-wake-up words.

[0180] The above is only the embodiment of the present application, and does not limit the patent scope of the present application. The equivalent structure or equivalent flow transformation made by the content of the specification and drawings of the present application, or direct or indirect application in other related technical fields, are also included in the patent protection scope of the present application. < / filler3> < / filler2> < / filler2> < / filler1> < / filler1> < / wakeword> < / wakeword> < / filler2> < / filler1> < / wakeword> < / eps>

Claims

1. A voice wake-up method, characterized in that, The voice wake-up method includes: Acquire acoustic feature flow; The first state sequence of the acoustic feature flow is extracted based on the first modeling unit, wherein the first state sequence is the annotation sequence corresponding to the first modeling unit; In response to the first-level decoding graph decoding the first state sequence as a preset wake word, the decoding acoustic features of the decoded preset wake word are obtained; The second state sequence of the decoded acoustic features is extracted based on the second modeling unit, wherein the second state sequence is the annotation sequence corresponding to the second modeling unit, and the second modeling unit and the first modeling unit are modeling units at different levels; In response to the second-level decoding diagram decoding the second state sequence as the preset wake-up word, a wake-up operation is performed on the voice device associated with the preset wake-up word; The decoding result of the first-level decoding graph corresponding to the first state sequence further includes a first path likelihood value, and the decoding result of the second-level decoding graph corresponding to the second state sequence further includes a second path likelihood value. In response to the second-level decoding diagram decoding the second state sequence as the preset wake-up word, a wake-up operation is performed on the voice device associated with the preset wake-up word, including: In response to the second-level decoding graph decoding the second state sequence as the preset wake-up word, and the absolute value of the difference between the first path likelihood value and the second path likelihood value being less than a preset threshold, a wake-up operation is performed on the voice device associated with the preset wake-up word.

2. The voice wake-up method according to claim 1, characterized in that, The length of the decoded acoustic feature of the preset wake-up word is the length of the weighted undirected graph corresponding to the preset wake-up word.

3. The voice wake-up method according to claim 1, characterized in that, The modeling unit levels include words, syllables, initials and finals, and states.

4. The voice wake-up method according to claim 1, characterized in that, Before extracting the first state sequence of the acoustic feature stream based on the first modeling unit, the voice wake-up method further includes: Acquire the training data and extract the training acoustic features from the training data; Based on the acoustic features to be trained, a sequence of acoustic features to be trained and a sequence of acoustic annotations to be trained are obtained, wherein the sequence of acoustic annotations to be trained is the annotation sequence of the first modeling unit; Obtain the mutual information between the acoustic feature sequence to be trained and the acoustic annotation sequence of the first modeling unit; The first modeling unit is trained using the mutual information to obtain the first modeling unit.

5. The voice wake-up method according to claim 4, characterized in that, The step of training the first modeling unit using the mutual information to obtain the first modeling unit includes: Using the mutual information, a target training function is constructed; The first modeling unit is trained by increasing the objective function value of the target training function until the objective function value reaches a preset threshold, thus obtaining the first modeling unit.

6. The voice wake-up method according to claim 3, characterized in that, Before extracting the first state sequence of the acoustic feature stream based on the first modeling unit, the voice wake-up method further includes: The first decoding graph is constructed using the first modeling unit, the context-dependent model, the pronunciation dictionary model, and the language model.

7. The voice wake-up method according to claim 4, characterized in that, Before extracting the acoustic features of the data to be trained, the voice wake-up method further includes: The training data is subjected to data augmentation processing, which includes adding reverb, noise, velocity perturbation and / or volume perturbation.

8. A voice wake-up device, characterized in that, The voice wake-up device includes a processor and a memory connected to the processor, wherein... The memory stores program instructions; The processor is used to execute program instructions stored in the memory to implement the voice wake-up method as described in any one of claims 1 to 7.

9. A computer-readable storage medium, characterized in that, The storage medium stores program instructions, which, when executed, implement the voice wake-up method as described in any one of claims 1 to 7.

Images