CN115457938A - Method, device, storage medium and electronic device for identifying wake-up words - Google Patents

Abstract

Embodiments of the present invention provide a method, device, storage medium, and electronic device for recognizing wake-up words, wherein the method includes: performing feature extraction on the target voice signal to obtain multi-frame acoustic feature vectors; The acoustic feature vector is processed to obtain the target processing result; the multi-frame acoustic feature vector is decoded through the decoding map to obtain the target decoding result; according to the target processing result and the target decoding result, the recognition result of the wake-up word in the target speech signal is determined. The present invention solves the problem in the related art that the accuracy of wake-up word recognition is low due to crosstalk between wake-up words.

Description

Method, device, storage medium and electronic device for identifying wake-up words

technical field

Embodiments of the present invention relate to the field of voice wake-up, and in particular, relate to a method, device, storage medium, and electronic device for recognizing wake-up words.

Background technique

In recent years, with the rapid development of information technology, technologies related to speech recognition have greatly facilitated and enriched people's lives. Perfect voice wake-up function is configured in smart home equipment, video conferencing equipment, home appliances, etc. The user can speak the wake word, wake up the device, and start the human-computer voice interaction with the device. Therefore, voice wake-up is an important part of voice interaction.

In the current wake-up word recognition, it is often necessary to define multiple wake-up words, and train these different wake-up words at the same time, and perform classification tasks in the same model, which will lead to crosstalk between multiple wake-up words, resulting in recognition The accuracy of the wake-up word is not high, which will increase the false wake-up rate of the device. Therefore, there is a problem in the prior art that the recognition accuracy of wake-up words is low due to crosstalk between wake-up words.

Aiming at the problem in the related art that the accuracy of wake-up word recognition is low due to the crosstalk between wake-up words, no effective solution has been proposed so far.

Contents of the invention

Embodiments of the present invention provide a method, device, storage medium, and electronic device for recognizing wake-up words, so as to at least solve the problem in the related art that the accuracy of wake-up word recognition is low due to crosstalk between wake-up words.

According to an embodiment of the present invention, a method for recognizing wake-up words is provided, including: performing feature extraction on a target speech signal to obtain a multi-frame acoustic feature vector; processing the multi-frame acoustic feature vector through a deep neural network, Obtain a target processing result; decode the multi-frame acoustic feature vectors through a decoding map to obtain a target decoding result; determine a recognition result of a wake-up word in the target speech signal according to the target processing result and the target decoding result.

In an exemplary embodiment, processing the multi-frame acoustic feature vectors through a deep neural network to obtain a target processing result includes: inputting the multi-frame acoustic feature vectors into a deep neural network, and processing the multi-frame acoustic feature vectors through the deep neural network The acoustic feature vectors of each frame in the multi-frame acoustic feature vectors are classified to obtain the phoneme posterior feature vectors corresponding to the acoustic feature vectors of each frame, wherein the target processing result includes the phoneme posterior feature vectors corresponding to the acoustic feature vectors of each frame test feature vector.

In an exemplary embodiment, the decoding the multi-frame acoustic feature vectors through a decoding graph to obtain a target decoding result includes: decoding the multi-frame acoustic feature vectors through multiple paths in the decoding graph, Obtaining a target path; determining the target path as the target decoding result.

In an exemplary embodiment, decoding the multi-frame acoustic feature vectors through multiple paths in the decoding graph to obtain the target path includes: using a token passing algorithm in the multiple paths in the decoding graph Determine the target path.

In an exemplary embodiment, the identifying the wake-up word to be recognized in the target voice signal according to the target processing result and the target decoding result includes: including the wake-up word to be recognized on the target path In the case of , the target phoneme posterior feature vector corresponding to the wake-up word to be recognized is determined in the phoneme posterior feature vector corresponding to the acoustic feature vector of each frame, wherein the target processing result includes the The phoneme posterior feature vector corresponding to the acoustic feature vector, the target decoding result includes the target path; the wake-up word to be recognized is identified through the target phoneme posterior feature vector.

In an exemplary embodiment, the identifying the wake-up word to be recognized through the target phoneme posterior feature vector includes: determining the relationship between the target phoneme posterior feature vector and a preset standard template Target distance: identifying the wake-up word to be recognized according to the relationship between the target distance and a preset standard distance.

In an exemplary embodiment, the identifying the wake-up word to be recognized according to the relationship between the target distance and a preset standard distance includes: If the difference between them is less than or equal to the preset threshold, the wake-up word to be recognized will be determined as the recognition result of the wake-up word in the target voice signal.

According to yet another embodiment of the present invention, there is also provided a device for recognizing wake-up words, including: an extraction module for feature extraction of the target voice signal to obtain multi-frame acoustic feature vectors; a processing module for The network processes the multi-frame acoustic feature vectors to obtain target processing results; the decoding module is used to decode the multi-frame acoustic feature vectors through the decoding map to obtain target decoding results; the determination module is used to obtain target decoding results according to the target Processing the result and the target decoding result to determine the recognition result of the wake-up word in the target voice signal.

According to yet another embodiment of the present invention, a computer-readable storage medium is also provided, and a computer program is stored in the computer-readable storage medium, wherein the computer program is configured to perform any one of the above methods when running Steps in the examples.

According to yet another embodiment of the present invention, there is also provided an electronic device, including a memory and a processor, wherein a computer program is stored in the memory, and the processor is configured to run the computer program to perform any of the above Steps in the method examples.

Through the present invention, the target decoding result is obtained by decoding the acoustic feature vector, and the target result is verified through the multi-frame posterior feature vector corresponding to the acoustic feature vector, and the target decoding result is not directly used as the recognition result, but is processed according to the target processing result After the target decoding result is verified, the recognition result of the wake-up word in the target voice signal is determined, which avoids the crosstalk between the wake-up words and makes the wake-up word in the target voice signal recognized as other wake-up words, thereby improving the accuracy of wake-up word recognition Rate.

Description of drawings

FIG. 1 is a block diagram of the hardware structure of a mobile terminal according to a method for identifying a wake-up word according to an embodiment of the present invention;

FIG. 2 is a flowchart of a method for identifying a wake-up word according to an embodiment of the present invention;

FIG. 3 is a flow chart of a method for identifying a wake-up word according to a specific embodiment of the present invention;

Fig. 4 is a structural block diagram of an apparatus for recognizing a wake-up word according to an embodiment of the present invention.

detailed description

Embodiments of the present invention will be described in detail below with reference to the drawings and in combination with the embodiments.

It should be noted that the terms "first" and "second" in the description and claims of the present invention and the above drawings are used to distinguish similar objects, but not necessarily used to describe a specific sequence or sequence.

The method embodiments provided in the embodiments of the present application may be executed in mobile terminals, computer terminals or similar computing devices. Taking running on a mobile terminal as an example, FIG. 1 is a block diagram of a hardware structure of a mobile terminal according to a method for recognizing wake-up words according to an embodiment of the present invention. As shown in Figure 1, the mobile terminal may include one or more (only one is shown in Figure 1) processors 102 (processor 102 may include but not limited to processing devices such as microprocessor MCU or programmable logic device FPGA, etc.) and a memory 104 for storing data, wherein the above-mentioned mobile terminal may also include a transmission device 106 and an input and output device 108 for communication functions. Those skilled in the art can understand that the structure shown in FIG. 1 is only for illustration, and it does not limit the structure of the above mobile terminal. For example, the mobile terminal may also include more or fewer components than those shown in FIG. 1 , or have a different configuration from that shown in FIG. 1 .

The memory 104 can be used to store computer programs, for example, software programs and modules of application software, such as the computer program corresponding to the method for recognizing wake-up words in the embodiment of the present invention, and the processor 102 runs the computer program stored in the memory 104, thereby Executing various functional applications and data processing is to realize the above-mentioned method. The memory 104 may include high-speed random access memory, and may also include non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some examples, the memory 104 may further include a memory that is remotely located relative to the processor 102, and these remote memories may be connected to the mobile terminal through a network. Examples of the aforementioned networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

The transmission device 106 is used to receive or transmit data via a network. The specific example of the above network may include a wireless network provided by the communication provider of the mobile terminal. In one example, the transmission device 106 includes a network interface controller (NIC for short), which can be connected to other network devices through a base station so as to communicate with the Internet. In one example, the transmission device 106 may be a radio frequency (Radio Frequency, RF for short) module, which is used to communicate with the Internet in a wireless manner.

In this embodiment, a method for identifying a wake-up word is provided. FIG. 2 is a flowchart of a method for identifying a wake-up word according to an embodiment of the present invention. As shown in FIG. 2 , the process includes the following steps:

Step S202, performing feature extraction on the target speech signal to obtain multi-frame acoustic feature vectors;

Step S204, processing the multi-frame acoustic feature vectors through a deep neural network to obtain a target processing result;

Step S206, decoding the multi-frame acoustic feature vectors through the decoding map to obtain a target decoding result;

Step S208, according to the target processing result and the target decoding result, determine the recognition result of the wake-up word in the target voice signal.

In the technical solution provided in the above step S202, the target voice signal is the voice signal collected by the voice acquisition device. Before the wake-up word recognition in the target voice signal is performed, it is necessary to extract effective acoustic features according to the waveform of the voice signal. Feature extraction is important for subsequent The accuracy of the wake word recognition system is extremely critical.

The model for extracting the acoustic feature vector may include at least one of the following: MFCC (Mel-FrequencyCepstral Coefficients, Mel cepstral coefficients), FBANK (Filter Bank, filter bank feature), PLP (Perceptual Linear Predictive, perceptual linear prediction), PCEN (Per-Channel EnergyNormalization, channel energy normalization features), etc.

In the technical solution provided in the above step S204, the deep neural network can be a deep neural network in a deep neural network-hidden Markov model architecture (Deep Neural Network-Hidden Markov Model, DNN-HMM), and the multi-frame acoustic feature vector Input it into the deep neural network to obtain the posterior feature vector of each frame of the acoustic feature vector in the multi-frame acoustic feature vector on the pre-defined k-type phonemes, that is, the target processing result.

In the technical solution provided by the above step S206, the above-mentioned decoding map is an HCLG decoding map, and the HCLG decoding map is a large resource map constructed by language models, dictionaries, context phonemes and HMMs. The multi-frame acoustic feature vectors are decoded in HCLG to obtain the target decoding result.

The HCLG decoding graph contains multiple paths, and one or more optimal paths are selected as the target decoding result through the decoding process, and each path includes the corresponding word-level results, for example, the most optimal path decoded on HCLG The optimal path is path 1, and the word-level result corresponding to path 1 is "please turn on the device".

In the technical solution provided in the above step S208, it is judged in the target decoding result whether the wake-up word is parsed, if the wake-up word is not parsed, it is determined that the target voice signal does not contain the wake-up word, and the device is not woken up; and if the wake-up word is parsed In the case of , determine whether the wake-up word decoded in the decoding map is the wake-up word contained in the target speech signal according to the target processing result and the target decoding result. E.g. The decoding result in the HCLG decoding diagram is "Please turn on the device", then the decoding result contains the wake-up word "turn on", and further combined with the target processing result and the target decoding result to confirm that the target voice signal contains the wake-up word "turn on", After confirmation, "ON" is used as the wake-up word recognized in the target voice signal.

Through the above steps, the target decoding result is obtained by decoding the acoustic feature vector, and the target result is verified through the multi-frame a posteriori feature vector corresponding to the acoustic feature vector. The target decoding result is not directly used as the recognition result, but is processed according to the target After the target decoding result is verified, the recognition result of the wake-up word in the target speech signal is determined. Since the acoustic feature vector is processed to obtain the phoneme posterior feature, the decoding result after decoding the multi-frame acoustic feature vector is combined with the phoneme posterior feature. The accuracy of the decoding result obtained in the decoding process is judged to reduce the error in the decoding result caused by the crosstalk between the wake-up words, thereby determining the recognition result of the wake-up word in the target voice signal, and avoiding the error caused by the wake-up word The crosstalk among them enables the recognition of the wake-up word in the target voice signal as other wake-up words, thereby improving the accuracy of wake-up word recognition.

In an optional embodiment, processing the multi-frame acoustic feature vectors through a deep neural network to obtain a target processing result includes: inputting the multi-frame acoustic feature vectors into a deep neural network, through the deep neural network Classifying each frame of the acoustic feature vectors in the multi-frame acoustic feature vectors to obtain the phoneme posterior feature vectors corresponding to the acoustic feature vectors of each frame, wherein the target processing result includes the phoneme corresponding to the acoustic feature vectors of each frame Posterior eigenvectors.

In this embodiment, K phonemes are predefined, and phoneme modeling is performed on wake-up words and non-wake-up words, and fine modeling is adopted for wake-up words, that is, one wake-up word contains multiple phonemes; Extensive modeling is adopted, that is, non-awakening words are used to model words, and a non-awakening word corresponds to a phoneme.

第一帧声学特征向量o₁对应的音素后验特征向量为

第二帧声学特征向量o₂对应的音素后验特征向量为

以此类推第n帧声学特征向量o_n对应的音素后验特征向量为

The input of the deep neural network is an acoustic feature vector of a frame, and the output is the phoneme posterior feature vector corresponding to the frame. For example, the multi-frame acoustic feature vector obtained by feature extraction of the target speech signal is: u _o ={o ₁ ,o ₂ ,...,o _n }, where n is the number of frames of multi-frame acoustic feature vectors, input the acoustic feature vectors of each frame into the deep neural network in turn to obtain the phoneme posterior feature vectors of the corresponding frames, and the multi-frame acoustic feature vectors correspond to Multi-frame phoneme posterior feature vectors:

The phoneme posterior feature vector corresponding to the acoustic feature vector o ₁ of the first frame is

The phoneme posterior feature vector corresponding to the acoustic feature vector o ₂ of the second frame is

By analogy, the phoneme posterior feature vector corresponding to the acoustic feature vector o _n of the nth frame is

A phoneme posterior feature vector PG represents the posterior probability distribution of the corresponding speech feature vector o on the predefined K phonemes {C ₁ , C ₂ ,...,C _k }, expressed as:

PG _o ＝{p(C ₁ |o), p(C ₂ |o),...p(C _k |o)}

Where p(C ₁ |o) is the posterior probability of feature vector o on the first type of phoneme, p(C ₂ |o) is the posterior probability of feature vector o on the second type of phoneme, and so on, p (C _k |o) is the posterior probability of feature vector o on the k-th phoneme.

The target processing result includes multi-frame phoneme posterior feature vectors, and each frame of phoneme posterior feature vectors in the multi-frame phoneme posterior feature vectors is in one-to-one correspondence with each frame of acoustic feature vectors in the multi-frame acoustic feature vectors.

In an optional embodiment, the decoding the multi-frame acoustic feature vectors through the decoding graph to obtain the target decoding result includes: decoding the multi-frame acoustic feature vectors through multiple paths in the decoding graph , to obtain a target path; and determine the target path as the target decoding result.

In this embodiment, the HCLG decoding graph contains multiple paths, and the sum of the outputs of all nodes on each path constitutes an output sentence or word. After constructing the HCLG decoding map, find the optimal path in the HCLG decoding map according to the multi-frame acoustic feature vector. The cost of the output label sequence on the optimal path on the speech to be recognized should be as small as possible, and the output taken out on the optimal path The label sequence is the word-level recognition result, and this process is decoding. The optimal multiple paths can also be found, and the identification result of the optimal multiple paths is called an N-best list.

In an optional embodiment, decoding the multi-frame acoustic feature vectors through multiple paths in the decoding graph to obtain the target path includes: using a token passing algorithm on the multiple paths in the decoding graph Determine the target path in .

In this embodiment, the token passing algorithm is to place tokens on the start nodes, one start node corresponds to one token, and if there are multiple start nodes, each start node places a token. Decode multi-frame acoustic feature vectors frame by frame, after decoding the first frame (decoding the first acoustic feature vector), pass the token on the starting node to the next node according to the decoded information, and calculate the transfer cost, Then decode the acoustic feature vector of the second frame, pass the token from the current node to the next node according to the decoded information, accumulate the transfer cost, decode all the acoustic feature vectors in turn, and check the token after decoding the last frame The state node where it is located traces back the path that the current token has traveled, and calculates and accumulates the transfer cost during each transfer. If a state node has multiple jumps, multiple copies of the token are passed on separately. After passing to the last frame, check the transfer cost of all tokens in the current decoding graph, and select the optimal path or paths according to the transfer cost. The lower the transfer cost, the better the corresponding path.

The optimal path refers to the global optimum, and the global optimum must be locally optimal, that is, if a path is globally optimal, then the path must be a locally optimal path through any state. Therefore, when multiple tokens are passed to the same state node, only the optimal token (the token with the smallest cumulative cost) can be kept.

In an optional embodiment, the identifying the wake-up word to be recognized in the target speech signal according to the target processing result and the target decoding result includes: including the wake-up word to be recognized on the target path In the case of words, the target phoneme posterior feature vector corresponding to the wake-up word to be recognized is determined in the phoneme posterior feature vector corresponding to the acoustic feature vector of each frame, wherein the target processing result includes the each The phoneme posterior feature vector corresponding to the frame acoustic feature vector, the target decoding result includes the target path; the wake-up word to be recognized is identified through the target phoneme posterior feature vector.

In this embodiment, it is judged whether the output sentence or word corresponding to the target path contains a wake-up word. One or more frames of phoneme posterior feature vectors corresponding to the wake-up word are found in the phoneme posterior feature vector of the target phoneme posterior feature vector, and whether the wake-up word is used as this recognition is determined according to the target phoneme posterior feature vector. the result of.

Wherein, the target phoneme posterior feature vector includes one or more frames of phoneme posterior feature vectors.

假设其中“请”对应于第一个音素后验特征向量

“开启”对应于第2～8帧音素后验特征向量

设备对应于第9和10帧音素后验特征向量PG_o9、PG_o10，则将{PG_o2,...,PG_o8}确定为目标音素后验特征向量。For example, the output sentence in the target path is "Please turn on the device", and "please" and "device" in this sentence are non-wake-up words, and "turn on" is a wake-up word. In this case, the target path is If it contains the wake-up word to be recognized, find the phoneme posterior feature vector corresponding to "open" in the multi-frame phoneme posterior feature vector in the target processing result, and determine it as the target phoneme posterior feature vector, assuming that the acoustic feature vector Classification is performed to obtain 10 frames of phoneme posterior feature vectors, expressed as:

Suppose where "please" corresponds to the first phoneme posterior feature vector

"On" corresponds to the phoneme posterior feature vectors of the 2nd to 8th frames

The device corresponds to the phoneme posterior feature vectors PG _o9 and PG _o10 of the 9th and 10th frames, and {PG _o2 ,...,PG _o8 } are determined as the target phoneme posterior feature vectors.

In an optional embodiment, the identifying the wake-up word to be recognized through the target phoneme posterior feature vector includes: determining the difference between the target phoneme posterior feature vector and a preset standard template The target distance; the wake-up word to be recognized is identified according to the relationship between the target distance and a preset standard distance.

In this embodiment, a corresponding phoneme posterior feature vector sequence (i.e. standard template) is preset for each wake-up word, and the relationship between the standard template corresponding to the wake-up word to be recognized and the target phoneme posterior feature vector sequence is calculated. The target distance, the relationship between the target distance and the preset standard distance determines whether to use the wake-up word to be recognized as the recognition result of the wake-up word this time.

It should be noted that the dynamic warping algorithm is used to calculate the target distance between the standard template and the target phoneme posterior feature vector sequence.

The phoneme posterior feature vector sequence corresponding to the standard template is u _x ={x ₁ ,x ₂ ,...,x _n }, u _y ={y ₁ ,y ₂ ,...,y _m }, n and m Indicate the number of frames of the two sequence phoneme posterior feature vectors respectively, establish a distance matrix D, and the element D(i, j) in the distance matrix represents the i-th frame vector in the standard template and the j-th sequence of the target phoneme posterior feature vector The distance between frame vectors, using the negative logarithmic inner product metric to represent the distance between the i-th frame vector in the standard template and the j-th frame vector of the target phoneme posterior feature vector sequence, that is, the distance between the i-th frame vector and the j-th frame vector The distance between is D(i,j)=-lg( _xi ·y _j ).

Let φ represent a possible correspondence between u _x and u _y :

φ(k)=(i _k , j _k ), k=1, 2, . . . , T.

Among them, T represents time, k is the independent variable of T, and the i-th frame vector in the u _x sequence corresponds to the j-th frame sequence in the u _y sequence at time k.

Find an optimal corresponding sequence φ' in the matrix D, and the minimum cumulative distortion value corresponding to the optimal sequence φ' is:

The minimum cumulative distortion value is determined as the target distance between the target phoneme posterior feature vector sequence and the preset standard template.

In an optional embodiment, the identifying the wake-up word to be recognized according to the relationship between the target distance and a preset standard distance includes: When the difference between them is less than or equal to the preset threshold, the wake-up word to be recognized is determined as the recognition result of the wake-up word in the target voice signal.

In this embodiment, the preset standard distance is calculated through the wake-up word standard template of the positive sample data of the wake-up word test set. Therefore, by calculating the distance between the posterior feature vector sequence corresponding to each test sample and the standard template, multiple matching distances can be obtained, and the standard distance can be obtained according to multiple matching distances. The standard distance can be obtained by calculating the average value of multiple matching distances, or can be determined by other methods, which are not limited here.

Taking the wake-up word as "open" as an example, there are 10 "open" test samples in the wake-up word test set. In these 10 test samples, "open" may be spoken by different accents and at different speeds. Obtain the posterior feature vector sequence of each test sample, and the standard template is the posterior feature vector sequence obtained from speech spoken in Mandarin and standard speech speed. The posterior feature vector sequence of each test sample is matched with the standard template to obtain the standard distance.

When the difference between the calculated target distance and the standard distance is less than or equal to the preset threshold, the wake-up word to be recognized in the target decoding result is determined as the final recognition result, that is, the wake-up word in the target voice signal recognition result. For example, the wake-up word to be recognized in the target decoding result "Please turn on the device" is "ON", and when the difference between the calculated target distance and the standard distance is less than or equal to the preset threshold, "ON" is determined as The recognition result is to determine that the target voice signal contains the wake-up word "on", and then the device performs the corresponding operation according to the wake-up word, that is, the device performs the turn-on operation.

Apparently, the above-described embodiments are only part of the embodiments of the present invention, not all of them.

The present invention is specifically described below in conjunction with embodiment:

Fig. 3 is a flow chart of the method for identifying wake-up words according to a specific embodiment of the present invention, as shown in Fig. 3,

S301: Acquire a speech signal, and extract multi-frame acoustic feature vectors;

Collect the voice signal through the voice signal acquisition device on the device, and extract the acoustic features from the voice signal;

S302: Using a deep neural network to classify multi-frame acoustic feature vectors, and obtain multi-frame phoneme posterior feature vectors;

The trained model is used to classify multi-frame acoustic feature frames. The trained model is a deep neural network-hidden Markov model (Deep Neural Network-Hidden Markov Model, DNN-HMM) architecture, and the modeling unit is a phoneme.

The input of the deep neural network is the acoustic feature of a frame, and the output is a phoneme posterior feature vector. Input an acoustic feature vector o, and output the posterior probability distribution of the feature vector on the pre-defined k classes {C ₁ ,C ₂ ,...,C _k } as follows:

PG _o ＝(p(C ₁ |o)p(C ₂ |o)...p(C _k |o))

where p(C _i |o) is the posterior probability of the feature vector o on the i-th class, where a class can be defined as any kind of speech unit, such as a phoneme. The phone-level posterior features are used in this patent.

Using the DNN-HMM model that has been trained, the DNN in it can obtain the phoneme posterior features in units of frames.

S303: Construct the HCLG decoding map, perform timing decoding, and obtain the optimal path;

In the DNN-HMM model, the decoding of the acoustic feature sequence is realized according to the HCLG decoding graph, and a list of recognition results (N-best list) is obtained through the optimal multiple paths found in the decoding graph. The construction of the decoding graph relies on a large resource graph composed of language models, dictionaries, context phonemes, and HMMs. The token passing algorithm is carried out by frame. When the execution reaches the last frame, the token passing ends. At this time, check the tokens in the termination state, take the optimal one or more tokens, and take out or backtrack according to the information on them. Find the path corresponding to these tokens, and thus get the recognition result. This path accumulates the likelihood values of the two parts of the acoustic model and the language model, assuming that the cumulative likelihood value of several frames decoding the wake word in the path with the highest likelihood value is _Ph .

S304: According to whether the wake-up word is decoded in the optimal path, perform phoneme posterior feature matching, and calculate the negative logarithmic inner product;

According to the result of HMM timing decoding, it is determined whether to perform phoneme posterior probability matching. Once the wake-up word is decoded in step 3, the phoneme posterior probability matching is performed according to the frame corresponding to the wake-up word on the path.

The phoneme posterior probability is the phoneme classification probability output by DNN. The timing matching uses the dynamic time warping algorithm to calculate the distance between the phoneme posterior probability sequence corresponding to the wake-up word and the standard template. Here, the inner product metric is used to calculate the distance between the two sequences .

The sequence matching of phoneme feature vectors adopts the following dynamic time warping algorithm. Given the feature vector sequences of two consecutive speech segments, u _x ={x ₁ ,x ₂ ,...,x _n }, u _y ={y ₁ ,y ₂ ,...,y _m }, n and m respectively represent the frame number of the two sequence feature vectors, by defining the distance between the feature vectors of speech frames, a distance matrix D is established, and φ represents a possible correspondence between u _x and u _y , φ( k)=(i _k , j _k ), k=1,2,...,T, where T represents time, k is an independent variable of T, at time k the i-th frame vector in u _x sequence is related to u corresponds to the jth frame sequence in the _y sequence.

Find an optimal corresponding sequence φ′ in the matrix D, so as to minimize the accumulated distortion value Dist _φ (u _x ,u _y ),

In this embodiment, the speech frame is represented by the phoneme posterior feature, and the negative logarithmic inner product measure is used, then:

The final distance between the sequence to be matched and the template sequence is denoted as P _d , and its value is the minimum cumulative distortion value Dist _φ (u _x ,u _y ).

S305: Use the positive sample data of the wake-up word test set to perform distance matching with the standard template of the wake-up word to obtain a distance threshold;

Through the DTW distance matching between the positive sample data of the wake-up word test set and the standard template of the wake-up word, the distance threshold (that is, the standard distance) can be calculated.

The calculation of the DTW distance can adopt the existing technology.

S306: Comparing the difference between the distance between the sequence to be matched and the template sequence and the distance threshold, and obtaining a wake-up result;

该差值小到超过人为设定的某一阈值，则判定当前唤醒词为识别到的唤醒词，输出唤醒结果。Compare the distance between the current sequence to be matched and the standard template sequence, and calculate

If the difference is so small as to exceed a certain threshold set manually, it is determined that the current wake-up word is the recognized wake-up word, and a wake-up result is output.

Through the description of the above embodiments, those skilled in the art can clearly understand that the method according to the above embodiments can be implemented by means of software plus a necessary general-purpose hardware platform, and of course also by hardware, but in many cases the former is Better implementation. Based on such an understanding, the essence of the technical solution of the present invention or the part that contributes to the prior art can be embodied in the form of software products, and the computer software products are stored in a storage medium (such as ROM/RAM, disk, CD) contains several instructions to enable a terminal device (which may be a mobile phone, a computer, a server, or a network device, etc.) to execute the methods described in various embodiments of the present invention.

In this embodiment, a device for recognizing wake-up words is also provided. FIG. 4 is a structural block diagram of a device for recognizing wake-up words according to an embodiment of the present invention. As shown in FIG. 4 , the device includes:

Extraction module 402, is used for carrying out feature extraction to target speech signal, obtains multi-frame acoustic feature vector;

A processing module 404, configured to process the multi-frame acoustic feature vectors through a deep neural network to obtain a target processing result;

A decoding module 406, configured to decode the multi-frame acoustic feature vectors through a decoding map to obtain a target decoding result;

The determination module 408 is configured to determine the recognition result of the wake-up word in the target voice signal according to the target processing result and the target decoding result.

In an optional embodiment, the above-mentioned device is further configured to input the multi-frame acoustic feature vectors into a deep neural network, and perform an acoustic feature vector on each frame of the multi-frame acoustic feature vectors through the deep neural network. Classify to obtain the phoneme posterior feature vectors corresponding to the acoustic feature vectors of each frame, wherein the target processing result includes the phoneme posterior feature vectors corresponding to the acoustic feature vectors of each frame.

In an optional embodiment, the above device is further configured to decode the multi-frame acoustic feature vectors through multiple paths in the decoding graph to obtain a target path; and determine the target path as the target decoding result .

In an optional embodiment, the above apparatus is further configured to determine the target path among the multiple paths in the decoding graph by using a token passing algorithm.

In an optional embodiment, the above-mentioned device is also used to determine, in the case that the target path contains the wake-up word to be recognized, in the phoneme posterior feature vector corresponding to each frame acoustic feature vector The target phoneme posterior feature vector corresponding to the wake-up word, wherein the target processing result includes the phoneme posterior feature vector corresponding to the acoustic feature vector of each frame, and the target decoding result includes the target path; through the The target phoneme posterior feature vector identifies the wake-up word to be identified.

In an optional embodiment, the above device is also used to determine the target distance between the target phoneme posterior feature vector and the preset standard template; according to the target distance between the target distance and the preset standard distance The relationship identifies the wake word to be identified.

In an optional embodiment, the above-mentioned device is further configured to, when the difference between the target distance and the preset target standard distance is less than or equal to a preset threshold, send the wake-up word to be recognized to It is determined as the recognition result of the wake-up word in the target voice signal.

It should be noted that the above-mentioned modules can be realized by software or hardware. For the latter, it can be realized by the following methods, but not limited to this: the above-mentioned modules are all located in the same processor; or, the above-mentioned modules can be combined in any combination The forms of are located in different processors.

Embodiments of the present invention also provide a computer-readable storage medium, in which a computer program is stored, wherein the computer program is set to execute the steps in any one of the above method embodiments when running.

In an exemplary embodiment, the above-mentioned computer-readable storage medium may include but not limited to: U disk, read-only memory (Read-Only Memory, referred to as ROM), random access memory (Random Access Memory, referred to as RAM) , mobile hard disk, magnetic disk or optical disk and other media that can store computer programs.

An embodiment of the present invention also provides an electronic device, including a memory and a processor, where a computer program is stored in the memory, and the processor is configured to run the computer program to perform the steps in any one of the above method embodiments.

In an exemplary embodiment, the electronic device may further include a transmission device and an input and output device, wherein the transmission device is connected to the processor, and the input and output device is connected to the processor.

For specific examples in this embodiment, reference may be made to the examples described in the foregoing embodiments and exemplary implementation manners, and details will not be repeated here in this embodiment.

Obviously, those skilled in the art should understand that each module or each step of the above-mentioned present invention can be realized by a general-purpose computing device, and they can be concentrated on a single computing device, or distributed in a network formed by multiple computing devices In fact, they can be implemented in program code executable by a computing device, and thus, they can be stored in a storage device to be executed by a computing device, and in some cases, can be executed in an order different from that shown here. Or described steps, or they are fabricated into individual integrated circuit modules, or multiple modules or steps among them are fabricated into a single integrated circuit module for implementation. As such, the present invention is not limited to any specific combination of hardware and software.

The above descriptions are only preferred embodiments of the present invention, and are not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and changes. Any modification, equivalent replacement, improvement, etc. made within the principle of the present invention shall be included in the protection scope of the present invention.

Claims (10)

1. A method for identifying wake-up words, comprising: Perform feature extraction on the target speech signal to obtain multi-frame acoustic feature vectors; Processing the multi-frame acoustic feature vectors through a deep neural network to obtain a target processing result; Decoding the multi-frame acoustic feature vectors through a decoding map to obtain a target decoding result; According to the target processing result and the target decoding result, determine the recognition result of the wake-up word in the target voice signal. 2. The method according to claim 1, characterized in that, the multi-frame acoustic feature vectors are processed by a deep neural network to obtain target processing results, including: The multi-frame acoustic feature vector is input into a deep neural network, and each frame acoustic feature vector in the multi-frame acoustic feature vector is classified by the deep neural network to obtain a phoneme posterior feature vector corresponding to each frame acoustic feature vector , wherein the target processing result includes the phoneme posterior feature vectors corresponding to the acoustic feature vectors of each frame. 3. The method according to claim 1, wherein said multi-frame acoustic feature vectors are decoded through a decoding map to obtain a target decoding result, comprising: Decoding the multi-frame acoustic feature vectors through multiple paths in the decoding graph to obtain a target path; The target path is determined as the target decoding result. 4. The method according to claim 3, wherein the multi-frame acoustic feature vector is decoded by multiple paths in the decoding graph to obtain the target path, including: The target path is determined among the plurality of paths in the decoding graph by a token passing algorithm. 5. The method according to any one of claims 1 to 4, wherein, according to the target processing result and the target decoding result, the wake-up word to be recognized in the target voice signal is identified ,include: In the case that the wake-up word to be identified is contained on the target path, the target phoneme posterior feature vector corresponding to the wake-up word to be identified is determined in the phoneme posterior feature vector corresponding to the acoustic feature vector of each frame, wherein the The target processing result includes the phoneme posterior feature vectors corresponding to the acoustic feature vectors of each frame, and the target decoding result includes the target path; The wake-up word to be recognized is recognized through the target phoneme posterior feature vector. 6. The method according to claim 5, wherein the identifying the wake-up word to be identified by the target phoneme posterior feature vector comprises: Determining the target distance between the target phoneme posterior feature vector and a preset standard template; The wake-up word to be recognized is recognized according to the relationship between the target distance and a preset standard distance. 7. The method according to claim 6, wherein the identifying the wake-up word to be identified according to the relationship between the target distance and a preset standard distance comprises: If the difference between the target distance and the preset target standard distance is less than or equal to a preset threshold, the wake-up word to be recognized is determined as the recognition result of the wake-up word in the target voice signal. 8. A device for recognizing wake-up words, comprising: The extraction module is used to extract the features of the target speech signal to obtain multi-frame acoustic feature vectors; A processing module, configured to process the multi-frame acoustic feature vectors through a deep neural network to obtain a target processing result; A decoding module, configured to decode the multi-frame acoustic feature vectors through a decoding map to obtain a target decoding result; A determining module, configured to determine the recognition result of the wake-up word in the target voice signal according to the target processing result and the target decoding result. 9. A computer-readable storage medium, characterized in that, a computer program is stored in the computer-readable storage medium, wherein when the computer program is executed by a processor, any one of claims 1 to 7 is implemented. The steps of the method. 10. An electronic device comprising a memory, a processor, and a computer program stored on the memory and operable on the processor, characterized in that the rights are realized when the processor executes the computer program The steps of the method described in any one of Claims 1 to 7.

Images