Robot personification voice work order alarm system based on voice instruction matching technology

Abstract

The invention relates to an anthropomorphic work order alarm system based on a voice instruction matching technology, and aims to improve work order processing efficiency and user experience. The system can receive and analyze a voice instruction of a user in real time through an advanced voice recognition technology, automatically generate a work order and carry out intelligent alarm. And the personification design enables the system to carry out natural conversation with the user and provide personalized services. The system not only can reduce the complexity of manual operation, but also can improve the response speed, is helpful for handling emergencies in time, and is suitable for work order management and alarm requirements of various industries.

Description

Technical Field

[0001] The present invention specifically relates to a humanized work order alarm system based on voice command matching technology, and belongs to the technical field of natural language processing. Background Art

[0002] In today's environment, the rapid development of artificial intelligence technology has promoted the widespread application of human-computer interaction design, enabling people to enjoy various services brought by machines in a more convenient and efficient way. For example, Alibaba's Tmall Genie allows users to interact with speakers through voice commands, and the speakers communicate with users through voice replies, greatly improving the convenience and diversity of services. Similarly, Siri in Apple's mobile phones can answer various questions and provide personalized services based on users' voice commands, which not only improves service quality but also reduces the difficulty for the elderly and children to use mobile phones. At the same time, in the field of property services, with the application of image recognition and artificial intelligence technology, robots are gradually able to self-discover problems. If property personnel need to observe the surrounding environment and check the work order content on the computer or mobile phone at the same time during inspection, it will inevitably reduce their work efficiency and response speed, and may even threaten the safety of staff in some cases. In order to solve this problem, the voice work order alarm system came into being. It can significantly improve service quality and response speed, facilitate property personnel to obtain information, and quickly handle work tasks, becoming an important tool to improve work efficiency. Summary of the invention

[0003] In view of the problems in the prior art, the present invention proposes a robot anthropomorphic voice work order alarm system based on voice and natural language processing. The solution integrates voice recognition technology, speech synthesis (TTS) technology and instruction matching technology. When the work order system receives a new work order, the system inserts it into the work order list according to the priority and determines whether a voice alarm is required. If necessary, the work order is placed in the voice alarm work order list according to the priority. The work order alarm system grabs the work order with the highest priority from the list for processing. First, the work order content is extracted, and the main complaint information is converted into voice using TTS technology and broadcast to the staff. After receiving the voice work order, the staff can start processing the task. During the processing, the staff can directly reply by voice to obtain more information. These replies will be converted into text through voice-to-text technology and processed through instruction matching technology. If the instruction matches successfully, the system will perform the corresponding operation; if the instruction is "task completed", the work order alarm system will notify the work order system to delete the work order from the voice alarm work order list and the work order list. Finally, the system will continue to grab new work orders from the work order system for processing.

[0004] In order to achieve the above object, the technical solution of the present invention is as follows: a humanized work order alarm system based on instruction matching, the alarm system comprising:

[0005] Voice recognition module: It can receive and process user voice commands in real time, and adopts advanced voice recognition algorithms to recognize different users' voice inputs with high accuracy.

[0006] Command matching module: matches the recognized voice commands according to the preset command library, and transmits the matching results to the system processing module.

[0007] Alarm generation module: After the command is matched successfully, the system automatically generates the corresponding alarm information, including the alarm type, time, location and related description to ensure the integrity and accuracy of the information.

[0008] Anthropomorphic voice feedback module: The system uses natural language generation technology to convert alarm information into natural and fluent voice, and promptly feedback to users to enhance the human-computer interaction experience.

[0009] Work order system: responsible for receiving newly added work orders, inserting them into the work order list according to priority, and judging whether the newly added work orders need voice alarm. If necessary, insert them into the voice alarm work order list.

[0010] Work order alarm system: read the work order with the highest priority in the voice alarm work order list in the work order system, input it into the alarm generation module, generate alarm voice, and issue an alarm.

[0011] The specific steps of the alarm method of the alarm system are as follows:

[0012] Step 1: Add a new work order and determine whether this work order requires a voice alarm.

[0013] Step 2: If voice alarm is not required, insert the work order list.

[0014] Step 3: If voice alarm is required, insert the voice alarm work order list first.

[0015] Step 4: The work order alarm system reads the work order with the highest priority in the voice alarm work order list.

[0016] Step 5: Extract the main complaint of the work order and convert the text of the main complaint of the work order into voice broadcast to the staff through the TTS system.

[0017] Step 6: After receiving the message, the staff will give a voice command if there is any new instruction or reply.

[0018] Step 7: The system converts the voice command into text through STT technology.

[0019] Step 8: The text input command matching system performs command matching and generates a reply based on the command matching result.

[0020] Step 9: The generated reply is converted into voice reply to the staff through the TTS system.

[0021] In the alarm system, an overlap rate calculation algorithm based on the relative position of text is adopted, which is as follows

[0022] a) When we get two texts, we need to calculate the text overlap rate of the two texts. One is text t1 and the other is text t2. We segment the two texts and generate keyword lists. The lengths of the two word lists are n1 and n2 respectively.

[0023] b) Find the intersection of the two vocabulary lists and get the overlapping vocabulary. The length of the vocabulary list of the overlapping vocabulary is n, so the overlap ratio is overlap_alpha = n / ((n1+n2) / 2),

[0024] c) Scan the words in the overlapping vocabulary lists, calculate the relative positions of the i-th word in the two vocabulary lists as l1 and l2, and then calculate the relative position difference as beta_i = |l1-l2|,

[0025] d) Calculate the overlap rate algorithm based on the relative position of the text:

[0026] overlap=overlap_alpha*(n-beta_1-beta_2-…-beta_i-…-beta_n) / n,

[0027] e) The text overlap rate based on the instruction is:

[0028] overlap_alpha2=n / n2

[0029] f) Algorithm for overlapping ratio of relative position based on instructions:

[0030] First, find the first and last positions of the words in the overlapping vocabulary list in the word list of t1, and form a new list of words at the first position, the last position, and all the words in the middle. Then, calculate the relative position difference beta2_i based on this new list:

[0031] overlap2=overlap_alpha2*(n-beta2_1-beta2_2-…-beta2_i-…-beta2_n) / n.

[0032] The alarm system uses a command matching algorithm, as follows:

[0033] a) First, after receiving the voice-to-text conversion, this text is q1,

[0034] b) Traverse each instruction in the instruction set Inst. If the text overlap rate of instruction I and q1 exceeds 90%, the instruction is matched successfully and the traversal ends.

[0035] c) If the text overlap ratio overlap_alpha2 of instruction I and q1 exceeds 80%, the instruction matches successfully, but the traversal continues.

[0036] d) If the text overlap ratio overlap2 of instruction I and q1 exceeds 70%, the instruction matches successfully, but the traversal continues.

[0037] e) If no suitable instruction is found after instruction matching, the matching fails.

[0038] f) If the match is successful based on the text overlap rate, the matched instruction is semantically matched with the original text. The original text is converted into a sentence vector through the embedding algorithm, and the distance is calculated with the instruction vector of the instruction. If the semantics also match, the instruction is successfully matched.

[0039] g) If the match is successful based on the text overlap ratio, but the semantic match is unsuccessful, the instruction match is sub-successful.

[0040] h) If there is more than one instruction that matches successfully or sub-successfully, all instructions that match successfully will be considered valid and output. If all instructions are sub-successful, the instruction with the highest matching degree will be output as sub-successful.

[0041] The alarm system uses a double verification mechanism of text-to-speech system, as follows:

[0042] a) Capture the highest priority work order from the voice alarm work order list and process it.

[0043] b) First, capture the main complaint of the work order and convert the text into speech using the text-to-speech (TTS) technology.

[0044] c) The converted voice to be broadcast is converted into text again using the speech-to-text model. In this way, we get two texts, one is the work order complaint text t1, and the other is the text t2 of the voice to be broadcast after speech-to-text conversion.

[0045] d) Match the two texts and calculate the text overlap ratio.

[0046] e) When the overlap is less than 90%, it is considered that the model has some problems in processing the text TTS or STT task, and the work order is returned to a non-voice alarm emergency work order, and the abnormal situation is pushed to the technician for processing.

[0047] When the overlap rate is greater than 90%, it is considered that the overlap rate is high. At this time, the two texts are converted into sentence vectors respectively, and then the cosine distance between the two sentence vectors is calculated. If the distance is small enough, the voice will give a normal alarm. If not, the work order content has been greatly distorted due to some individual words in the method, which is also considered an invalid conversion and the work order is returned.

[0048] The alarm system uses text-to-speech TTS technology, as follows:

[0049] a) First, convert the text into phoneme text through the phoneme dictionary,

[0050] b) The next step is to predict the duration of each phoneme. This can be done effectively using a segmentation model. The input for segmentation model training is an audio and the phoneme text corresponding to the audio. The output is the starting position of each phoneme in the audio. The trained phoneme duration model can output the duration of each phoneme.

[0051] c) Fundamental frequency prediction model: The fundamental frequency prediction model first predicts whether a phoneme is pronounced. If so, it predicts its fundamental frequency.

[0052] d) Audio synthesis model. The audio synthesis model synthesizes the final speech text based on the previous phoneme text, predicting the duration of each phoneme and the fundamental frequency of each phoneme.

[0053] The alarm system uses speech-to-text STT technology, as follows:

[0054] a) Speech preprocessing: pre-emphasis, framing, windowing, short-time Fourier transform (STFT), mel filtering, and de-averaging. After such a series of operations, the speech will be divided into small segments. Theoretically, 10ms of speech information is one segment, so a long segment of speech information will eventually be divided into x1, x2, x3, ..., x n The output of this information fragment sequence is Chinese characters or English letters, y=( <sos>,y1,...,y S , <eos>),y S is any character in the prediction range, such as English letters, Chinese characters,

[0055] b) The output is conditional probability:

[0056] P(y|x)=Π t P(y t |x,y ＜t ),

[0057] x is the sound, y is ＜t is the text before the letter or word to be predicted, y i It is any letter or word in y.

[0058] train:

[0059] c) Model architecture: Both seq2seq models are available. Listen, Attend and Spell is an encoder-decoder model. First, the segmented speech fragments pass through a pyramid-shaped BLSTM, and then output the feature vector set h. Then, it passes through a temporal network to generate an attention vector based on h and the intermediate output of the model. Finally, it is decoded to obtain y. t , and y t It will also be used as the next prediction y t+1 Input.

[0060] The alarm system adopts a voice command processing algorithm, which is as follows:

[0061] a) Listen to the voice commands and convert them into text content using STT technology.

[0062] b) After receiving the staff's reply, the work order alarm system will match the instructions.

[0063] c) After the command is successfully matched, a reply will be generated, and the TTS text-to-speech process will be used to generate voice and reply to the staff

[0064] d) If the command matches successfully, further voice confirmation is performed to ask the staff whether this is the command to be issued.

[0065] Compared with the prior art, the present invention has the following advantages: the scheme integrates the latest TTS text-to-speech model and STT speech-to-text model, constructs a work order alarm system, provides a human-computer interaction mechanism, sets an instruction set for employees and the alarm system to interact with each other by voice, greatly improves the immediacy of work order processing, and improves the convenience of employees in processing work orders. The present application designs an algorithm for the overlap rate of text based on the relative position of words, and proposes an algorithm for the overlap rate based on instructions. Based on the brief characteristics of instructions, this overlap rate calculation method has more scene advantages. The scheme designs an instruction matching algorithm, applies multiple strategies (based on text overlap rate and semantic matching) to match instructions, and also takes into account the situation of multiple instructions. When the text overlap rate reaches the matching degree, semantic matching is performed again, which can greatly reduce the computing resources consumed by matching. After eliminating noise, semantic matching is performed again, which is also more conducive to semantic matching to output correct results. In order to ensure the correctness of text-to-speech conversion, so that there will not be too many mistakes when broadcasting to staff, a double verification mechanism is designed for the text-to-speech system. In addition, the TTS text-to-speech system designed by the present invention provides double verification to ensure the effectiveness and accuracy of the voice alarm. The command matching algorithm takes into account the compatibility of the text overlap rate matching algorithm and the semantic matching algorithm. While considering the efficiency, it also ensures the matching quality as much as possible. The human-computer interaction system is the same, with multiple verifications and layered processing, trying to make the work order system anthropomorphic, humanized, and intelligent, so as to provide better services. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] Figure 1 This is the design details of the speech-to-text TTS technology.

[0067] Figure 2 This is a design diagram of the two-factor authentication mechanism for text-to-speech.

[0068] Figure 3 This is the operation flow chart of instruction matching after the staff responds.

[0069] Figure 4 This is the main operation logic diagram of the work order system and work order alarm system. DETAILED DESCRIPTION

[0070] In order to deepen the understanding of the present invention, the present embodiment is described in detail below with reference to the accompanying drawings.

[0071] Example 1: See Figure 1-Figure 4 , a humanized order alarm system based on instruction matching, the alarm system comprising:

[0072] Voice recognition module: It can receive and process user voice commands in real time, and adopts advanced voice recognition algorithms to recognize different users' voice inputs with high accuracy.

[0073] Command matching module: matches the recognized voice commands according to the preset command library, and transmits the matching results to the system processing module.

[0074] Alarm generation module: After the command is matched successfully, the system automatically generates the corresponding alarm information, including the alarm type, time, location and related description to ensure the integrity and accuracy of the information.

[0075] Anthropomorphic voice feedback module: The system uses natural language generation technology to convert alarm information into natural and fluent voice, and promptly feedback to users to enhance the human-computer interaction experience.

[0076] Work order system: responsible for receiving newly added work orders, inserting them into the work order list according to priority, and judging whether the newly added work orders need voice alarm. If necessary, insert them into the voice alarm work order list.

[0077] Work order alarm system: read the work order with the highest priority in the voice alarm work order list in the work order system, input it into the alarm generation module, generate alarm voice, and issue an alarm.

[0078] The specific steps of the alarm method of the alarm system are as follows:

[0079] Step 1: Add a new work order and determine whether this work order requires a voice alarm.

[0080] Step 2: If voice alarm is not required, insert the work order list.

[0081] Step 3: If voice alarm is required, insert the voice alarm work order list first.

[0082] Step 4: The work order alarm system reads the work order with the highest priority in the voice alarm work order list.

[0083] Step 5: Extract the main complaint of the work order and convert the text of the main complaint of the work order into voice broadcast to the staff through the TTS system.

[0084] Step 6: After receiving the message, the staff will give a voice command if there is any new instruction or reply.

[0085] Step 7: The system converts the voice command into text through STT technology.

[0086] Step 8: The text input command matching system performs command matching and generates a reply based on the command matching result.

[0087] Step 9: The generated reply is converted into voice reply to the staff through the TTS system.

[0088] When adding a new work order, sort the work orders by priority and determine whether it is a work order that can be alerted by voice. If so, insert it into the voice alert work order list according to the priority.

[0089] 1. Instruction set design:

[0090] In the work order alert system, there will be an instruction set, which will be continuously expanded by technicians based on failure cases after the work order alert system is triggered. This instruction set will refer to the context, preset the possible responses of the staff, the services that may be required, and the services that can be provided by the staff. Figure 4 .

[0091] 2. Overlap rate calculation algorithm based on the relative position of text:

[0092] a) When we get two texts, we need to calculate the text overlap rate of the two texts, one is text t1 and the other is text t2, perform text segmentation on the two texts, and generate keyword vocabulary lists. The lengths of the two vocabulary lists are n1 and n2 respectively.

[0093] b) Find the intersection of the two vocabulary lists and get the overlapping vocabulary. The length of the vocabulary list of the overlapping vocabulary is n, then the overlap ratio is overlap_alpha = n / ((n1+n2) / 2).

[0094] c) Scan the words in the overlapping vocabulary lists, calculate the relative positions of the i-th word in the two vocabulary lists as l1 and l2, and then calculate the relative position difference as beta_i = |l1-l2|,

[0095] d) Calculate the overlap rate algorithm based on the relative position of the text:

[0096] overlap=overlap_alpha*(n-beta_1-beta_2-…-beta_i-…-beta_n) / n

[0097] e) The text overlap rate based on the instruction is (this text overlap rate is mainly based on the characteristics of the instruction. In fact, as long as the instruction is found in the target text, the task is completed):

[0098] overlap_alpha2=n / n2

[0099] f) Algorithm for overlapping ratio of relative position based on instructions:

[0100] First, find the first and last positions of the words in the overlapping vocabulary list in the word list of t1, and form a new list of words at the first position, the last position, and all the words in the middle. Then, calculate the relative position difference beta2_i based on this new list:

[0101] overlap2=overlap_alpha2*(n-beta2_1-beta2_2-…-beta2_i-…-beta2_n) / n

[0102] 3. Instruction matching algorithm:

[0103] a) First, after receiving the voice-to-text, this text is q1.

[0104] b) Traverse each instruction in the instruction set Inst. If the text overlap rate of instruction I and q1 exceeds 90%, the instruction is matched successfully and the traversal ends.

[0105] c) If the text overlap ratio overlap_alpha2 between instruction I and q1 exceeds 80%, the instruction matching is successful, but the traversal continues.

[0106] d) If the text overlap ratio overlap2 between instruction I and q1 exceeds 70%, the instruction is matched successfully, but the traversal continues.

[0107] e) If no suitable instruction is matched after the instruction matching is completed, the matching fails.

[0108] f) If the match is successful based on the text overlap rate, the matched instruction is semantically matched with the original text. The original text is converted into a sentence vector through the embedding algorithm, and the distance is calculated with the instruction vector of the instruction. When the semantics also match, the instruction is successfully matched.

[0109] g) If the match is successful based on the text overlap ratio, but the semantic match is unsuccessful, the command match is sub-successful

[0110] h) If there is more than one instruction that matches successfully or sub-successfully, all instructions that match successfully will be considered valid and output. If all instructions are sub-successful, the instruction with the highest matching degree will be output as sub-successful.

[0111] 4. Double verification mechanism for text-to-speech system

[0112] In order to ensure the accuracy of the text-to-speech conversion and to avoid major errors when reporting to the staff, our text-to-speech system has designed a double verification mechanism:

[0113] a) Capture the highest priority work order from the voice alarm work order list and process it

[0114] b) First, capture the main complaint of the work order and convert the text into speech using text-to-speech (TTS) technology.

[0115] c) The converted voice to be broadcast is converted into text again using the speech-to-text model. In this way, we get two texts, one is the work order complaint text t1, and the other is the text t2 of the voice to be broadcast after speech-to-text conversion.

[0116] d) Match the two texts and calculate the text overlap ratio

[0117] e) When the overlap is less than 90%, it is considered that the model has certain problems when processing the text TTS or STT task, and the work order is returned to a non-voice alarm emergency work order, and the abnormal situation is pushed to the technical staff for processing.

[0118] f) When the overlap rate is greater than 90%, it is considered that the overlap rate is high. At this time, the two texts are converted into sentence vectors, and the cosine distance between the two sentence vectors is calculated. If the distance is small enough, the voice will give a normal alarm. If not, the work order content is greatly distorted due to some words in the method, which is also considered an invalid conversion and the work order is returned. Figure 2

[0119] 5. Text-to-speech (TTS) technology:

[0120] a) First, convert the text into phoneme text through a phoneme dictionary (the main reason for this is that although many characters are the same, the pronunciation may be different in different scenarios, or in different texts, and in different words. Therefore, before converting text into speech, the text should be converted into phoneme text first).

[0121] b) The next step is to predict the duration of each phoneme. This can be done effectively using a segmentation model. The input for segmentation model training is an audio and the phoneme text corresponding to the audio. The output is the starting position of each phoneme in the audio. The trained phoneme duration model can output the duration of each phoneme.

[0122] c) Fundamental frequency prediction model: The fundamental frequency prediction model first predicts whether a phoneme is pronounced, and if so, predicts its fundamental frequency. Figure 1

[0123] d) Audio synthesis model. The audio synthesis model synthesizes the final speech text based on the previous phoneme text, predicting the duration of each phoneme and the fundamental frequency of each phoneme.

[0124] 6. Speech-to-text STT technology:

[0125] a) Speech preprocessing: pre-emphasis, framing, windowing, short-time Fourier transform (STFT), mel filtering, and de-averaging. After such a series of operations, the speech will be divided into small segments. Theoretically, 10ms of speech information is one segment. So a long segment of speech information will eventually be divided into x1, x2, x3, ..., x n This is a sequence of information fragments. The output is Chinese characters or English letters, such as y=( <sos>,y1,...,y S , <eos>),y S It is any character within the prediction range, such as English letters, Chinese characters, etc.

[0126] b) The output is conditional probability:

[0127] [(y|x)=Π t P(y t |x,y ＜t )

[0128] x is the sound, y is ＜t is the text before the letter or word to be predicted, y i It is any letter or word in y.

[0129] train:

[0130] c) Model architecture: Both seq2seq models are available, such as Listen, Attend and Spell, which is an encoder-decoder model. First, the segmented speech fragments will pass through a pyramid-shaped BLSTM, and then output the feature vector set h, and then pass through a temporal network to generate an attention vector based on h and the intermediate output of the model, and finally decode to get y t , and y t It will also be used as the next prediction y t+1 Input.

[0131] 7. Voice command processing algorithm

[0132] If the staff has further needs, or simply wants to tell the work order alarm system that he has received the instructions, these can be responded directly by voice to improve work efficiency:

[0133] a) Listen to voice commands and convert them into text using STT technology

[0134] b) After receiving the staff's reply, the work order alarm system will match the instructions.

[0135] c) After the command is successfully matched, a reply will be generated, and the TTS text-to-speech process will be used to generate voice and reply to the staff

[0136] d) If the command matches successfully, further voice confirmation is performed to ask the staff whether this is the command they want to issue. Figure 3 .

[0137] It should be noted that the above embodiments are not intended to limit the protection scope of the present invention, and equivalent changes or substitutions made on the basis of the above technical solutions all fall within the protection scope of the claims of the present invention.< / eos> < / sos> < / eos> < / sos>

Claims

1. A robot anthropomorphic voice work order alarm system based on voice command matching technology, characterized in that: The alarm system comprises: Voice recognition module: It can receive and process user voice commands in real time, and adopts advanced voice recognition algorithms to recognize different users' voice inputs with high accuracy. Command matching module: matches the recognized voice commands according to the preset command library, and transmits the matching results to the system processing module. Alarm generation module: After the command is matched successfully, the system automatically generates the corresponding alarm information, including the alarm type, time, location and related description to ensure the integrity and accuracy of the information. Anthropomorphic voice feedback module: The system uses natural language generation technology to convert alarm information into natural and fluent voice, and promptly feedback to users to enhance the human-computer interaction experience. Work order system: responsible for receiving newly added work orders, inserting them into the work order list according to priority, and judging whether the newly added work orders need voice alarm. If necessary, insert them into the voice alarm work order list. Work order alarm system: read the work order with the highest priority in the voice alarm work order list in the work order system, input it into the alarm generation module, generate alarm voice, and issue an alarm.

2. The anthropomorphic voice work order alarm system based on voice command matching technology according to claim 1 is characterized in that: The specific steps of the alarm method of the alarm system are as follows: Step 1: Add a new work order and determine whether this work order requires a voice alarm. Step 2: If voice alarm is not required, insert the work order list. Step 3: If voice alarm is required, insert the voice alarm work order list first. Step 4: The work order alarm system reads the work order with the highest priority in the voice alarm work order list. Step 5: Extract the main complaint of the work order and convert the text of the main complaint of the work order into voice broadcast to the staff through the TTS system. Step 6: After receiving the message, the staff will give a voice command if there is any new instruction or reply. Step 7: The system converts the voice command into text through STT technology. Step 8: The text input command matching system performs command matching and generates a reply based on the command matching result. Step 9: The generated reply is converted into voice reply to the staff through the TTS system.

3. The anthropomorphic voice work order alarm system based on voice command matching technology according to claim 1 is characterized in that: In the alarm system, an overlap rate calculation algorithm based on the relative position of text is adopted, which is as follows a) When we get two texts, we need to calculate the text overlap rate of the two texts. One is text t1 and the other is text t2. We segment the two texts and generate keyword lists. The lengths of the two word lists are n1 and n2 respectively. b) Find the intersection of the two vocabulary lists and get the overlapping vocabulary. The length of the vocabulary list of the overlapping vocabulary is n, so the overlap rate is overlap_alpha = n / ((n1+n2) / 2). c) Scan the words in the overlapping vocabulary lists, calculate the relative positions of the i-th word in the two vocabulary lists as l1 and l2, and then calculate the relative position difference as beta_i = |l1-l2|, d) Calculate the overlap rate algorithm based on the relative position of the text: overlap=overlap_alpha*(n-beta_1-beta_2-…-beta_i-…-beta_n) / n, e) The text overlap rate based on the instruction is: overlap_alpha2=n / n2 f) Algorithm for overlapping ratio of relative position based on instructions: First, find the first and last positions of the words in the overlapping vocabulary list in the word list of t1, and form a new list of words at the first position, the last position, and all the words in the middle. Then, calculate the relative position difference beta2_i based on this new list: overlap2=overlap_alpha2*(n-beta2_1-beta2_2-…-beta2_i-…-beta2_n) / n.

4. The anthropomorphic voice work order alarm system based on voice command matching technology according to claim 1 is characterized in that: The alarm system uses a command matching algorithm, as follows: a) First, after receiving the voice-to-text conversion, this text is q1, b) Traverse each instruction in the instruction set Inst. If the text overlap rate of instruction I and q1 exceeds 90%, the instruction is matched successfully and the traversal ends. c) If the text overlap ratio overlap_alpha2 of instruction I and q1 exceeds 80%, the instruction matches successfully, but the traversal continues. d) If the text overlap ratio overlap2 of instruction I and q1 exceeds 70%, the instruction matches successfully, but the traversal continues. e) If no suitable instruction is found after instruction matching, the matching fails. f) If the match is successful based on the text overlap rate, the matched instruction is semantically matched with the original text. The original text is converted into a sentence vector through the embedding algorithm, and the distance is calculated with the instruction vector of the instruction. If the semantics also match, the instruction is successfully matched. g) If the match is successful based on the text overlap ratio, but the semantic match is unsuccessful, the instruction match is sub-successful. h) If there is more than one instruction that matches successfully or sub-successfully, all instructions that match successfully will be considered valid and output. If all instructions are sub-successful, the instruction with the highest matching degree will be output as sub-successful.

5. The anthropomorphic voice work order alarm system based on voice command matching technology according to claim 1 is characterized in that: The alarm system uses a double verification mechanism of text-to-speech system, as follows: a) Capture the highest priority work order from the voice alarm work order list and process it. b) First, capture the main complaint of the work order and convert the text into speech using the text-to-speech (TTS) technology. c) The converted voice to be broadcast is converted into text again using the speech-to-text model. In this way, we get two texts, one is the work order complaint text t1, and the other is the text t2 of the voice to be broadcast after speech-to-text conversion. d) Match the two texts and calculate the text overlap ratio. e) When the overlap is less than 90%, it is considered that the model has some problems in processing the text TTS or STT task, and the work order is returned to a non-voice alarm emergency work order, and the abnormal situation is pushed to the technician for processing. When the overlap rate is greater than 90%, it is considered that the overlap rate is high. At this time, the two texts are converted into sentence vectors respectively, and then the cosine distance between the two sentence vectors is calculated. If the distance is small enough, the voice will give a normal alarm. If not, the work order content has been greatly distorted due to some individual words in the method, which is also considered an invalid conversion and the work order is returned.

6. The anthropomorphic voice work order alarm system based on voice command matching technology according to claim 1 is characterized in that: The alarm system uses text-to-speech TTS technology, as follows: a) First, convert the text into phoneme text through the phoneme dictionary, b) The next step is to predict the duration of each phoneme. This can be done effectively using a segmentation model. The input for segmentation model training is an audio and the phoneme text corresponding to the audio. The output is the starting position of each phoneme in the audio. The trained phoneme duration model can output the duration of each phoneme. c) Fundamental frequency prediction model: The fundamental frequency prediction model first predicts whether a phoneme is pronounced. If so, it predicts its fundamental frequency. d) Audio synthesis model. The audio synthesis model synthesizes the final speech text based on the previous phoneme text, predicting the duration of each phoneme and the fundamental frequency of each phoneme.

7. The anthropomorphic voice work order alarm system based on voice command matching technology according to claim 1 is characterized in that: The alarm system uses speech-to-text STT technology, as follows: a) Speech preprocessing: pre-emphasis, framing, windowing, short-time Fourier transform (STFT), mel filtering, and de-averaging. After such a series of operations, the speech will be divided into small segments. Theoretically, 10ms of speech information is one segment, so a long segment of speech information will eventually be divided into x1, x2, x3, ..., x n The output of this information fragment sequence is Chinese characters or English letters, y=( <sos>,y1,...,y S , <eos>),y S is any character in the prediction range, such as English letters, Chinese characters,< / eos> < / sos> b) The output is conditional probability: P(y|x)=Π t P(yt|x,y ＜t ), x is the sound, y is ＜t is the text before the letter or word to be predicted, y i is any letter or word in y, train: c) Model architecture: Both seq2seq models are available. Listen, Attend and Spell is an encoder-decoder model. First, the segmented speech fragments pass through a pyramid-shaped BLSTM, and then output the feature vector set h. Then, it passes through a temporal network to generate an attention vector based on h and the intermediate output of the model. Finally, it is decoded to obtain y. t , and y t It will also be used as the next step to predict y t+1 Input.

8. The anthropomorphic voice work order alarm system based on voice command matching technology according to claim 1 is characterized in that: The alarm system adopts a voice command processing algorithm, which is as follows: a) Listen to the voice commands and convert them into text content using STT technology. b) After receiving the staff's reply, the work order alarm system will match the instructions. c) After the command is successfully matched, a reply will be generated, and the TTS text-to-speech process will be used to generate voice and reply to the staff d) If the command matches successfully, further voice confirmation is performed to ask the staff whether this is the command they want to issue.

Images