An AI voice linkage-based keyword driving method

Abstract

This invention specifically relates to a keyword-driven method based on AI voice interaction, encompassing the fields of speech recognition and intelligent interaction technologies. The method includes: continuously acquiring speech signals played by an instruction scheduling system deployed in the production control area via an audio acquisition device; performing frame-by-frame processing and filter bank energy feature extraction on the speech signals to generate an acoustic feature sequence. In this invention, the audio acquisition device acquires the speech signals and utilizes an acoustic channel as an information transmission medium across security partitions. This solves the technical problem that network isolation prevents multiple service systems from directly communicating through network data interfaces, achieving the technical effect of cross-partition information transmission without altering the security partition network architecture.

Description

Technical Field

[0001] This invention relates to the field of speech recognition and intelligent linkage technology, and in particular to a keyword-driven method based on AI speech linkage. Background Technology

[0002] In the power plant information system, based on the safety zoning principle stipulated in the safety protection regulations for power secondary systems, the command and dispatch system is deployed in safety zone I of the production control area, while the flood control command system and the OnCall safety early warning system are deployed in the management information area. Network isolation exists between different safety zones, preventing direct communication between various business systems via network data interfaces, resulting in a lack of a unified linkage mechanism between multiple systems.

[0003] In the existing business process, the command and dispatch system notifies the dispatcher via voice and text commands before remotely starting power generation. The dispatcher needs to pay attention to the start-up command in real time and manually log in to the flood control command system to enter a broadcast notification after receiving the command to remind people in the downstream riverbank area to evacuate. The existing technology uses manual operation to realize cross-system information transmission and uses a keyword matching method based on a fixed vocabulary for string rule comparison.

[0004] The aforementioned existing technologies have the following technical problems: dispatchers may delay in discovering the start-up command due to objective reasons, resulting in a long response delay; there is a lack of automated collaborative scheduling capabilities between multiple independent business systems; the acoustic environment of the power plant command center is complex, and the superposition of equipment noise and multiple sound sources interferes with speech recognition; traditional fixed word list matching methods cannot handle speech variations such as connected speech and non-standard pronunciation. Summary of the Invention

[0005] The purpose of this invention is to propose a keyword-driven method based on AI voice interaction in order to solve the above-mentioned problems.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: A keyword-driven method based on AI voice interaction includes: The audio acquisition device continuously acquires the voice signal played by the command scheduling system deployed in the production control area, performs frame processing and filter bank energy feature extraction on the voice signal, and generates an acoustic feature sequence. The acoustic feature sequence is input into a pre-trained bidirectional long short-term memory network for temporal modeling to obtain the phoneme posterior probability vector corresponding to each frame in the acoustic feature sequence and generate the phoneme posterior probability sequence. The phoneme posterior probability sequence is decoded based on the connection time-series classification criterion to obtain the recognizable character sequence; The identified character sequence is matched with a preset keyword mapping table. When the identified character sequence contains a target keyword in the keyword mapping table, a linkage instruction is generated according to the linkage rule corresponding to the target keyword. Based on the target system identifier in the linkage instruction, the linkage instruction is distributed to the corresponding target system through the internal communication link of the management information area. After receiving the linkage instruction, the target system executes the corresponding linkage operation.

[0007] Preferably, before performing frame segmentation processing on the speech signal, the method further includes: The acquired raw speech waveform is processed by speech activity detection to identify time segments containing valid speech signals, remove silent segments and pure noise segments, and obtain valid speech segments. The effective speech segment is subjected to frequency domain noise suppression processing. The gain of the spectral components of the effective speech segment is adjusted based on noise power spectrum estimation to obtain the noise-reduced speech waveform. The noise-reduced speech waveform is used as the input for subsequent frame segmentation processing.

[0008] Preferably, the step of performing frame segmentation processing and filter bank energy feature extraction on the speech signal to generate an acoustic feature sequence includes: The original speech waveform is divided into frames according to preset frame length and frame shift parameters to obtain a speech frame sequence consisting of multiple speech frames, with overlapping areas between adjacent frames; A Hamming window function is applied to each frame signal in the speech frame sequence, and a short-time Fourier transform is performed to obtain the spectral representation of each frame. The spectral representation is then energy-integrated through a Mel-scale triangular filter bank and the logarithm is taken to obtain the filter bank energy feature vector corresponding to each frame. The energy feature vectors of the filter bank in each frame are arranged in chronological order to form an acoustic feature sequence.

[0009] Preferably, the bidirectional Long Short-Term Memory (LSTM) network consists of a forward LSTM layer, a backward LSTM layer, and a linear mapping layer. The forward LSTM layer processes the features of each frame in forward chronological order and outputs a forward hidden state at each time step. The backward LSTM layer processes the features of each frame in reverse chronological order and outputs a backward hidden state at each time step. The forward and backward hidden states at the same time step are concatenated to obtain the joint hidden state at that time step. The joint hidden state is input to the linear mapping layer, which maps the joint hidden state into a real vector with a dimension equal to the total number of phoneme categories. The real vector is then normalized using a softmax activation function to obtain the posterior probability vector of the phonemes for the corresponding frame.

[0010] Preferably, each LSTM unit within the forward LSTM layer and the backward LSTM layer adopts a three-gated structure, including a forget gate, an input gate, and an output gate; during the update calculation at each time step: The forget gate calculates the forgetting coefficient based on the hidden state of the previous time step and the input features of the current frame, using the forget gate weight matrix and the forget gate bias vector through the sigmoid activation function. The input gate calculates the input coefficients based on the hidden state of the previous time step and the input features of the current frame through the input gate weight matrix and the input gate bias vector using the sigmoid activation function. At the same time, it calculates the candidate cell state through the candidate cell state weight matrix and the candidate cell state bias vector using the hyperbolic tangent activation function. The cell state is updated by adding the element-wise product of the forgetting coefficient and the cell state at the previous time step to the element-wise product of the input coefficient and the candidate cell state. The output gate calculates the output coefficients based on the hidden state of the previous time step and the input features of the current frame using the output gate weight matrix and the output gate bias vector through the sigmoid activation function. The output coefficients are then multiplied element-wise with the updated cell state by the hyperbolic tangent activation function to obtain the hidden state of the current time step.

[0011] Preferably, the bidirectional long short-term memory network adopts a multi-layer stacked structure, including at least two bidirectional long short-term memory layers; the joint hidden state sequence of the lower-level bidirectional long short-term memory layers serves as the input feature sequence of the higher-level bidirectional long short-term memory layers, and is passed layer by layer for higher-level temporal feature abstraction; the calculation of the phoneme posterior probability vector is based on the joint hidden state output by the highest-level bidirectional long short-term memory layer. Before inputting the acoustic feature sequence into the bidirectional long short-term memory network, the filter bank energy feature vectors of each frame in the acoustic feature sequence are Z-score normalized.

[0012] Preferably, the decoding process of the phoneme posterior probability sequence based on the connection time-series classification criterion to obtain the recognizable character sequence includes: Decoding is performed using a beam search strategy combined with hot word enhancement processing; a hot word list containing industry-specific terms is pre-established, which includes key terms related to the instruction scheduling scenario and their corresponding phoneme sequences; During the beam search process, multiple candidate paths are maintained, and each candidate path has a corresponding cumulative log probability score. When the current partial phoneme sequence of a candidate path matches the prefix phoneme sequence of a word in the hot word list, the cumulative score of the candidate path is increased by a preset hot word reward score. After the beam search is completed, the candidate path with the highest cumulative score is selected, and a folding operation is performed on the candidate path, that is, merging consecutively repeated identical symbols and removing whitespace characters to obtain the recognized character sequence.

[0013] Preferably, the step of performing keyword matching further includes: Obtain the frame-level probability score corresponding to the position of the target keyword in the recognized character sequence, and calculate the average of the frame probability scores corresponding to the target keyword as the keyword confidence score; The keyword confidence score is compared with a preset confidence threshold. The linkage instruction is generated only when the keyword confidence score is greater than or equal to the confidence threshold. When the keyword confidence score is less than the confidence threshold, no linkage instruction is generated, and the current recognition result is recorded for subsequent verification. The confidence threshold is determined by traversing the candidate threshold interval on a verification set containing recordings of actual acoustic environments, calculating the false trigger rate and missed trigger rate of the linkage command for each candidate threshold, and selecting the candidate value that minimizes the weighted comprehensive index of the false trigger rate and the missed trigger rate.

[0014] Preferably, after distributing the linkage instruction to the corresponding target system, the method further includes: Start a timeout timer. If the execution status information returned by the target system is not received within the preset timeout period, the linkage instruction will be resent. The number of resentments shall not exceed the preset maximum number of resentments. If no execution status information is received after reaching the maximum number of retransmissions, a linkage anomaly alarm is generated, the linkage anomaly alarm is recorded in the local log, and pushed to the operation and maintenance personnel through a preset notification channel. Each record in the keyword mapping table contains the target keyword, the associated target system identifier, and the corresponding linkage instruction template; the linkage instruction contains the target system identifier, instruction type, instruction content, and trigger time information.

[0015] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are: 1. This invention acquires voice signals through an audio acquisition device and uses an acoustic channel as an information transmission medium across security partitions. It solves the technical problem that multiple service systems cannot communicate directly through network data interfaces due to network isolation, and achieves the technical effect of cross-partition information transmission without changing the security partition network architecture.

[0016] 2. This invention solves the technical problem of interference caused by equipment noise and multiple sound source superposition on the temporal relationship of speech features in the complex acoustic environment of a power plant command center by using a bidirectional long short-term memory network for bidirectional temporal modeling, and achieves the technical effect of improving the accuracy of phoneme posterior probability sequences in complex noise environments; by using a connection-time classification criterion for end-to-end decoding, it solves the technical problem that traditional fixed vocabulary matching methods cannot handle speech variations such as connected speech and non-standard pronunciation, and achieves the technical effect of adaptively handling keyword recognition under different speech rates and accents. Attached Figure Description

[0017] Further details, features, and advantages of this application are disclosed in the following description of exemplary embodiments in conjunction with the accompanying drawings, in which: Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0018] Several embodiments of this application will now be described in more detail with reference to the accompanying drawings to enable those skilled in the art to implement this application. This application may be embodied in many different forms and for various purposes and should not be limited to the embodiments set forth herein. These embodiments are provided to make this application thorough and complete, and to fully convey the scope of this application to those skilled in the art. The embodiments described do not limit this application.

[0019] Unless otherwise defined, all terms used herein (including technical and scientific terms) shall have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains. It will be further understood that terms such as those defined in commonly used dictionaries shall be interpreted as having a meaning consistent with their meaning in the relevant field and / or the context of this specification, and shall not be interpreted in an idealized or overly formal sense unless expressly defined herein.

[0020] Example 1 Its specific implementation method is combined with the appendix Figure 1 Please provide a detailed explanation.

[0021] In this embodiment, it includes: A hydroelectric power plant (hereinafter referred to as "Plant A") deployed local voice processing terminal equipment in its command center to achieve automatic linkage between the command and dispatch system, flood control command system, and OnCall safety early warning system. Plant A is located in a cascade hydroelectric power station group in a certain river basin. There are residential areas along its downstream riverbank, and downstream personnel must be notified to evacuate in advance via flood control broadcasts before the generator units are started. On the morning of July 15, 20XX, the dispatch center issued a start-up command through the command and dispatch system. The audio acquisition device of the local voice processing terminal continuously monitored the acoustic environment of the command center, processed the voice signal played by the command and dispatch system in real time, and triggered the flood control broadcast linkage process.

[0022] The audio acquisition device of the local voice processing terminal continuously acquires the voice signal broadcast by the loudspeaker of the command and dispatch system in the command center at a sampling rate of 16000Hz. At 09:47:32 on July 15, 20XX, the audio acquisition device captured a voice signal lasting approximately 2.1 seconds. After voice activity detection, the original voice waveform identified the valid voice segment (starting at 0.08 seconds and ending at 2.06 seconds). After removing the first and last silent segments, it entered frequency domain noise suppression processing to obtain the noise-reduced voice waveform.

[0023] The denoised speech waveform is divided into frames with a frame length of 25 milliseconds and a frame shift of 10 milliseconds. The effective speech segment duration is approximately 1.98 seconds, and the number of frames is... The estimates are as follows: ; A total of 196 speech frames were generated. After applying a Hamming window to each frame, a short-time Fourier transform was performed, and then the energy was integrated and the logarithm was taken through 80 Mel-scale triangular filters to obtain an 80-dimensional Fbank feature vector for each frame, which finally constituted an acoustic feature sequence of 196 frames.

[0024] Table 1 Key parameters and data for acoustic feature extraction process

[0025] Before feeding the acoustic feature sequence into the bidirectional long short-term memory network, the 80-dimensional Fbank feature vectors of 196 frames are Z-score normalized. Taking the 3rd dimension feature component of frame 98 as an example, let the mean of this dimension over the entire sequence be... The standard deviation is The standardized value is: ; The standardized acoustic feature sequence is input into a two-layer stacked bidirectional long short-term memory network. At each time step in the network... The first layer of forward LSTM units sequentially performs the forget gate, input gate, cell state update, and output gate calculations. (Based on frame 98...) Taking a certain example, the forget gate output suppresses historical interference information introduced by background device noise, the input gate selectively writes the speech features of the current frame into the cell state, and the output gate combines the updated cell state to obtain the forward hidden state. After processing in reverse chronological order by the inverse LSTM layer, the inverse hidden state is obtained. The two are concatenated to obtain the first layer joint hidden state, which serves as the input to the second layer bidirectional long short-term memory layer. The joint hidden state output by the second layer is processed by a linear mapping layer and a softmax function to obtain the phoneme posterior probability vector of the 98th frame. After all 196 frames are processed, a phoneme posterior probability sequence is generated. The dimension of the probability vector for each frame is the number of phoneme categories plus one (including whitespace). In this implementation, the total number of phoneme categories is set to 218, so the dimension of the probability vector for each frame is 219.

[0026] Table 2 shows examples of posterior probability vectors for phonemes in some frames (displaying the three phoneme categories with the highest probabilities).

[0027] Using the phoneme posterior probability sequence as input, a temporal classification decoder employs a beam search strategy (beam width set to 10) to select phonemes or whitespace characters with higher probabilities for each frame, generating frame-level output paths. Simultaneously, a pre-established hot word list contains entries such as "power on," "power off," "flood prevention," and "flood discharge," along with their corresponding phoneme sequences. During the beam search process, when the current prefix phoneme sequence of a candidate path matches the prefix phoneme sequence of the hot word "power on," a hot word bonus score (preset to 0.35) is added to the cumulative logarithmic probability score of that candidate path, giving it higher priority in the candidate path ranking.

[0028] After the beam search is completed, the candidate path with the highest cumulative score is selected to perform a folding operation: first, consecutive repeating identical phoneme symbols are merged, then whitespace symbols are removed, the folded phoneme sequence is mapped to the corresponding Chinese characters, and finally a recognized character sequence is generated.

[0029] Table 3 Examples of Connection Temporal Classification Decoding Folding Process

[0030] The system identifies the character sequence "Please start generating power immediately for Unit XX" and performs substring matching search against a preset keyword mapping table, matching the target keyword "start". The system reads the linkage instruction template corresponding to "start" and obtains the frame-level probability score of the frame segment (frames 90 to 115) corresponding to the two characters "start" in the identified character sequence, and calculates the keyword confidence score.

[0031] Suppose that the highest probability scores of the corresponding phonemes in each of the 26 frames from frame 90 to frame 115 are known sequentially, and their average values ​​are calculated as follows (for simplification, representative values ​​are used): ; The confidence threshold is preset to 0.65, because If the confidence level condition is met, execute the linkage instruction generation operation. Fill the linkage instruction template with the target keyword "power on" and the current timestamp "July 15, 20XX, 09:47:34" to generate the linkage instruction.

[0032] Table 4 Keyword Mapping Table Matching Results and Linkage Instruction Content

[0033] Based on the target system identifier in the linkage command, the local voice processing terminal device determines the distribution destination address to be the internal IP address (IP-HF-001) of the flood control command system within the management information area's local area network. It then encapsulates the linkage command into a communication message format that the flood control command system can parse and sends it through the management information area's internal communication link. After receiving and parsing the message, the flood control command system automatically triggers the preset flood control broadcast content and broadcasts it to the downstream riverbank area. After the broadcast is completed, the flood control command system returns execution status information to the local voice processing terminal device. Before the timeout timer (preset timeout of 5 seconds) is triggered, the local voice processing terminal device receives the execution success status information at 09:47:36 on July 15, 20XX, and the linkage execution result is recorded in the local log.

[0034] Table 5 Record of Linkage Command Distribution and Execution Results

[0035] In this example, the data starts from the initial raw speech waveform. After speech activity detection and noise suppression, the effectively denoised speech is obtained. Frame segmentation and Fbank feature extraction generate 196 frames of 80-dimensional acoustic feature sequences. After Z-score standardization, the data is input into a bidirectional long short-term memory network to generate a 219-dimensional phoneme posterior probability sequence frame by frame. The probability sequence is then folded and mapped by a temporal classification decoder combined with hot word enhancement beam search, outputting the recognized character sequence "Please start generating power immediately for Unit XX". After the keyword matching hits "start", and passes the confidence check (0.7143≥0.65), the linkage instruction template is automatically filled to generate a flood control broadcast notification instruction for the flood control command system. Finally, the instruction is distributed to the flood control command system for execution via the internal communication link of the management information area. The data in the entire process maintains strict logical connection and consistency between each step, realizing end-to-end automated processing from voice acquisition to multi-system linkage.

[0036] The foregoing has only described certain exemplary embodiments of the present invention by way of illustration. Undoubtedly, those skilled in the art can modify the described embodiments in various ways without departing from the spirit and scope of the present invention. Therefore, the foregoing drawings and descriptions are illustrative in nature and should not be construed as limiting the scope of protection of the claims of the present invention.

[0037] It should be understood that in the various embodiments of this application, the order of the above-mentioned processes does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0038] In addition, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0039] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A keyword-driven method based on AI voice interaction, characterized in that, include: The audio acquisition device continuously acquires the voice signal played by the command scheduling system deployed in the production control area, performs frame processing and filter bank energy feature extraction on the voice signal, and generates an acoustic feature sequence. The acoustic feature sequence is input into a pre-trained bidirectional long short-term memory network for temporal modeling to obtain the phoneme posterior probability vector corresponding to each frame in the acoustic feature sequence and generate the phoneme posterior probability sequence. The phoneme posterior probability sequence is decoded based on the connection time-series classification criterion to obtain the recognizable character sequence; The identified character sequence is matched with a preset keyword mapping table. When the identified character sequence contains a target keyword in the keyword mapping table, a linkage instruction is generated according to the linkage rule corresponding to the target keyword. Based on the target system identifier in the linkage instruction, the linkage instruction is distributed to the corresponding target system through the internal communication link of the management information area. After receiving the linkage instruction, the target system executes the corresponding linkage operation.

2. The keyword-driven method based on AI voice interaction according to claim 1, characterized in that, Before performing frame segmentation processing on the speech signal, the method further includes: The acquired raw speech waveform is processed by speech activity detection to identify time segments containing valid speech signals, remove silent segments and pure noise segments, and obtain valid speech segments. The effective speech segment is subjected to frequency domain noise suppression processing. The gain of the spectral components of the effective speech segment is adjusted based on noise power spectrum estimation to obtain the noise-reduced speech waveform. The noise-reduced speech waveform is used as the input for subsequent frame segmentation processing.

3. The keyword-driven method based on AI voice interaction according to claim 1, characterized in that, The speech signal is subjected to frame segmentation and filter bank energy feature extraction to generate an acoustic feature sequence, including: The original speech waveform is divided into frames according to preset frame length and frame shift parameters to obtain a speech frame sequence consisting of multiple speech frames, with overlapping areas between adjacent frames; A Hamming window function is applied to each frame signal in the speech frame sequence, and a short-time Fourier transform is performed to obtain the spectral representation of each frame. The spectral representation is then energy-integrated through a Mel-scale triangular filter bank and the logarithm is taken to obtain the filter bank energy feature vector corresponding to each frame. The energy feature vectors of the filter bank in each frame are arranged in chronological order to form an acoustic feature sequence.

4. The keyword-driven method based on AI voice interaction according to claim 1, characterized in that, The bidirectional Long Short-Term Memory (LSTM) network consists of a forward LSTM layer, a backward LSTM layer, and a linear mapping layer. The forward LSTM layer processes the features of each frame in forward chronological order and outputs a forward hidden state at each time step. The backward LSTM layer processes the features of each frame in reverse chronological order and outputs a backward hidden state at each time step. The forward and backward hidden states at the same time step are concatenated to obtain the joint hidden state for that time step. The joint hidden state is then input to the linear mapping layer, which maps the joint hidden state into a real vector with a dimension equal to the total number of phoneme categories. This real vector is then normalized using a softmax activation function to obtain the posterior probability vector of the phonemes for the corresponding frame.

5. The keyword-driven method based on AI voice interaction according to claim 1, characterized in that, Each LSTM unit within both the forward and backward LSTM layers employs a three-gated structure, including a forget gate, an input gate, and an output gate; during the update calculation at each time step: The forget gate calculates the forgetting coefficient based on the hidden state of the previous time step and the input features of the current frame through the forget gate weight matrix and the forget gate bias vector and the sigmoid activation function. The input gate calculates the input coefficients based on the hidden state of the previous time step and the input features of the current frame through the input gate weight matrix and the input gate bias vector using the sigmoid activation function. At the same time, it calculates the candidate cell state through the candidate cell state weight matrix and the candidate cell state bias vector using the hyperbolic tangent activation function. The cell state is updated by adding the element-wise product of the forgetting coefficient and the cell state at the previous time step to the element-wise product of the input coefficient and the candidate cell state. The output gate calculates the output coefficients based on the hidden state of the previous time step and the input features of the current frame using the output gate weight matrix and the output gate bias vector through the sigmoid activation function. The output coefficients are then multiplied element-wise with the updated cell state by the hyperbolic tangent activation function to obtain the hidden state of the current time step.

6. The keyword-driven method based on AI voice interaction according to claim 1, characterized in that, The bidirectional long short-term memory network adopts a multi-layer stacked structure, containing at least two bidirectional long short-term memory layers; the joint hidden state sequence of the lower-level bidirectional long short-term memory layers serves as the input feature sequence of the higher-level bidirectional long short-term memory layers, which is passed layer by layer and subjected to higher-level temporal feature abstraction. The calculation of the phoneme posterior probability vector is based on the joint hidden state output by the highest-level bidirectional long short-term memory layer; Before inputting the acoustic feature sequence into the bidirectional long short-term memory network, the filter bank energy feature vectors of each frame in the acoustic feature sequence are Z-score normalized.

7. The keyword-driven method based on AI voice interaction according to claim 1, characterized in that, Decoding the phoneme posterior probability sequence based on a connection-time classification criterion yields a sequence of recognizable characters, including: Decoding is performed using a beam search strategy combined with hot word enhancement processing; a hot word list containing industry-specific terms is pre-established, which includes key terms related to the instruction scheduling scenario and their corresponding phoneme sequences; During the beam search process, multiple candidate paths are maintained, and each candidate path has a corresponding cumulative log probability score. When the current partial phoneme sequence of a candidate path matches the prefix phoneme sequence of a word in the hot word list, the cumulative score of the candidate path is increased by a preset hot word reward score. After the beam search is completed, the candidate path with the highest cumulative score is selected, and a folding operation is performed on the candidate path, that is, merging consecutively repeated identical symbols and removing whitespace characters to obtain the recognized character sequence.

8. The keyword-driven method based on AI voice interaction according to claim 1, characterized in that, Keyword matching also includes: Obtain the frame-level probability score corresponding to the position of the target keyword in the recognized character sequence, and calculate the average of the frame probability scores corresponding to the target keyword as the keyword confidence score; The keyword confidence score is compared with a preset confidence threshold. The linkage instruction is generated only when the keyword confidence score is greater than or equal to the confidence threshold. When the keyword confidence score is less than the confidence threshold, no linkage instruction is generated, and the current recognition result is recorded for subsequent verification. The confidence threshold is determined by traversing the candidate threshold interval on a verification set containing recordings of actual acoustic environments, calculating the false trigger rate and missed trigger rate of the linkage command for each candidate threshold, and selecting the candidate value that minimizes the weighted comprehensive index of the false trigger rate and the missed trigger rate.

9. The keyword-driven method based on AI voice interaction according to claim 1, characterized in that, After distributing the linkage command to the corresponding target system, the process also includes: Start a timeout timer. If the execution status information returned by the target system is not received within the preset timeout period, the linkage instruction will be resent. The number of resentments shall not exceed the preset maximum number of resentments. If no execution status information is received after reaching the maximum number of retransmissions, a linkage anomaly alarm is generated, the linkage anomaly alarm is recorded in the local log, and pushed to the operation and maintenance personnel through a preset notification channel. Each record in the keyword mapping table contains the target keyword, the associated target system identifier, and the corresponding linkage instruction template; the linkage instruction contains the target system identifier, instruction type, instruction content, and trigger time information.

Images