Vehicle-mounted voice interaction control method and system, electronic device, and storage medium

Abstract

The present application relates to the technical field of voice interaction, and provides a vehicle-mounted voice interaction control method and system, electronic equipment and a storage medium, comprising: acquiring current cabin perception data in a vehicle, determining the number of people in the vehicle according to the cabin perception data; determining the current interaction rejection strategy based on the number of people in the vehicle; receiving target voice data input for a wake-up sound area, performing intent recognition on the target voice data to obtain a conversation intent; and generating a recall response instruction or a rejection interception instruction for the target voice data based on the conversation intent and the interaction rejection strategy. The present application dynamically adjusts the interaction rejection strategy by combining the number of people in the vehicle, and comprehensively judges by fusing the conversation intent, thereby realizing accurate differentiation and differential control of non-human-human conversation and human-human conversation in a complex vehicle-mounted environment, effectively reducing the disturbance to normal communication of users, and improving the accuracy and user experience of vehicle-mounted voice interaction.

Description

Technical Field

[0001] This invention relates to the field of voice interaction technology, and in particular to an in-vehicle voice interaction control method, system, electronic device and storage medium. Background Technology

[0002] With the widespread application of large language models in in-vehicle voice interaction systems, compared to the rigid and limited casual conversation capabilities of traditional small models, large models possess extensive and powerful generation capabilities, enabling them to respond broadly to user input. However, unlike large model dialogues on web pages or mobile devices, which have strong human-computer interaction windows or triggering conditions, voice-based interaction solutions in the in-vehicle interaction space inevitably require the determination of non-human-computer dialogue versus human-computer dialogue, thus needing to intervene with the large model at the appropriate time to reduce user confusion. Therefore, how to accurately distinguish between non-human-computer dialogue and human-computer dialogue in user voice input to avoid undue system intervention is a crucial problem that urgently needs to be solved.

[0003] Currently, the determination of dialogue types within vehicle cockpits typically relies solely on textual semantic analysis. The system judges based on the text content; if it is determined to be a typical non-human-computer dialogue, it directly rejects the dialogue and does not respond; if it is determined to be a human-computer dialogue, it allows the dialogue and a response is provided by a large model.

[0004] However, relying solely on plain text content for judgment is too simplistic. In complex in-vehicle interaction environments (such as when users express their subjective preferences or engage in casual conversation), due to the high degree of overlap in language expression, plain text analysis often fails to accurately discern the user's true interaction intent. This can easily lead to misjudgments between non-human-computer dialogue and human-computer dialogue, resulting in a significant risk of over-calling and causing disturbance to the user. Summary of the Invention

[0005] This invention provides a vehicle-mounted voice interaction control method, system, electronic device, and storage medium to address the shortcomings of existing vehicle-mounted voice interaction systems that rely solely on plain text for dialogue judgment, which can easily lead to false responses and frequent disruptions to normal user communication.

[0006] This invention provides a vehicle-mounted voice interaction control method, comprising: Acquire current cabin perception data and determine the number of occupants based on the cabin perception data; Based on the number of people inside the vehicle, determine the current interaction rejection strategy; Receive target speech data input to the wake-up voice zone, perform intent recognition on the target speech data, and obtain the dialogue intent; Based on the dialogue intent and the interaction rejection strategy, a recall response instruction or rejection interception instruction is generated for the target voice data, and a voice interaction response is executed according to the recall response instruction or silence is maintained according to the rejection interception instruction.

[0007] According to a vehicle-mounted voice interaction control method provided by the present invention, determining the current interaction rejection strategy based on the number of occupants in the vehicle includes: If the number of people in the vehicle is a single person, the interaction rejection strategy is determined to be the first rejection strategy. When there are multiple people inside the vehicle, the interaction rejection strategy is determined to be the second rejection strategy. The conditions under which the second denial strategy triggers denial-of-identity interception are stricter than those under which the first denial strategy triggers denial-of-identity interception.

[0008] According to the present invention, an in-vehicle voice interaction control method is provided, wherein the dialogue scenario corresponding to the dialogue intent includes at least one of a preset skill-oriented scenario, a strong human-computer dialogue scenario, a weak human-computer dialogue scenario, and a casual conversation scenario centered around the speaker. The step of generating a recall response instruction or a rejection interception instruction for the target voice data based on the dialogue intent and the interaction rejection strategy includes: If the dialogue scenario corresponding to the dialogue intent matches a preset skill-oriented scenario or a strong human-computer dialogue scenario, a recall response instruction for the target voice data is generated. If the dialogue scenario corresponding to the dialogue intent matches a weak human-computer dialogue scenario or a casual conversation scenario around the speaker, and if the interaction rejection strategy is the first rejection strategy, then the recall response instruction is generated; if the interaction rejection strategy is the second rejection strategy, then the rejection interception instruction is generated.

[0009] According to a vehicle-mounted voice interaction control method provided by the present invention, after generating the denial-of-identity interception command, the method further includes: Obtain the target text corresponding to the target voice data, and temporarily store the target text that triggers the denial-of-identity interception command and the dialogue intent; Obtain candidate speech data for the next round of input, and identify the candidate text and candidate dialogue intent corresponding to the candidate speech data; If the candidate dialogue intent matches the temporarily stored dialogue intent Figure 1 If the candidate text and the target text are similar and the similarity between them meets a preset similarity threshold, then a recall response instruction is generated for the candidate speech data.

[0010] The in-vehicle voice interaction control method provided by the present invention further includes: Entity word detection is performed on the target text corresponding to the target voice data to determine whether the target text contains a preset brand wake-up word; If the target text contains the brand wake-up word, and the target text presents a combination of any instruction and the brand wake-up word, then the current interaction rejection strategy's interception logic is blocked, and a recall response instruction for the target voice data is generated, wherein the combination is either the brand wake-up word preceding or following the brand wake-up word.

[0011] According to a vehicle-mounted voice interaction control method provided by the present invention, the cockpit perception data includes voice endpoint detection data for each voice zone, wherein each voice zone includes the wake-up voice zone and at least one non-wake-up voice zone, and the step of determining the number of occupants in the vehicle based on the cockpit perception data includes: Based on the speech endpoint detection data of each speech region, the first speech endpoint time of the wake-up speech region and the second speech endpoint time of the at least one non-wake-up speech region are obtained. The first voice endpoint time is compared with the second voice endpoint time to determine the temporal overlap relationship between the first voice endpoint time and the second voice endpoint time on the time axis; The number of people inside the vehicle is determined based on the aforementioned temporal overlap relationship.

[0012] According to a vehicle-mounted voice interaction control method provided by the present invention, determining the number of occupants in the vehicle based on the temporal overlap relationship includes: If, during the interaction rounds of the wake-up voice zone, a synchronous overlap is detected between the first voice endpoint time and the second voice endpoint time, or if the second voice endpoint time exists within a preset time window preceding the first voice endpoint time, then it is determined that there are multiple people in the vehicle.

[0013] The in-vehicle voice interaction control method provided by the present invention further includes: Extract historical text from the multi-voice region mixed history stack and calculate the contextual semantic relevance between the historical text and the target text corresponding to the target speech data. The multi-voice region mixed history stack is constructed based on the speech text corresponding to all continuously identified voice regions in the vehicle according to the time series, and is used to capture the dialogue relay features across voice regions. If the context semantic relevance meets the preset relevance threshold, then the target voice data is determined to be dialogue data between people in the vehicle, and a denial-of-identity interception instruction is generated for the target voice data; If the contextual semantic relevance does not meet the preset relevance threshold, the target speech data is arbitrated based on the interaction content in the wake-up voice region independent history stack to generate the recall response instruction or the denial-of-identification interception instruction. The wake-up voice region independent history stack is constructed based on the speech text corresponding to the continuously identified wake-up voice region in a time sequence and is used to determine the contextual coherence within a single voice region when the contextual semantic relevance does not meet the preset relevance threshold.

[0014] The in-vehicle voice interaction control method provided by the present invention further includes: Extract multidimensional acoustic features from the target speech data, wherein the multidimensional acoustic features include at least two of the following: fundamental frequency features, rhythmic features, spectral features, and linguistic features; Based on the multidimensional acoustic features, humming recognition and classification are performed to determine the humming classification result corresponding to the target speech data; If the humming classification result indicates that the target speech data is any one of wordless humming audio, humming audio with words, or singing along audio, then a rejection and interception instruction is generated for the target speech data.

[0015] According to a vehicle-mounted voice interaction control method provided by the present invention, the cockpit perception data further includes in-vehicle image data, and the method further includes: The in-vehicle image data is identified to extract the visual behavior features of the target person, which include at least one of head posture features, gaze focus features, hand grip features, and lip movement features. Based on the visual behavioral characteristics, determine the current interaction state of the target person; If the current interaction state is any one of the following: making a phone call, interacting with a handheld device, chatting with multiple people, interacting with a child, or interacting with a pet, then it is determined that the current situation is a non-human-computer dialogue scenario, and a rejection and interception instruction is generated for the target voice data.

[0016] The present invention also provides an in-vehicle voice interaction control system, comprising: The environmental perception module is used to acquire the current cabin perception data inside the vehicle and determine the number of people inside the vehicle based on the cabin perception data. The strategy determination module is used to determine the current interaction rejection strategy based on the number of people in the vehicle; The intent recognition module is used to receive target voice data input to the wake-up voice zone, perform intent recognition on the target voice data, and obtain the dialogue intent; The arbitration execution module is used to generate a recall response instruction or a rejection interception instruction for the target voice data based on the dialogue intent and the interaction rejection strategy, and to execute a voice interaction response according to the recall response instruction or remain silent according to the rejection interception instruction.

[0017] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor executes the computer program to implement the in-vehicle voice interaction control method as described above.

[0018] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the vehicle voice interaction control method as described above.

[0019] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements any of the in-vehicle voice interaction control methods described above.

[0020] The in-vehicle voice interaction control method, system, electronic device, and storage medium provided by this invention first accurately determines the number of occupants in the vehicle by acquiring cabin perception data, breaking the limitation of traditional in-vehicle voice interaction systems that rely solely on simple text semantics for dialogue determination. Based on this, the invention can dynamically and flexibly match an appropriate interaction rejection strategy according to whether the in-vehicle environment is single-person or multi-person. Subsequently, after receiving voice input from the wake-up voice zone and recognizing the specific dialogue intent, the system integrates this dialogue intent with the currently dynamically determined interaction rejection strategy, performing comprehensive arbitration to determine whether to generate a recall response command or a rejection interception command. This mechanism effectively avoids system erroneous responses in multi-person scenarios, reduces the risk of over-recall in the voice interaction system, minimizes frequent interruptions to normal communication among in-vehicle occupants, and improves the accuracy and user experience of in-vehicle voice interaction. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in this invention or related technologies, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 This is one of the flowcharts of the in-vehicle voice interaction control method provided by the present invention; Figure 2 This is a schematic diagram of the speech endpoint detection data for each vocal range provided by the present invention; Figure 3This is a schematic diagram of the process for vehicle-mounted humming recognition provided by the present invention; Figure 4 This is a schematic diagram comparing the features of humming and normal speech provided by the present invention; Figure 5 This is a schematic diagram of the continuous and repeated enhanced recall process provided by the present invention; Figure 6 This is the second flowchart of the in-vehicle voice interaction control method provided by the present invention; Figure 7 This is a schematic diagram of the structure of the in-vehicle voice interaction control system provided by the present invention; Figure 8 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention. It should be noted that all actions involving the acquisition of signals, information, or data in this invention are performed in accordance with the relevant data protection laws and policies of the country where the invention is located, and with authorization from the owner of the corresponding device.

[0024] With the continuous development of artificial intelligence technology, large language models, due to their extensive and powerful generation capabilities, are increasingly being applied to in-vehicle voice interaction systems. This overcomes the rigidity and narrowness of traditional small models' casual conversation capabilities, thus better responding to users' natural language input. Unlike large model dialogues on web pages or mobile devices, which typically have clear, strong human-computer interaction windows or trigger conditions, in the in-vehicle interaction space, interaction is primarily voice-based. Therefore, in-vehicle voice interaction systems inevitably need to accurately distinguish between non-human-computer dialogue and human-computer dialogue in the received voice, so as to intervene with a large model response at the appropriate time, enhancing the fluency of the dialogue while reducing user annoyance.

[0025] Currently, the technologies used to distinguish between non-human-machine dialogue and human-machine dialogue in vehicle cockpits typically only analyze and determine the dialogue at the text level. Specifically, when the system determines through text semantics that the voice input belongs to a typical non-human-machine dialogue, it directly adopts a rejection strategy and does not respond; when it determines that the voice input belongs to a typical human-machine dialogue, it directly allows the dialogue, and the large model provides the subsequent response.

[0026] However, in practical applications, the above solutions have been found to have significant shortcomings. Specifically, the logic of relying solely on text content for dialogue determination is too simplistic and cannot adapt to the complex and ever-changing in-vehicle interaction environment. In particular, it cannot solve the problem of misjudgment caused by the significantly increased complexity of user dialogue in multi-person scenarios.

[0027] For example, when a user inputs a voice expression of subjective preference (such as "I really like beef hot pot"), it's reasonable to interpret this as a conversation with the system and trigger a response if only one person is in the car. However, in a car with multiple people, such expressions of subjective preference highly overlap with normal non-human-machine dialogue between occupants. In this case, if the system still directly identifies it as a human-machine dialogue based solely on the text content and triggers a response, there is a significant risk of over-recalling, which could disrupt normal user communication and negatively impact the user's in-car voice interaction experience.

[0028] In addition, although there are solutions based on dynamic switching of multi-zone rejection modes in related technologies, they only focus on the management of rejection modes at the zone level and lack fine-grained scene layering of dialogue intent; there are also solutions that use multimodal features to train rejection models, but they usually use end-to-end black box models for unified judgment, and lack interpretable classification of different dialogue scenarios and interactive error correction mechanisms for false rejection.

[0029] To address this, the present invention provides an in-vehicle voice interaction control method. By acquiring cabin perception data to determine the number of people in the vehicle, and dynamically adjusting the interaction rejection strategy based on the number of people in the vehicle, the method then performs corresponding recall or rejection operations based on the identified dialogue intent. This changes the single mode of relying solely on text dialogue for rejection determination in related technologies. It can flexibly adjust the tightness of rejection according to the actual scene in the vehicle (single or multiple people), thereby effectively reducing the risk of over-recall caused by misjudging non-human-machine dialogue as human-machine dialogue in multi-person scenarios, reducing disturbance to users, improving the accuracy of in-vehicle voice interaction and user experience, and thus overcoming the above-mentioned defects.

[0030] This invention provides a method for controlling in-vehicle voice interaction, addressing the shortcomings of existing in-vehicle voice interaction systems that rely solely on plain text for dialogue determination, leading to erroneous responses and frequent disruptions to normal user communication. It also overcomes the deficiencies of related multi-zone rejection schemes, which lack fine-grained scene layering of dialogue intent and interaction-level rejection error correction capabilities. The method provided in this invention logically divides rejection into three dimensions: scene rejection, principle rejection, and interaction rejection, achieving precise control through multi-dimensional collaborative determination. The technical solution provided by this invention will be described in detail below.

[0031] Figure 1 This is one of the flowcharts illustrating the in-vehicle voice interaction control method provided by the present invention, such as... Figure 1 As shown, the method includes: Step 110: Obtain the current cabin perception data inside the vehicle, and determine the number of people inside the vehicle based on the cabin perception data.

[0032] Specifically, cabin perception data refers to data that reflects the state of the in-vehicle environment, especially the distribution and status of occupants. This data can be acquired through various sensors deployed inside the vehicle. For example, cabin perception data can include cabin sensing data from gravity sensors or pressure sensors inside the seats. By detecting the pressure on each seat, it can be determined whether there is only the driver in the vehicle or multiple people, including the front passenger or rear passengers.

[0033] In addition, cabin perception data can also include visual image data from in-vehicle cameras or multi-zone audio data from in-vehicle microphone arrays. Based on the acquired cabin perception data, the system can analyze and determine the number of people currently in the vehicle, such as one or more people.

[0034] Step 120: Determine the current interaction rejection strategy based on the number of people in the vehicle.

[0035] Specifically, the interaction rejection strategy refers to the set of rules that an in-vehicle voice interaction system uses to decide whether to respond (i.e., recall) or ignore (i.e., reject) a user's voice input. Because the complexity of user voice input and the dialogue partners differ significantly between single-person and multi-person scenarios, the strictness of the strategy needs to be dynamically adjusted based on the number of people in the vehicle.

[0036] If it's determined that the vehicle is occupied by a single person, the voice input is highly likely directed at the in-vehicle system (i.e., human-machine dialogue). Therefore, the system can adopt a relatively lenient interaction rejection strategy; that is, as long as the dialogue is complete and fluent, the system tends to initiate a response. Conversely, if it's determined that the vehicle is occupied by multiple people, since casual conversation (not human-machine dialogue) is highly likely among the occupants, blindly responding would be disruptive. In this case, the system will adopt a stricter interaction rejection strategy, tightening the rejection criteria. A response will only be initiated when the dialogue is both complete and fluent, and contains clear and meaningful instructions or question-and-answer intentions.

[0037] Step 130: Receive target speech data input to the wake-up voice zone, perform intent recognition on the target speech data, and obtain the dialogue intent.

[0038] Specifically, the wake-up sound zone typically refers to a specific area within the vehicle that triggers a voice interaction wake-up phrase (such as "Hello, Xinghuo"), such as the driver's sound zone or the passenger's sound zone. When the system is in a wake-up state, it receives target voice data input by the user from that wake-up sound zone.

[0039] Subsequently, the system utilizes Natural Language Understanding (NLU) technology or Large Language Model (LLM) to perform text conversion and semantic parsing on the target speech data, thereby identifying the user's expressed dialogue intent and the corresponding target text. Here, the dialogue intent may be a clear vehicle control skill operation (such as "open the car window"), a knowledge question, or it may be an ambiguous casual statement, an expression of personal opinion, or an emotional outburst.

[0040] Step 140: Based on the dialogue intent and the interaction rejection strategy, generate a recall response instruction or rejection interception instruction for the target voice data, and execute a voice interaction response according to the recall response instruction or remain silent according to the rejection interception instruction.

[0041] Specifically, after obtaining the dialogue intent and determining the applicable interaction rejection policy, the system makes a comprehensive decision. If the current dialogue intent meets the release conditions under the determined interaction rejection policy, the system generates a recall response command. Subsequently, the in-vehicle system generates a corresponding response or executes a corresponding vehicle control action (i.e., executes a voice interaction response) based on the command. If the dialogue intent meets the interception conditions under the current interaction rejection policy, for example, under a multi-user tightened rejection policy, if the target voice data is identified as casual conversation between users, the system generates a rejection interception command. At this time, the in-vehicle system remains silent and does not make any voice or interface response to avoid disturbing the user's normal communication.

[0042] The method provided in this invention first accurately determines the number of occupants in the vehicle by acquiring cabin perception data, breaking the limitation of traditional in-vehicle voice interaction systems that rely solely on plain text semantics for dialogue determination. Based on this, the invention can dynamically and flexibly match an appropriate interaction rejection strategy according to whether the in-vehicle environment is single-person or multi-person. Subsequently, after receiving voice input from the wake-up voice zone and recognizing the specific dialogue intent, the system integrates this dialogue intent with the currently dynamically determined interaction rejection strategy, performing comprehensive arbitration to determine whether to generate a recall response command or a rejection interception command. This mechanism effectively avoids system erroneous responses in multi-person scenarios, reduces the risk of over-recall in the voice interaction system, minimizes frequent interruptions to normal communication among in-vehicle occupants, and improves the accuracy and user experience of in-vehicle voice interaction.

[0043] Based on any of the above embodiments, this invention provides a scheme for assisting in determining the number of occupants in a vehicle based on multi-zone acoustic features, thereby more accurately adjusting the rejection intensity. Specifically, the cabin perception data includes voice endpoint detection data for each zone, wherein each zone includes the wake-up zone and at least one non-wake-up zone. Accordingly, step 110 specifically includes: Step 111: Based on the speech endpoint detection data of each speech region, obtain the first speech endpoint time of the wake-up speech region and the second speech endpoint time of the at least one non-wake-up speech region; Step 112: Compare the first voice endpoint time with the second voice endpoint time to determine the temporal overlap relationship between the first voice endpoint time and the second voice endpoint time on the time axis; Step 113: Determine the number of people inside the vehicle based on the time sequence overlap relationship.

[0044] Specifically, the in-vehicle microphone array can independently collect acoustic signals from various sound zones within the vehicle (such as the driver's sound zone, passenger's sound zone, left rear sound zone, right rear sound zone, etc.). The system performs Voice Activity Detection (VAD) on these signals to determine when someone is speaking in each sound zone. During this process, the system defines the sound zone that triggered the interaction as the wake-up sound zone (e.g., the driver's sound zone), and defines the remaining sound zones that were not directly involved in the wake-up as the non-wake-up sound zones (e.g., the sound zone where the passenger or rear seats are located).

[0045] By parsing VAD data, the system can obtain the start and end times of the voice output in the wake-up voice zone, i.e., the first voice endpoint time, which includes VAD beginTime and VAD endTime. At the same time, it can obtain the start and end times of the voice output in other non-wake-up voice zones (i.e., the second voice endpoint time).

[0046] Subsequently, the system compares these endpoint times on the same timeline to analyze whether there are temporal overlaps such as synchronous or consecutive speech. If only the wake-up sound zone generates acoustic information (i.e., the isolated first voice endpoint time), the system tends to determine that the current conversation is a single person (or only one person is interacting with the vehicle's infotainment system); however, if complex temporal overlaps are detected between multiple sound zones, it can be physically determined that there are multiple people in the vehicle, and it is very likely that a multi-person chat interaction is taking place.

[0047] The method provided in this invention introduces voice endpoint detection data from multiple voice zones in the vehicle and uses the temporal overlap of the phonation times of different voice zones on the time axis to determine the number of people in the vehicle and their interaction status. This overcomes the limitations of relying solely on text to determine multi-person scenarios and can more objectively and accurately identify multi-person chat scenarios from an acoustic information perspective, providing reliable data support for subsequent tightening of rejection strategies.

[0048] Based on any of the above embodiments, step 113 specifically includes: If, during the interaction rounds of the wake-up voice zone, a synchronous overlap is detected between the first voice endpoint time and the second voice endpoint time, or if the second voice endpoint time exists within a preset time window preceding the first voice endpoint time, then it is determined that there are multiple people in the vehicle.

[0049] Specifically, Figure 2 This is a schematic diagram of the speech endpoint detection data for each vocal range provided by the present invention, such as... Figure 2 As shown, assume that sound zone 1 is the wake-up sound zone, and sound zones 2, 3, and 4 are all non-wake-up sound zones. The system performs timing monitoring within one or two rounds of wake-up interaction. If the system detects that while the wake-up sound zone (i.e., sound zone 1, such as during time S12) is speaking, the VAD beginTime and VAD endTime of the non-wake-up sound zone (sound zone 2, such as during time S21) have a synchronous overlap on the time axis (i.e., there are VADs of the non-wake-up sound zone at the same time), this feature of two or more people speaking at the same time strongly suggests that multiple people are interacting in the vehicle. Therefore, the system determines that there are multiple people in the vehicle and adopts a tightening rejection strategy accordingly.

[0050] Secondly, even if there is no completely synchronized overlap of non-wake-up voice zones between two rounds of interaction in the wake-up voice zone, if the system detects that there is a second voice endpoint time (e.g., VAD beginTime and VAD endTime in the S31 time zone of voice zone 3) in a preset time window (e.g., within X seconds before the first voice endpoint time (e.g., VAD beginTime and VAD endTime in the S13 time zone of voice zone 1) before the first voice endpoint time (e.g., VAD beginTime and VAD endTime in the S31 time zone of voice zone 3), it indicates that the user in the non-wake-up voice zone has just finished speaking and the user in the wake-up voice zone is speaking immediately afterward. This close sequential timing also conforms to the rhythm of multi-person chat. The system also determines that there is multi-person interaction at this time and takes a tightening rejection strategy accordingly.

[0051] Apart from the two situations mentioned above, if no valid timing activity in the non-wake-up tone zone is detected, it is assumed that the user is engaging in a one-on-one conversation, and the system will relax the rejection policy, i.e., adopt a relaxed rejection policy.

[0052] The method provided in this invention, by setting strict time overlap judgment rules, including complete synchronous overlap and preceding time window overlap, can accurately capture the acoustic features of common interruptions, interruptions, or close conversations during multi-person chats, thereby accurately determining the multi-person interaction state in the vehicle and greatly improving the reliability of rejection judgment in multi-person scenarios.

[0053] Based on any of the above embodiments, step 120 specifically includes: If the number of people in the vehicle is a single person, the interaction rejection strategy is determined to be the first rejection strategy. When there are multiple people inside the vehicle, the interaction rejection strategy is determined to be the second rejection strategy. The conditions under which the second denial strategy triggers denial-of-identity interception are stricter than those under which the first denial strategy triggers denial-of-identity interception.

[0054] Specifically, after determining the number of people in the vehicle through cabin perception data (such as seat sensors), the system configures different policy modes with varying degrees of leniency. The aforementioned first rejection strategy refers to a relaxed rejection strategy that favors recall, primarily applied in single-person scenarios. In single-person scenarios, the user's voice is mostly directed at the vehicle's infotainment system; therefore, as long as the user's dialogue is complete and fluent, the system will directly recall users according to the target recall rate.

[0055] The second rejection strategy is a more stringent rejection approach, leaning towards interception, and is primarily used in multi-person scenarios. Because communication between people is frequent in multi-person scenarios, to avoid false alarms, the system raises the rejection threshold, for example, requiring a rejection rate of over 95%. In this case, simply having a complete and fluent conversation is not enough to trigger a response; it's also necessary to determine whether the conversation has a clear and relevant meaning to the vehicle's infotainment system before initiating a recall. This second rejection strategy has significantly stricter interception conditions than the first rejection strategy.

[0056] The method provided in this invention sets a first rejection strategy and a second rejection strategy for single-person and multi-person scenarios respectively, and clarifies the difference in the tightness of the interception conditions between the two. This allows the system to more flexibly adapt to the interaction characteristics under different personnel distributions, ensuring the response sensitivity when used by a single person, while also shielding against false wake-ups caused by idle chatter in multi-person scenarios, thereby improving the system's intelligence level in dealing with complex cockpit environments.

[0057] Based on any of the above embodiments, the dialogue scenario corresponding to the dialogue intent includes at least one of a preset skill-oriented scenario, a strong human-computer dialogue scenario, a weak human-computer dialogue scenario, and a casual conversation scenario centered around the speaker; correspondingly, in step 140, generating a recall response instruction or a rejection interception instruction for the target voice data based on the dialogue intent and the interaction rejection strategy includes: Step 141: If the dialogue scenario corresponding to the dialogue intent matches a preset skill-oriented scenario or a strong human-computer dialogue scenario, generate a recall response instruction for the target voice data. Step 142: If the dialogue scenario corresponding to the dialogue intent matches a weak human-computer dialogue scenario or a casual conversation scenario around the speaker, and if the interaction rejection strategy is the first rejection strategy, then the recall response instruction is generated; if the interaction rejection strategy is the second rejection strategy, then the rejection interception instruction is generated.

[0058] Specifically, the system first needs to classify the dialogue scenarios corresponding to the dialogue intent. These dialogue scenarios may include, but are not limited to, preset skill-oriented scenarios, chat dialogue scenarios centered around the speaker, and contextual dialogue scenarios. Among them, contextual dialogue scenarios include strong human-computer dialogue scenarios and weak human-computer dialogue scenarios.

[0059] Here, the preset skill-oriented scenario refers to a scenario in which the large model predefines a specific skill range and response strategy, including but not limited to vehicle control skills (such as "open the car window"), knowledge Q&A (such as "introduce the scenery of Huangshan"), language understanding, content generation (such as "help me write a review of a movie"), logical reasoning, mathematical calculation, code generation, and high emotional intelligence and friendly responses.

[0060] A strong human-computer dialogue scenario refers to a contextual dialogue scenario that contains explicit or implicit references to the preceding topic. For example, if the preceding topic is "What are some good places to eat in City A?", the current dialogue intent is an explicit interrogative sentence, "How's the cost of living there?". Another example is "What are some fun things to do in City A?", where the current dialogue intent is an explicit declarative sentence, "I highly recommend some must-see spots there," or an implicit interrogative sentence, "What's good to eat?", or an implicit declarative sentence, "Places suitable for taking photos."

[0061] For the two types of scenarios with clear interaction intent or strong contextual association, the system will uniformly allow passage regardless of whether there is one or more people in the vehicle, and directly generate a recall response instruction.

[0062] For scenarios where the intent is less direct, the system will differentiate its approach based on the current rejection strategy. Weak human-computer dialogue scenarios refer to those where there is no explicit or implicit reference to the preceding topic; they are often evaluations or descriptions surrounding a particular topic, primarily in declarative sentences (e.g., "The name of this dish is really scary!" or "I heard that they've recently waived admission fees there"). Casual conversation scenarios centered around the speaker refer to communication where the speaker, "I," is the information origin, aimed at self-expression, identity building, or seeking external interaction. Their core characteristic is high self-direction; the content closely revolves around the subject's inner world and external behavior, mainly including "my state," "my behavior," "my desires and preferences," self-declarations, inquiries about related information, and expression of opinions (e.g., "What color is my cat?" or "I really like beef hot pot").

[0063] When the system recognizes that the dialogue intent belongs to one of these two scenarios, if the first rejection strategy (i.e., single-person state) is currently being implemented, the system will appropriately relax the restrictions and generate a recall response command; however, if the second rejection strategy (i.e., multi-person state) is currently being implemented, the system will determine that these statements or expressions of subjective preferences are highly likely to be casual conversations between people in the vehicle. In order to avoid over-recall risk and disturbing users, the system will execute an interception, generate a rejection interception command, and remain silent.

[0064] Furthermore, for non-vehicle-machine dialogue subjects whose grammatical expressions are incomplete, incoherent, meaningless, or have some distinctive features, the system directly rejects the requests in both single-person and multi-person scenarios.

[0065] The method provided in this invention subdivides dialogue intent into multiple refined scenarios such as preset skill orientation, strong human-computer dialogue, weak human-computer dialogue, and casual conversation around the speaker, and performs cross-combination judgment with single / multi-person rejection strategies. This not only ensures the absolute recall of explicit instructions, but also solves the problem of frequent misjudgment of subjective casual conversation statements in multi-person scenarios, significantly enhancing the interpretability and accuracy of the system's dialogue logic.

[0066] Based on any of the above embodiments, the method further includes: Extract historical text from the multi-voice region mixed history stack and calculate the contextual semantic relevance between the historical text and the target text corresponding to the target speech data. The multi-voice region mixed history stack is constructed based on the speech text corresponding to all continuously identified voice regions in the vehicle according to the time series, and is used to capture the dialogue relay features across voice regions. If the context semantic relevance meets the preset relevance threshold, then the target voice data is determined to be dialogue data between people in the vehicle, and a denial-of-identity interception instruction is generated for the target voice data; If the contextual semantic relevance does not meet the preset relevance threshold, the target speech data is arbitrated based on the interaction content in the wake-up voice region independent history stack to generate the recall response instruction or the denial-of-identification interception instruction. The wake-up voice region independent history stack is constructed based on the speech text corresponding to the continuously identified wake-up voice region in a time sequence and is used to determine the contextual coherence within a single voice region when the contextual semantic relevance does not meet the preset relevance threshold.

[0067] It should be noted that, based on the above embodiments, this embodiment of the invention further provides an intent arbitration scheme based on a dual history stack architecture for complex contextual interactions in multi-voice zone scenarios, so as to accurately identify the conversation or interruption behavior between users in the vehicle.

[0068] Specifically, the system maintains two dimensions of historical dialogue memory stacks simultaneously during operation. The multi-voice-zone hybrid history stack is a global memory stack constructed by mixing and arranging the voice text recognized from all voice zones in the vehicle (such as driver, front passenger, and rear seats) in a strict chronological order. For example, the chronological arrangement is: driver t11 → system reply A11 → front passenger t31 → driver t13. This hybrid history stack is specifically used to capture the relay characteristics of cross-voice-zone dialogue, that is, to identify the internal communication between different people. The wake-up-zone independent history stack, on the other hand, is a local memory stack constructed by extracting only the continuous recognized text and system replies from the currently wake-up-zone (such as the driver's voice zone) in a chronological order. For example, the chronological arrangement is: driver t11 → system reply A11 → driver t13. Its core function is to determine the contextual coherence within a single voice zone when the contextual semantic relevance does not meet a preset relevance threshold (i.e., after excluding the possibility of cross-voice-zone chat), thereby accurately conveying the current user's multi-turn continuous interaction intent.

[0069] Upon receiving the target text corresponding to the target voice data input from the wake-up voice zone (e.g., t13 just spoken by the driver), the system prioritizes calculating the contextual semantic relevance of the multi-voice zone mixed history stack. Specifically, the system extracts the immediately preceding historical text from the mixed stack (e.g., t31 just spoken by the passenger) and calculates the contextual semantic relevance between the two. For example, if the passenger just said, "That car in front is driving really slowly," and the driver immediately follows with, "Indeed, I was thinking of overtaking," the system calculates that the driver's target text and the passenger's historical text are highly semantically related, meeting a preset relevance threshold. At this point, the system determines that the current speaker's words are responding to other people in the car, belonging to typical in-car conversation data (i.e., non-human-machine dialogue). Therefore, it generates a rejection interception command and directly executes the interception to prevent the vehicle's system from erroneously interrupting and disturbing the user.

[0070] Conversely, if the driver's speech is completely unrelated to the passenger's previous casual conversation—meaning the contextual semantic relevance does not meet the preset relevance threshold—the system assumes the driver is not conversing with the passenger. In this case, the system will backtrack and perform secondary arbitration based on the interaction content in the wake-up voice region's independent history stack. Combining this with the driver's own historical context—for example, if the driver's previous sentence was "Navigate to City B," and the current sentence is "Give me a detailed introduction to the local specialties"—the system will determine whether the target speech is a multi-turn relevant instruction according to the predetermined single-person or multi-person rejection strategy, thereby generating a corresponding recall response instruction or rejection interception instruction.

[0071] The method provided in this invention constructs a multi-zone mixed history stack and a wake-up zone independent history stack in parallel. It prioritizes the screening of contextual semantic relevance based on the mixed time sequence of the entire vehicle. This method can accurately identify and filter out strongly related casual conversations such as cross-zone conversations and interactions between people in multi-person scenarios. This solves the technical pain point of confusing the context of human-computer interaction and non-human-computer interaction in complex spaces and improves the accuracy and intelligence of rejection judgment in multi-zone interaction scenarios.

[0072] Based on any of the above embodiments, and considering the common and high-frequency scenario in the in-vehicle environment where users hum or sing along, which is prone to false wake-ups or garbled commands, this embodiment of the invention provides a filtering scheme based on acoustic low-level multi-dimensional features and a deep learning model. The method further includes: Extract multidimensional acoustic features from the target speech data, wherein the multidimensional acoustic features include at least two of the following: fundamental frequency features, rhythmic features, spectral features, and linguistic features; Based on the multidimensional acoustic features, humming recognition and classification are performed to determine the humming classification result corresponding to the target speech data; If the humming classification result indicates that the target speech data is any one of wordless humming audio, humming audio with words, or singing along audio, then a rejection and interception instruction is generated for the target speech data.

[0073] Specifically, Figure 3 This is a schematic diagram of the in-vehicle humming recognition process provided by the present invention, as follows: Figure 3As shown, after the vehicle microphone array acquires the audio stream, the system first performs audio preprocessing, including noise reduction using RNNoise (Recurrent Neural Network for Audio Noise Reduction) or DTLN (Dual-Signal Transformation LSTM Network) models, and crucial echo cancellation (AEC) processing to prevent the music played by the vehicle audio system from being misinterpreted as humming. Then, the system frames the audio and uses VAD to extract the effective human voice target speech data.

[0074] Next, the system extracts multidimensional acoustic features of the target speech data in parallel, specifically including four dimensions of differentiated features: fundamental frequency features, rhythmic features, spectral features, and linguistic features. Among them, the fundamental frequency features are used to distinguish between the continuous pitch changes in humming and the discrete pitch jumps in speech; the rhythmic features are used to identify the regular beats in humming and the differences between natural pauses in speech; and the spectral features can effectively distinguish between normal conversational speech rich in high-frequency bandwidth noise and humming audio that is characterized by low frequencies and strong harmonics.

[0075] Figure 4 This is a schematic diagram comparing the features of humming and normal speech provided by the present invention, such as... Figure 4 As shown, a comparison of the fundamental frequency (F0) trajectories reveals that the fundamental frequency trajectory of normal speech is short and frequently interrupted by consonants; while the fundamental frequency trajectory of humming presents a smooth and continuous melodic curve. The system can extract fundamental frequency features such as the mean, range, fundamental frequency variation, and proportion of vocal segments using the CREPE (Convolutional Representation for Pitch Estimation) or YIN algorithms.

[0076] A comparison of syllable duration reveals that syllables in normal speech are short and uneven, and combined with... Figure 4A comparison of energy envelopes reveals that, due to the abundance of voiceless consonants and semantically driven pauses in normal speech, its normalized energy exhibits frequent and steep energy troughs on the time axis (i.e., energy values ​​often sharply drop to near zero). Humming, on the other hand, follows musical rhythms. Because the user maintains a continuous voiced resonance state for an extended period during vocalization, its energy envelope exhibits a smooth, full, and sustained high-energy fluctuation, lacking the abrupt energy cutoffs seen in speech. Based on this, the system not only utilizes onset detection (to capture energy abrupt changes) and beat tracking algorithms to extract the distribution of time intervals between adjacent syllables, beat period, and beat intensity, but also extracts the coherence and fluctuation differences of this energy envelope trajectory, using these as core temporal features to distinguish humming from speech.

[0077] The comparison of spectral energy distribution shows that normal speech has a wide spectrum and abundant high-frequency components (consonants significantly increase high-frequency energy); while humming energy is concentrated in the low frequency and has a clear and continuous harmonic structure. Based on this, the system extracts 39-dimensional MFCCs (Mel-Frequency Cepstral Coefficients), the spectral centroid, and spectral flatness, which measures the sharpness of the spectrum.

[0078] Furthermore, the discriminative role of linguistic features lies in providing reverse corroboration at the level of phoneme pronunciation and semantic confidence. Humming (especially wordless humming) often lacks the regular alternation of consonants and vowels, mainly consisting of vowels and nasals, and lacks complete grammatical logic. The system can obtain the recognition confidence, vowel ratio, and nasal ratio of speech through a lightweight ASR (Automatic Speech Recognition) front-end (humming is mostly dominated by vowels and nasals, and ASR confidence is naturally low), thus serving as a highly discriminative auxiliary signal to distinguish it from normal speech from the dimension of human vocal habits.

[0079] The features of the above four dimensions are extracted and concatenated in each frame to form a T×54 temporal feature matrix, where T is the number of time frames and 54 is the total feature dimension.

[0080] Subsequently, the system performs humming recognition and classification based on the concatenated feature vectors. Considering the limitations of real-time performance and memory in in-vehicle scenarios, the system introduces a lightweight CNN-BiLSTM (Convolutional Neural Networks-Bidrectional Long Short-Term Memory, a hybrid neural network combining convolutional neural networks and bidirectional long short-term memory networks) architecture as the main classification model. That is, the classification decision is based on the temporal classification performed by the CNN-BiLSTM model, and its output layer includes four categories of labels: normal speech, humming without words, humming with words, and singing along. Specifically, the system inputs the aforementioned T×54 feature matrix into the model. First, the model's CNN feature extraction layer (including three one-dimensional convolutional modules) captures local spectral patterns (such as harmonic structures and formant morphology), performing high-dimensional mapping of short-term features. Then, the output is fed into the model's BiLSTM layer to model temporal dependencies, accurately capturing long-term melody progressions and rhythmic cycles. Subsequently, the model's Self-Attention layer assigns higher weights to key frames in the temporal sequence (such as melody turning points and beat initiation points). Finally, through a fully connected layer and a Softmax activation function, the system outputs the probability distribution of the target speech data belonging to each category. This distribution constitutes the final humming classification result.

[0081] During the decision-making phase, the system arbitrates based on the classification result. If the humming classification result indicates that the target voice data is any of the following: wordless humming audio (the user hums along to the melody without lyrics), humming audio with lyrics (the user sings along to the lyrics softly), or singing along audio (used to sing loudly along to the car's music), for example, if the sum or maximum of the predicted probabilities of these three categories exceeds a set threshold, the system determines that the current user has not made a genuine interaction intention. At this time, even if lyric fragments such as "open" or "close" are recognized, the system will forcibly generate a rejection interception command for the target voice data, keeping the system silent and not performing any vehicle control operations or semantic responses.

[0082] The method provided in this invention, by deeply exploring the essential differences between humming and normal speech in terms of acoustic features such as fundamental frequency, rhythm, spectrum, and linguistics, and combining the CNN-BiLSTM temporal classification model for multi-dimensional feature fusion analysis and accurate recognition, effectively solves the technical problem that users singing or humming in in-vehicle music scenarios are easily misidentified as vehicle control commands by voice assistants. Without affecting the response rate of normal voice commands, it achieves the filtering of humming audio, thereby improving the robustness and user experience of the in-vehicle voice system in complex entertainment scenarios.

[0083] Based on any of the above embodiments, this invention provides a visual modality-assisted rejection scheme, further improving the accuracy of the system in identifying non-human-computer dialogue in complex scenarios. Specifically, the cockpit perception data also includes in-vehicle image data, and correspondingly, the method further includes: The in-vehicle image data is identified to extract the visual behavior features of the target person, which include at least one of head posture features, gaze focus features, hand grip features, and lip movement features. Based on the visual behavioral characteristics, determine the current interaction state of the target person; If the current interaction state is any one of the following: making a phone call, interacting with a handheld device, chatting with multiple people, interacting with a child, or interacting with a pet, then it is determined that the current situation is a non-human-computer dialogue scenario, and a rejection and interception instruction is generated for the target voice data.

[0084] Specifically, in addition to seat sensor and acoustic data, the in-vehicle system can also collect real-time in-vehicle image data using cameras deployed inside the vehicle. When the system is activated by voice, the computer vision module simultaneously analyzes this image data to extract the visual behavioral characteristics of the target occupant. These characteristics can include head posture features, gaze focus features, hand grip features, etc.

[0085] Among them, head posture features can be used to determine whether the head of the target person (such as the driver) is looking straight ahead, looking down, or turning to the other side of the vehicle (such as the passenger seat), or even whether there are nodding or shaking actions.

[0086] Eye focus features are used to track where a target person's gaze falls, such as whether they are looking at the road ahead, the central control screen (vehicle infotainment system), staring intently at a certain area in their hand (such as a mobile phone screen), or making eye contact with other people in the vehicle.

[0087] Hand grip features are used to identify the target person's hand movements and states, such as whether they are holding a mobile device similar to a mobile phone with both hands or one hand, with their arms naturally bent, or whether they are making continuous auxiliary gestures that follow the rhythm of speech.

[0088] Lip movement features are used to capture the frequency and amplitude of the opening and closing of the lips of a target person, as well as whether they are synchronized with the rhythm of speech, or to detect whether there are alternating lip opening and closing movements among multiple people in a vehicle, thereby helping to determine whether continuous speech communication is taking place.

[0089] The system combines one or more of the aforementioned visual features to determine the current interaction state of the target person. For example, when the system detects that the target person has one hand close to their ear or is making obvious talking motions, it determines that they are currently making or receiving a phone call; when the system detects that the target person's head is lowered, their gaze is fixed on their hand area, and their hand is in a holding posture for an electronic device, it determines that they are currently in a handheld device interaction state (i.e., the user may be playing a game, watching a video, or using the phone's voice function); when the system detects that two people in the car are facing each other, their gazes are making eye contact or lingering, and their lips are alternately opening and closing or their bodies are actively tilting, it determines that they are currently in a multi-person chat social state. The alternating opening and closing of the lips refers to the alternating opening and closing movements of the lips of both parties, indicating that the two parties are engaged in dialogue. In addition, when the system detects that there is a child in the car and the target person's head is pointing in the direction of the child, their hand is making a stroking motion, or their lips are making alternating opening and closing movements, it determines that they are currently interacting with a child; when the system detects that there is a pet in the car and the target person's head is pointing in the direction of the pet, their hand is making a stroking motion or holding posture, it determines that they are currently interacting with a pet.

[0090] When the system confirms that the target person is in any of the above states, it means that the voice emitted by the target person at this time is highly likely to be communicating with the person on the other end of the phone, an application on the phone, other people in the car, a child, or a pet. For example, when the system detects that the user is holding the phone with both hands and arms naturally bent, head down and phone position below head, and eyes fixed on the screen, it determines that the user is in a handheld device interaction state. Therefore, the system directly determines from a visual perspective that the current situation is not a human-computer dialogue scenario. To prevent the system from mistakenly answering or interfering with the user's normal social activities, the system will forcibly generate a rejection and interception command for the target voice data, keeping the in-vehicle system silent.

[0091] The method provided in this invention introduces in-vehicle visual image modalities for rejection decision-making and accurately captures external behaviors such as the user's head, gaze, and hands. The system can more comprehensively understand the user's current true interaction intent, thereby making accurate interception judgments in typical high-frequency scenarios that do not require vehicle system intervention, such as the user making or receiving phone calls, playing on their mobile phone, chatting face-to-face, and interacting with children or pets. This achieves a leap from content rejection to higher-dimensional scenario rejection.

[0092] Based on any of the above embodiments, in order to prevent the pre-emptive rejection strategy from incorrectly rejecting the user's true intent, this embodiment of the invention further provides a continuous and repeated reinforcement recall mechanism based on interaction logic, namely, an experience fallback error correction mechanism. Specifically, after generating the rejection interception instruction, the method further includes: Obtain the target text corresponding to the target voice data, and temporarily store the target text that triggers the denial-of-identity interception command and the dialogue intent; Obtain candidate speech data for the next round of input, and identify the candidate text and candidate dialogue intent corresponding to the candidate speech data; If the candidate dialogue intent matches the temporarily stored dialogue intent Figure 1 If the candidate text and the target text are similar and the similarity between them meets a preset similarity threshold, then a recall response instruction is generated for the candidate speech data.

[0093] Specifically, Figure 5 This is a schematic diagram of the continuous and repeated enhanced recall process provided by the present invention, such as... Figure 5 As shown, when the system's basic decision engine determines that the conversation is not being recognized (i.e., generating a rejection interception command) based on the received voice input, the system does not directly discard the conversation. Instead, it stores the corresponding target text and conversation intent in a short-term cache. Judging from users' habitual expressions, in some scenarios, if the system does not respond, users tend to emphasize their intentions by repeating themselves, thus conveying a stronger desire to engage in human-computer dialogue.

[0094] Subsequently, the system will listen to and acquire candidate speech data from the next round of input, and parse out candidate text and candidate dialogue intent. At this point, the system will enter... Figure 5 The conditional judgment process is as follows: First, it determines whether the interaction falls within the scenarios that the vehicle's infotainment system can respond to, and then it marks the number of occurrences. If a similar request is detected more than once, the system further extracts the dialogue from the two instances and compares them to verify whether the candidate dialogue intent is consistent with the temporarily stored dialogue intent. It also determines whether the grammatical and sentence structure similarity between the two sentences meets a preset similarity threshold (e.g., ...). Figure 5 The system explicitly sets a similarity match of >85%. For example, if the previous sentence "Have you ever been to Huangshan?" is rejected, and the user immediately repeats "Have you ever been to Huangshan?"; or if the previous sentence "I'm considering taking a spontaneous trip with you" is rejected, and the user immediately says "I'm considering taking a spontaneous trip with you?", the intent of these two sets of dialogues is completely consistent, and the similarity is extremely high. When the above conditions are met, the system determines that the user is emphasizing and correcting errors. At this point, the forced recall logic is triggered, no longer hindered by the previous rejection strategy, and a recall response instruction is directly generated for the candidate voice data, thus providing the corresponding reply.

[0095] The method provided in this invention utilizes short-term memory caching to temporarily store rejected dialogues and dynamically calculates the similarity between text and intent to perform forced recall when the user repeatedly makes the same request. This fully respects the user's error correction habits in real-world scenarios, effectively compensates for the false interception that may be caused by the pre-rejection strategy, and thus improves the error tolerance of voice interaction and the overall user experience.

[0096] Based on any of the above embodiments, and considering users' clearly targeted interaction habits in multi-person chat scenarios, this embodiment of the invention provides a dedicated recall channel based on strong association with brand word entities. Specifically, the method further includes: Entity word detection is performed on the target text corresponding to the target voice data to determine whether the target text contains a preset brand wake-up word; If the target text contains the brand wake-up word, and the target text presents a combination of any instruction and the brand wake-up word, then the current interaction rejection strategy's interception logic is blocked, and a recall response instruction for the target voice data is generated, wherein the combination is either the brand wake-up word preceding or following the brand wake-up word.

[0097] Specifically, in multi-person chat scenarios, users' targeted dialogue often carries certain characteristic object information. For example, in non-human-computer dialogue, they might say, "You don't want to go eat hot pot either, right, Lao Zhang?" Applying this logic to human-computer interaction dialogue, users will also habitually use the car's system's specific names. The system uses an entity recognition module to detect entity words in the target text, looking for whether a preset brand wake-up word is embedded within it (e.g., the car system's preset name is "Spark").

[0098] If the system detects that the target text contains the brand's wake word and recognizes that the entire sentence presents a combination of any command and the brand wake word, the system considers this a clear strong human-computer dialogue. This combination includes both brand word prefixes (e.g., a command preceded by the name: "Spark, I think I can win first place in this competition") and brand word suffixes (e.g., a command followed by the name: "I really don't like cloudy days, Spark"). Once this specific sentence combination characteristic is met, the system will grant it the highest priority response, directly blocking the current interaction rejection strategy's interception logic. This means ignoring the tightened rejection judgment or general intent rejection judgment in multi-person scenarios, performing a forced state reversal, and ultimately generating a recall response command for the target voice data to provide the corresponding reply.

[0099] The method provided in this invention captures brand wake-up words and their pre- and post-word combinations in text through entity word detection, and establishes an unconditionally mandatory response channel for statements with clear in-vehicle navigation system targeting characteristics. This effectively avoids the false rejection of such instructions with strong interactive intentions in highly complex scenarios such as multi-person chat, thereby further enhancing the accuracy of interaction determination.

[0100] Based on any of the above embodiments, this invention provides an in-vehicle voice interaction control method for in-vehicle human-machine interaction rejection. This method dynamically adjusts the interaction rejection strategy by combining the number of people in the vehicle, integrates dialogue intent scene layering, and achieves cross-sound zone context arbitration by constructing a dual-history stack architecture of multi-sound zone hybrid history stack and wake-up sound zone independent history stack. It provides interaction-level error correction capability through continuous repetition reinforcement recall mechanism and brand word entity bypass channel, and achieves multimodal scene-level pre-interception through visual behavior state classification and acoustic humming recognition. It forms a four-stage cascaded decision system from multimodal environment perception, intent layering, context arbitration to interaction error correction fallback, realizing accurate differentiation and differentiated control of non-human-machine dialogue and human-machine dialogue in complex in-vehicle environments, thereby effectively reducing interference with normal user communication and improving the accuracy and user experience of in-vehicle voice interaction.

[0101] Figure 6 This is the second flowchart of the in-vehicle voice interaction control method provided by the present invention, as shown below. Figure 6 As shown, the method specifically includes the following steps: The first stage involves multimodal state perception and environment initialization. When a user speaks inside the vehicle, the system first enters the multimodal state perception and environment initialization phase. The system not only receives the user's voice input through the microphone array, but also simultaneously retrieves cabin perception data, including seat gravity sensor data, VAD timing overlap data for each sound zone, and in-vehicle image data captured by the camera.

[0102] At this stage, the system rapidly completes environmental modeling at the underlying level. On one hand, it determines whether the current situation inside the vehicle is a single person (applying a lenient first rejection policy) or a multi-person situation (applying a tighter second rejection policy); on the other hand, if it directly identifies the user as being in a state of making or receiving a phone call, interacting with a handheld device, or engaging in face-to-face social chat through visual features, the system will mark it as a non-human-computer interaction at the forefront to prepare for subsequent interception.

[0103] The second phase involves constructing independent and hybrid history stacks and verifying strong contextual relevance. After environment initialization, the process moves to the stage of constructing independent and hybrid history stacks and context verification. The system constructs a hybrid history stack of multiple voice regions based on the time sequence of the text recognized from all multi-voice regions in the vehicle, and extracts the context for comparison.

[0104] If the system calculates and finds that the current voice text is highly semantically related to the historical text just spoken by people in other voice regions in the hybrid stack (i.e., it is judged to be related), it determines that the current conversation is a cross-voice region chat between people in the vehicle. At this time, the system skips the subsequent semantic analysis and directly flows to the bottom-most silent node, generates a denial-of-identity interception instruction, and remains silent.

[0105] The third stage involves dynamic intent recognition and fusion decision arbitration. If no strong correlation is found during the hybrid stack verification (i.e., the result is negative), the process proceeds to the dynamic intent recognition and fusion decision arbitration stage. In this stage, the system utilizes a Large Language Model (LLM) to perform deep intent parsing on the text in the independent historical stack of the wake-up voice region, such as distinguishing between vehicle control skills, strong human-computer dialogue, weak human-computer dialogue, and casual conversation based on subjective preferences. Simultaneously, it combines underlying acoustic features (such as the humming classification results extracted by CNN-BiLSTM) for fusion judgment.

[0106] The system cross-matches the identified intent category with the loose / tight strategy (single / multi-person) set in the first stage. If the intent is clear and meets the release conditions (such as a skill instruction), the basic arbitration determines that it is not rejected (i.e., it is judged as no), and the system directly issues and executes the interactive response. If the intent is ambiguous, belongs to subjective chatter and is under the multi-person tightening strategy, or is judged by the acoustic model as humming or singing audio, the system judges it as yes (i.e., the basic arbitration determines that it is rejected and blocked).

[0107] In the fourth stage, dual-path interaction logic strengthens recall and final execution. To prevent the tightening of policies during the basic arbitration phase from mistakenly intercepting genuine needs, the process does not end immediately after a basic arbitration decision of denial. Instead, it enters a dual-path interaction logic-enhanced recall phase. The system will activate two fallback detection channels in parallel: one detects whether the user has used brand wake-up phrase combinations (such as any command + brand word or brand word + any command), and the other detects whether the user is making highly similar, continuously repetitive error correction expressions based on temporary data.

[0108] If any of the above strong interaction intention characteristics are detected, the system determines that it is true, ignores the previous interception judgment, triggers enhanced recall, and flows to the interaction response execution node for recall; if the user does not perform the above enhanced interaction behavior, the system determines that it is false, finally confirms the interception, and flows to the silent node to not disturb the user.

[0109] The method provided in this invention has the following advantages compared to existing solutions that only switch rejection modes based on voice region, only distinguish between valid / invalid voice requests, or use an end-to-end black-box model for interactive rejection: (1) By introducing a multimodal perspective inside the vehicle, the problem of insufficient consideration caused by single-dimensional judgment is avoided. When a single person is identified inside the vehicle, a relatively lenient recall strategy is given, and when multiple people are identified inside the vehicle, a more stringent rejection strategy is given, thereby flexibly adjusting the tightness of rejection; (2) By dividing users’ casual conversations into scenarios, flexible configuration is allowed to allow or deny specific types of conversations directly, and different recall and denial strategies are given for single and multi-person states, thereby enriching the granularity of single text denial. (3) Introduce the acoustic signal features of the whole vehicle to more accurately determine whether the current dialogue is a one-person dialogue or a multi-person chat, so as to give a more accurate rejection judgment result; at the same time, for some specific scenes with obvious acoustic features, such as humming a song, to make a scene-based rejection strategy judgment. (4) Introducing the visual signal features of the whole vehicle, the scene is determined by assisting in judging the user's behavior, thereby helping to determine whether the dialogue at this moment is a non-human-machine dialogue or a clear human-machine dialogue; making comprehensive decisions based on multimodal information rather than a single text, and giving a rejection judgment by using the comprehensive judgment of vision, acoustics and text, which is more accurate and more comprehensive, greatly improving the intelligent voice interaction experience in the whole cabin scenario (especially the highly complex multi-person co-ride and multimedia entertainment scenarios).

[0110] The in-vehicle voice interaction control system provided by the present invention is described below. The in-vehicle voice interaction control system described below can be referred to in correspondence with the in-vehicle voice interaction control method described above.

[0111] Based on any of the above embodiments Figure 7 This is a schematic diagram of the structure of the in-vehicle voice interaction control system provided by the present invention, as shown below. Figure 7 As shown, the system includes: The environmental perception module 710 is used to acquire current cabin perception data inside the vehicle and determine the number of people inside the vehicle based on the cabin perception data. The strategy determination module 720 is used to determine the current interaction rejection strategy based on the number of people in the vehicle; The intent recognition module 730 is used to receive target voice data input to the wake-up sound zone, perform intent recognition on the target voice data, and obtain the dialogue intent; Arbitration execution module 740 is used to generate a recall response instruction or a rejection interception instruction for the target voice data based on the dialogue intent and the interaction rejection strategy, and to execute a voice interaction response according to the recall response instruction or remain silent according to the rejection interception instruction.

[0112] The device provided in this invention first accurately determines the number of occupants in the vehicle by acquiring cabin perception data, breaking the limitation of traditional in-vehicle voice interaction systems that rely solely on plain text semantics for dialogue determination. Based on this, the invention can dynamically and flexibly match an appropriate interaction rejection strategy according to whether the in-vehicle environment is single-person or multi-person. Subsequently, after receiving voice input from the wake-up voice zone and recognizing the specific dialogue intent, the system integrates this dialogue intent with the currently dynamically determined interaction rejection strategy, performing comprehensive arbitration to determine whether to generate a recall response command or a rejection interception command. This mechanism effectively avoids system erroneous responses in multi-person scenarios, reduces the risk of over-recall in the voice interaction system, minimizes frequent interruptions to normal communication among in-vehicle occupants, and improves the accuracy and user experience of in-vehicle voice interaction.

[0113] Based on any of the above embodiments, the strategy determination module is specifically used for: If the number of people in the vehicle is a single person, the interaction rejection strategy is determined to be the first rejection strategy. When there are multiple people inside the vehicle, the interaction rejection strategy is determined to be the second rejection strategy. The conditions under which the second denial strategy triggers denial-of-identity interception are stricter than those under which the first denial strategy triggers denial-of-identity interception.

[0114] Based on any of the above embodiments, the dialogue scenario corresponding to the dialogue intent includes at least one of a preset skill-oriented scenario, a strong human-computer dialogue scenario, a weak human-computer dialogue scenario, and a casual conversation scenario centered around the speaker. Accordingly, the arbitration execution module is specifically used for: If the dialogue scenario corresponding to the dialogue intent matches a preset skill-oriented scenario or a strong human-computer dialogue scenario, a recall response instruction for the target voice data is generated. If the dialogue scenario corresponding to the dialogue intent matches a weak human-computer dialogue scenario or a casual conversation scenario around the speaker, and if the interaction rejection strategy is the first rejection strategy, then the recall response instruction is generated; if the interaction rejection strategy is the second rejection strategy, then the rejection interception instruction is generated.

[0115] Based on any of the above embodiments, the device further includes an enhanced recall module, which, after generating the denial-of-identity interception command, is used to: Obtain the target text corresponding to the target voice data, and temporarily store the target text that triggers the denial-of-identity interception command and the dialogue intent; Obtain candidate speech data for the next round of input, and identify the candidate text and candidate dialogue intent corresponding to the candidate speech data; If the candidate dialogue intent matches the temporarily stored dialogue intent Figure 1 If the candidate text and the target text are similar and the similarity between them meets a preset similarity threshold, then a recall response instruction is generated for the candidate speech data.

[0116] Based on any of the above embodiments, the enhanced recall module is further configured to: Entity word detection is performed on the target text corresponding to the target voice data to determine whether the target text contains a preset brand wake-up word; If the target text contains the brand wake-up word, and the target text presents a combination of any instruction and the brand wake-up word, then the current interaction rejection strategy's interception logic is blocked, and a recall response instruction for the target voice data is generated, wherein the combination is either the brand wake-up word preceding or following the brand wake-up word.

[0117] Based on any of the above embodiments, the cockpit perception data includes voice endpoint detection data for each voice region, wherein each voice region includes the wake-up voice region and at least one non-wake-up voice region, and the environment perception module is specifically used for: Based on the speech endpoint detection data of each speech region, the first speech endpoint time of the wake-up speech region and the second speech endpoint time of the at least one non-wake-up speech region are obtained. The first voice endpoint time is compared with the second voice endpoint time to determine the temporal overlap relationship between the first voice endpoint time and the second voice endpoint time on the time axis; The number of people inside the vehicle is determined based on the aforementioned temporal overlap relationship.

[0118] Based on any of the above embodiments, the environment perception module is specifically used for: If, during the interaction rounds of the wake-up voice zone, a synchronous overlap is detected between the first voice endpoint time and the second voice endpoint time, or if the second voice endpoint time exists within a preset time window preceding the first voice endpoint time, then it is determined that there are multiple people in the vehicle.

[0119] Based on any of the above embodiments, the arbitration enforcement module is further configured to: Extract historical text from the multi-voice region mixed history stack and calculate the contextual semantic relevance between the historical text and the target text corresponding to the target speech data. The multi-voice region mixed history stack is constructed based on the speech text corresponding to all continuously identified voice regions in the vehicle according to the time series, and is used to capture the dialogue relay features across voice regions. If the context semantic relevance meets the preset relevance threshold, then the target voice data is determined to be dialogue data between people in the vehicle, and a denial-of-identity interception instruction is generated for the target voice data; If the contextual semantic relevance does not meet the preset relevance threshold, the target speech data is arbitrated based on the interaction content in the wake-up voice region independent history stack to generate the recall response instruction or the denial-of-identification interception instruction. The wake-up voice region independent history stack is constructed based on the speech text corresponding to the continuously identified wake-up voice region in a time sequence and is used to determine the contextual coherence within a single voice region when the contextual semantic relevance does not meet the preset relevance threshold.

[0120] Based on any of the above embodiments, the arbitration enforcement module is further configured to: Extract multidimensional acoustic features from the target speech data, wherein the multidimensional acoustic features include at least two of the following: fundamental frequency features, rhythmic features, spectral features, and linguistic features; Based on the multidimensional acoustic features, humming recognition and classification are performed to determine the humming classification result corresponding to the target speech data; If the humming classification result indicates that the target speech data is any one of wordless humming audio, humming audio with words, or singing along audio, then a rejection and interception instruction is generated for the target speech data.

[0121] Based on any of the above embodiments, the cockpit perception data further includes in-vehicle image data, and the arbitration execution module is further used for: The in-vehicle image data is identified to extract the visual behavior features of the target person, which include at least one of head posture features, gaze focus features, hand grip features, and lip movement features. Based on the visual behavioral characteristics, determine the current interaction state of the target person; If the current interaction state is any one of the following: making a phone call, interacting with a handheld device, chatting with multiple people, interacting with a child, or interacting with a pet, then it is determined that the current situation is a non-human-computer dialogue scenario, and a rejection and interception instruction is generated for the target voice data.

[0122] Figure 8 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 8As shown, the electronic device may include a processor 810, a communication interface 820, a memory 830, and a communication bus 840, wherein the processor 810, the communication interface 820, and the memory 830 communicate with each other through the communication bus 840. The processor 810 can call logical instructions in the memory 830 to execute an in-vehicle voice interaction control method, which includes: acquiring current cabin perception data in the vehicle and determining the number of occupants based on the cabin perception data; determining a current interaction rejection strategy based on the number of occupants; receiving target voice data input to the wake-up voice zone, performing intent recognition on the target voice data to obtain a dialogue intent; generating a recall response instruction or a rejection interception instruction for the target voice data based on the dialogue intent and the interaction rejection strategy, and executing a voice interaction response according to the recall response instruction or remaining silent according to the rejection interception instruction.

[0123] Furthermore, the logical instructions in the aforementioned memory 830 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to related technologies, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0124] On the other hand, the present invention also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the in-vehicle voice interaction control method provided by the above methods. The method includes: acquiring current cabin perception data in the vehicle and determining the number of occupants in the vehicle based on the cabin perception data; determining a current interaction rejection strategy based on the number of occupants in the vehicle; receiving target voice data input to the wake-up voice zone and performing intent recognition on the target voice data to obtain a dialogue intent; generating a recall response instruction or a rejection interception instruction for the target voice data based on the dialogue intent and the interaction rejection strategy, and executing a voice interaction response according to the recall response instruction or remaining silent according to the rejection interception instruction.

[0125] In another aspect, the present invention also provides a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the in-vehicle voice interaction control method provided by the above methods. The method includes: acquiring current cabin perception data within the vehicle; determining the number of occupants based on the cabin perception data; determining a current interaction rejection strategy based on the number of occupants; receiving target voice data input to a wake-up sound zone; performing intent recognition on the target voice data to obtain a dialogue intent; generating a recall response instruction or a rejection interception instruction for the target voice data based on the dialogue intent and the interaction rejection strategy; and executing a voice interaction response according to the recall response instruction or remaining silent according to the rejection interception instruction.

[0126] The system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0127] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the parts that contribute to the related technology, can be embodied in the form of software products. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0128] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A vehicle-mounted voice interaction control method, characterized in that, include: Acquire current cabin perception data and determine the number of occupants based on the cabin perception data; Based on the number of people inside the vehicle, determine the current interaction rejection strategy; Receive target speech data input to the wake-up voice zone, perform intent recognition on the target speech data, and obtain the dialogue intent; Based on the dialogue intent and the interaction rejection strategy, a recall response instruction or rejection interception instruction is generated for the target voice data, and a voice interaction response is executed according to the recall response instruction or silence is maintained according to the rejection interception instruction.

2. The in-vehicle voice interaction control method according to claim 1, characterized in that, The step of determining the current interaction rejection strategy based on the number of people inside the vehicle includes: If the number of people in the vehicle is a single person, the interaction rejection strategy is determined to be the first rejection strategy. When there are multiple people inside the vehicle, the interaction rejection strategy is determined to be the second rejection strategy. The conditions under which the second denial strategy triggers denial-of-identity interception are stricter than those under which the first denial strategy triggers denial-of-identity interception.

3. The in-vehicle voice interaction control method according to claim 2, characterized in that, The dialogue scenario corresponding to the dialogue intent includes at least one of the following: a preset skill-oriented scenario, a strong human-computer dialogue scenario, a weak human-computer dialogue scenario, and a casual conversation scenario centered around the speaker. The step of generating a recall response instruction or a rejection interception instruction for the target voice data based on the dialogue intent and the interaction rejection strategy includes: If the dialogue scenario corresponding to the dialogue intent matches a preset skill-oriented scenario or a strong human-computer dialogue scenario, a recall response instruction for the target voice data is generated. If the dialogue scenario corresponding to the dialogue intent matches a weak human-computer dialogue scenario or a casual conversation scenario around the speaker, and if the interaction rejection strategy is the first rejection strategy, then the recall response instruction is generated; if the interaction rejection strategy is the second rejection strategy, then the rejection interception instruction is generated.

4. The in-vehicle voice interaction control method according to claim 1, characterized in that, After generating the denial-of-identity interception command, the method further includes: Obtain the target text corresponding to the target voice data, and temporarily store the target text that triggers the denial-of-identity interception command and the dialogue intent; Obtain candidate speech data for the next round of input, and identify the candidate text and candidate dialogue intent corresponding to the candidate speech data; If the candidate dialogue intent is consistent with the temporarily stored dialogue intent, and the similarity between the candidate text and the target text meets a preset similarity threshold, then a recall response instruction is generated for the candidate voice data.

5. The in-vehicle voice interaction control method according to claim 1, characterized in that, Also includes: Entity word detection is performed on the target text corresponding to the target voice data to determine whether the target text contains a preset brand wake-up word; If the target text contains the brand wake-up word, and the target text presents a combination of any instruction and the brand wake-up word, then the current interaction rejection strategy's interception logic is blocked, and a recall response instruction for the target voice data is generated, wherein the combination is either the brand wake-up word preceding or following the brand wake-up word.

6. The in-vehicle voice interaction control method according to claim 1, characterized in that, The cockpit perception data includes voice endpoint detection data for each voice region, wherein each voice region includes the wake-up voice region and at least one non-wake-up voice region. Determining the number of occupants based on the cockpit perception data includes: Based on the speech endpoint detection data of each speech region, the first speech endpoint time of the wake-up speech region and the second speech endpoint time of the at least one non-wake-up speech region are obtained. The first voice endpoint time is compared with the second voice endpoint time to determine the temporal overlap relationship between the first voice endpoint time and the second voice endpoint time on the time axis; The number of people inside the vehicle is determined based on the aforementioned temporal overlap relationship.

7. The in-vehicle voice interaction control method according to claim 6, characterized in that, Determining the number of occupants in the vehicle based on the temporal overlap relationship includes: If, during the interaction rounds of the wake-up voice zone, a synchronous overlap is detected between the first voice endpoint time and the second voice endpoint time, or if the second voice endpoint time exists within a preset time window preceding the first voice endpoint time, then it is determined that there are multiple people in the vehicle.

8. The in-vehicle voice interaction control method according to any one of claims 1 to 7, characterized in that, Also includes: Extract historical text from the multi-voice region mixed history stack and calculate the contextual semantic relevance between the historical text and the target text corresponding to the target speech data. The multi-voice region mixed history stack is constructed based on the speech text corresponding to all continuously identified voice regions in the vehicle according to the time series, and is used to capture the dialogue relay features across voice regions. If the context semantic relevance meets the preset relevance threshold, then the target voice data is determined to be dialogue data between people in the vehicle, and a denial-of-identity interception instruction is generated for the target voice data; If the contextual semantic relevance does not meet the preset relevance threshold, the target speech data is arbitrated based on the interaction content in the wake-up voice region independent history stack to generate the recall response instruction or the denial-of-identification interception instruction. The wake-up voice region independent history stack is constructed based on the speech text corresponding to the continuously identified wake-up voice region in a time sequence and is used to determine the contextual coherence within a single voice region when the contextual semantic relevance does not meet the preset relevance threshold.

9. The in-vehicle voice interaction control method according to any one of claims 1 to 7, characterized in that, Also includes: Extract multidimensional acoustic features from the target speech data, wherein the multidimensional acoustic features include at least two of the following: fundamental frequency features, rhythmic features, spectral features, and linguistic features; Based on the multidimensional acoustic features, humming recognition and classification are performed to determine the humming classification result corresponding to the target speech data; If the humming classification result indicates that the target speech data is any one of wordless humming audio, humming audio with words, or singing along audio, then a rejection and interception instruction is generated for the target speech data.

10. The in-vehicle voice interaction control method according to any one of claims 1 to 7, characterized in that, The cockpit perception data also includes in-vehicle image data, and the method further includes: The in-vehicle image data is identified to extract the visual behavior features of the target person, which include at least one of head posture features, gaze focus features, hand grip features, and lip movement features. Based on the visual behavioral characteristics, determine the current interaction state of the target person; If the current interaction state is any one of the following: making a phone call, interacting with a handheld device, chatting with multiple people, interacting with a child, or interacting with a pet, then it is determined that the current situation is a non-human-computer dialogue scenario, and a rejection and interception instruction is generated for the target voice data.

11. A vehicle-mounted voice interaction control system, characterized in that, include: The environmental perception module is used to acquire the current cabin perception data inside the vehicle and determine the number of people inside the vehicle based on the cabin perception data. The strategy determination module is used to determine the current interaction rejection strategy based on the number of people in the vehicle; The intent recognition module is used to receive target voice data input to the wake-up voice zone, perform intent recognition on the target voice data, and obtain the dialogue intent; The arbitration execution module is used to generate a recall response instruction or a rejection interception instruction for the target voice data based on the dialogue intent and the interaction rejection strategy, and to execute a voice interaction response according to the recall response instruction or remain silent according to the rejection interception instruction.

12. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the in-vehicle voice interaction control method as described in any one of claims 1 to 10.

13. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the in-vehicle voice interaction control method as described in any one of claims 1 to 10.

Images