Voice interaction method, device and electronic equipment

Abstract

The application relates to a voice interaction method, device and electronic equipment, and a method applied to a streaming media playing device, which comprises the following steps: acquiring an input voice signal, and acquiring a current environmental noise intensity, a running state and a running scene; determining a target voiceprint matching threshold, and performing voiceprint identity verification; determining a current instruction receiving state of the streaming media playing device; when the voiceprint identity verification is passed and the instruction receiving state indicates that the input voice signal is allowed to be processed, performing semantic analysis on the input voice signal to obtain a candidate instruction; determining a current effective instruction range from a preset scene semantic library; when the candidate instruction belongs to the effective instruction range, acquiring a confidence score of the candidate instruction, determining a target confidence threshold; and when the confidence score is greater than or equal to the target confidence threshold, controlling an operation corresponding to the candidate instruction to be executed. The application can realize the technical effects of wake-up-free interaction, reduced mis-triggering risk, accurate instruction response and safe operation.

Description

Technical Field

[0001] This application relates to the field of voice interaction technology, and in particular to a voice interaction method, device, and electronic device. Background Technology

[0002] With the widespread adoption of streaming media playback devices such as smart TVs, mobile smart screens, and smart projectors, voice interaction has become the core human-computer interaction method for these devices. While the current mainstream "wake word + voice command" interaction mode can control the risk of accidental triggering to some extent, it suffers from cumbersome interaction steps and easily interrupts users' viewing, work, and other workflows.

[0003] To simplify the interaction process, wake-free voice interaction solutions have emerged in the industry. However, existing solutions mostly employ fixed command whitelists or fixed recognition thresholds, resulting in issues such as lack of user identity verification, insufficient scenario adaptability, inability to effectively filter invalid triggers caused by unauthorized user voices or environmental noise, and a lack of security checks for high-risk operations. These solutions struggle to balance the convenience, command response accuracy, and operational security of wake-free interaction. Therefore, how to reduce the risk of false triggers in wake-free scenarios while simultaneously ensuring both interaction convenience and operational security is a pressing technical problem that needs to be solved in this field. Summary of the Invention

[0004] In view of the above, this application provides a voice interaction method, apparatus, and electronic device to solve at least one problem existing in the background art.

[0005] In a first aspect, embodiments of this application provide a voice interaction method applied to a streaming media playback device, the method comprising:

[0006] The system acquires the input voice signal and obtains the current ambient noise level, operating status, and operating scenario of the streaming media playback device.

[0007] The target voiceprint matching threshold is determined based on the ambient noise intensity, and the voiceprint identity is verified on the input voice signal based on the target voiceprint matching threshold; the current instruction receiving status of the streaming media playback device is determined based on the operating status.

[0008] When the voiceprint identity verification is successful and the instruction receiving status indicates that the input voice signal can be processed, the input voice signal is semantically parsed to obtain candidate instructions; based on the operating scenario, the current range of valid instructions is determined from a preset scenario semantic library.

[0009] When the candidate instruction falls within the range of valid instructions, obtain the confidence score of the candidate instruction, and determine the target confidence threshold based on the running scenario;

[0010] When the confidence score is greater than or equal to the target confidence threshold, the streaming media playback device is controlled to execute the operation corresponding to the candidate instruction.

[0011] Optionally, determining the target voiceprint matching threshold based on the ambient noise intensity includes: determining the target voiceprint matching threshold based on the ambient noise intensity through a first mapping relationship; wherein the first mapping relationship includes a negative correlation between the ambient noise intensity and the target voiceprint matching threshold.

[0012] Optionally, the step of verifying the voiceprint identity of the input speech signal based on the target voiceprint matching threshold includes: extracting voiceprint features of the input speech signal using a convolutional neural network (CNN) model deployed locally on the streaming media playback device; performing similarity matching between the voiceprint features and a preset voiceprint library deployed locally on the streaming media playback device to obtain a voiceprint matching score; and determining that the voiceprint identity verification is passed when the voiceprint matching score is greater than or equal to the target voiceprint matching threshold.

[0013] Optionally, determining the current instruction receiving state of the streaming media playback device based on the operating state includes: when the operating state is active, determining that the instruction receiving state indicates that processing any input voice signal is allowed; when the operating state is dormant, determining that the instruction receiving state indicates that comparing the input voice signal with a preset voice signal set is allowed, and when the input voice signal belongs to the preset voice signal set, processing the input voice signal is allowed; when the operating state is silent, determining that the instruction receiving state indicates that processing any input voice signal is refused.

[0014] Optionally, the step of semantically parsing the input speech signal to obtain candidate instructions includes: obtaining historical interaction instructions within a preset time window; constructing a context association chain containing the running scenario, the historical interaction instructions, and the input speech signal using a semantic parsing model based on the Transformer architecture; and performing semantic parsing on the input speech signal based on the context association chain to obtain the candidate instructions.

[0015] Optionally, when the input speech signal is identified as an ambiguous instruction, the semantic parsing model combines the context association chain to perform intent reasoning and completion on the input speech signal to obtain the candidate instruction.

[0016] Optionally, the preset scenario semantic library includes a set of valid instructions corresponding to the running scenario and invalid instruction filtering rules; the set of valid instructions is used to determine whether the candidate instruction belongs to the range of valid instructions, and the invalid instruction filtering rules are used to verify and intercept the candidate instruction that does not belong to the range of valid instructions; wherein, the scenario semantic library supports user-defined extended instructions.

[0017] Optionally, obtaining the confidence score of the candidate instruction includes: extracting the speech clarity of the input speech signal through the semantic parsing model; and obtaining the confidence score of the candidate instruction based on the speech clarity, the matching degree between the candidate instruction and the running scenario, and the logical consistency of the context association chain.

[0018] Optionally, determining the target confidence threshold based on the operating scenario includes: determining the target confidence threshold based on the environmental noise intensity through a second mapping relationship; wherein the second mapping relationship includes a positive correlation between the environmental noise intensity and the target confidence threshold.

[0019] Optionally, before controlling the streaming media playback device to execute the operation corresponding to the candidate instruction, the method further includes: acquiring at least one type of detection data from the streaming media playback device; and when confirming that the detection data satisfies the preset execution conditions corresponding to the candidate instruction, controlling the streaming media playback device to execute the operation corresponding to the candidate instruction; wherein, the detection data satisfying the preset execution conditions corresponding to the candidate instruction includes at least one of the following: when the candidate instruction is a valid instruction related to distance, the user viewing distance contained in the detection data is within a preset distance range; when the candidate instruction is a valid instruction related to screen state, the real-time screen state contained in the detection data matches the preset screen state set corresponding to the candidate instruction; when the candidate instruction is a valid instruction related to posture matching, the user posture data contained in the detection data matches the preset posture set corresponding to the candidate instruction.

[0020] Optionally, controlling the streaming media playback device to execute the operation corresponding to the candidate instruction includes: determining whether the candidate instruction belongs to a preset high-risk instruction set; when the candidate instruction belongs to the high-risk instruction set, triggering a risk verification process for the candidate instruction, the risk verification process including: controlling the streaming media playback device to output an inquiry prompt message, and when a confirmation signal is received, controlling the streaming media playback device to execute the operation corresponding to the candidate instruction.

[0021] Secondly, embodiments of this application provide a voice interaction device applied to a streaming media playback device, comprising: a data acquisition module, a verification module, and an execution module. The verification module includes a first verification unit, a second verification unit, and a third verification unit. The data acquisition module is configured to acquire an input voice signal and acquire the current ambient noise intensity, operating status, and operating scenario of the streaming media playback device. The first verification unit is configured to determine a target voiceprint matching threshold based on the ambient noise intensity and perform voiceprint identity verification on the input voice signal based on the target voiceprint matching threshold. Based on the operating status, the current instruction receiving status of the streaming media playback device is determined. The second verification unit is configured to perform semantic parsing on the input voice signal to obtain candidate instructions when the voiceprint identity verification is passed and the instruction receiving status indicates that processing of the input voice signal is allowed; and to determine the current range of valid instructions from a preset scenario semantic library based on the operating scenario; the third verification unit is configured to obtain the confidence score of the candidate instruction when the candidate instruction belongs to the range of valid instructions, and to determine a target confidence threshold based on the operating scenario; and the instruction execution module is configured to control the streaming media playback device to execute the operation corresponding to the candidate instruction when the confidence score is greater than or equal to the target confidence threshold.

[0022] Thirdly, embodiments of this application provide an electronic device including a processor, the processor being configured to invoke instructions to cause the electronic device to execute the voice interaction method provided in any of the preceding first aspects.

[0023] The voice interaction method, apparatus, and electronic device provided in this application have at least the following technical effects:

[0024] By combining voiceprint identity verification that dynamically adapts to the environment with instruction permission control based on device operating status, pre-access filtering is completed before the semantic parsing of voice commands. This reduces invalid command processing caused by unauthorized users and environmental noise, lowers the risk of accidental triggering, and reduces unnecessary computing power consumption.

[0025] Meanwhile, by limiting the effective command range corresponding to the operating scenario and dynamically adapting the confidence verification, the probability of misjudging command intent across scenarios is reduced while eliminating wake word dependence and simplifying the interaction process. This balances the convenience of wake-free interaction, the accuracy of command response, and operational security. Attached Figure Description

[0026] Figure 1 A flowchart illustrating the voice interaction method provided in an embodiment of this application;

[0027] Figure 2A schematic diagram of the voiceprint identity verification process in the voice interaction method provided in the embodiments of this application;

[0028] Figure 3 A schematic diagram illustrating the semantic parsing process in the voice interaction method provided in this application embodiment;

[0029] Figure 4 This is a schematic diagram of the structure of the voice interaction device provided in the embodiments of this application. Detailed Implementation

[0030] To make the technical solution and beneficial effects of this application more apparent and understandable, a detailed description is provided below by listing specific embodiments. The accompanying drawings are not necessarily drawn to scale, and local features may be enlarged or reduced to more clearly show the details of the local features; unless otherwise defined, the technical and scientific terms used herein have the same meanings as those in the technical field to which this application pertains.

[0031] The embodiments in this application are not exhaustive, but merely illustrative of some embodiments, and are not intended to limit the scope of protection of this disclosure. Unless otherwise specified, each step in a particular embodiment can be implemented as an independent embodiment, and the steps can be arbitrarily combined. For example, a solution after removing some steps in a particular embodiment can also be implemented as an independent embodiment, and the order of the steps in a particular embodiment can be arbitrarily interchanged. Furthermore, the optional implementation methods in a particular embodiment can be arbitrarily combined; moreover, the embodiments can be arbitrarily combined, for example, some or all steps of different embodiments can be arbitrarily combined, and a particular embodiment can be arbitrarily combined with the optional implementation methods of other embodiments.

[0032] In the various embodiments of this application, unless otherwise specified or in case of logical conflict, the terminology and / or descriptions of the various embodiments are consistent and can be referenced by each other. The technical features of different embodiments can be combined to form new embodiments according to their inherent logical relationship.

[0033] In the description of the embodiments of this application, it should be understood that the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Therefore, features defined as "first" and "second" may explicitly or implicitly include one or more features. In the description of the embodiments of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0034] Currently, streaming media playback devices with audio and video playback capabilities, such as smart TVs, mobile smart screens (commonly known as "bestie phones"), and smart projectors, generally adopt a "wake word + voice command" interaction mode for their voice interaction functions. In this mode, users need to first call a preset fixed wake word to trigger the device's voice monitoring function, and then issue control commands. Although this mode can filter out most invalid voices through the wake word and control the risk of accidental triggering to a certain extent, it has the problems of cumbersome interaction steps and easy interruption of users' viewing, office work, and other usage processes.

[0035] In related technologies, to simplify the interaction process, one or a combination of the following methods can be used:

[0036] The scenario-specific approach combined with a fixed command whitelist involves pre-setting fixed command whitelists for different usage scenarios. The device monitors voice signals in real time, and when the detected voice content matches a command within the whitelist, it executes the corresponding action. While this solution eliminates the mandatory requirement for a wake word in specific scenarios, it suffers from the inability to distinguish the identity of the user from whom the command originates. Unauthorized users (such as visitors or minors) can also use commands to activate the device. Furthermore, this solution cannot filter voice segments in the device's own playback content or ambient conversations that match whitelist commands, easily triggering unexpected actions and posing a risk of accidental activation. Additionally, the command whitelist is fixed and cannot be dynamically adjusted according to the scenario, resulting in poor scenario adaptability.

[0037] A fixed recognition threshold combined with a single-modal voice matching scheme: This scheme uses pre-set fixed voice recognition and voiceprint matching thresholds. The device matches and recognizes the voice signal based on the fixed voiceprint matching threshold and real-time acquired voice signals. If the match is successful, the corresponding operation is executed. This scheme has simple logic, but it is not suitable for scenarios with frequently changing environments. For example, in high-noise environments, the fixed voiceprint matching threshold may cause the user's voice signal to be unrecognized or misrecognized, affecting the accuracy of recognition. Furthermore, this scheme may experience misinterpretation of commands across different scenarios, lacking additional security checks on commands. For high-risk operations, misrecognition may lead to damage to user rights.

[0038] To address the challenges of balancing ease of interaction, accuracy of command response, and operational security in wake-free voice interaction for streaming media playback devices, this application proposes a voice interaction method. By constructing a three-layer filtering system—identity and status confirmation, scene semantic reasoning, and confidence verification—this method leverages the synergy of dynamic voiceprint identity verification, device operating status permission management, operational scene command range limitation, and dynamic command confidence adaptation. This reduces the risk of accidental triggering in wake-free voice interaction while maintaining both command recognition accuracy and operational security.

[0039] refer to Figure 1As shown, this embodiment provides a voice interaction method applied to a streaming media playback device, the method including:

[0040] Step S110: Obtain the input voice signal and obtain the current ambient noise intensity, operating status, and operating scenario of the streaming media playback device.

[0041] This step is the basic data acquisition step of the method, used to synchronously acquire the input voice signal of voice interaction, as well as the operating status and operating scenario data of the streaming media playback device required for subsequent steps, providing real-time data acquisition.

[0042] Specifically, streaming media playback devices refer to electronic devices with audio and video decoding and playback capabilities, as well as voice acquisition and processing capabilities, including but not limited to smart TVs, mobile smart screens (commonly known as "bestie phones"), and smart projectors; input voice signals are acquired in real time through the microphone array built into the streaming media playback device, serving as audio signals that carry the user's control intentions; ambient noise intensity refers to the sound pressure level of the background noise in the environment where the streaming media playback device is located, measured in decibels (dB), which is acquired and calculated by the device's audio acquisition components and used to characterize the degree of acoustic interference in the current environment; operating status is determined by reading the operating parameters of the device system, including but not limited to screen on / off status, foreground application running status, and system do-not-disturb mode on / off status; operating scenarios are determined by identifying the type of foreground application and the characteristics of the currently playing content, including but not limited to movie playback scenarios, meeting PPT presentation scenarios, tutorial playback scenarios, and system desktop standby scenarios.

[0043] In some examples, the data collected in this step is stored locally on the streaming media playback device and not uploaded to a cloud server. This ensures real-time data processing while reducing the risk of leakage of user privacy data such as voice and usage behavior. In other examples, this step preprocesses the input voice signal and sensor data using data noise reduction algorithms to remove interfering data.

[0044] Step S120: Determine the target voiceprint matching threshold based on the ambient noise intensity, and perform voiceprint identity verification on the input voice signal based on the target voiceprint matching threshold; determine the current instruction receiving status of the streaming media playback device based on the operating status.

[0045] This step is the admission verification step of the method, which includes two parallel or serial processing branches: voiceprint identity verification and instruction reception status determination. It is used to intercept invalid input voice signals through admission verification. Only when the verification results of both branches meet the requirements are the corresponding input voice signals allowed to enter the subsequent processing stage. This can effectively reduce the consumption of computing power of streaming media playback devices by invalid input voice signals and reduce the risk of false triggering.

[0046] Specifically, the target voiceprint matching threshold is the minimum voiceprint similarity requirement used to determine whether the user corresponding to the input voice signal is a pre-authorized user of the streaming media playback device. In practice, the target voiceprint matching threshold is dynamically determined based on the ambient noise intensity obtained in step S110 to balance the accuracy and error tolerance of voiceprint recognition under different acoustic environments. Based on the determined target voiceprint matching threshold, voiceprint identity verification is performed on the input voice signal to verify whether the user is a pre-authorized user of the device. Only when the verification passes is the user considered to have voice control permissions for the streaming media playback device. Simultaneously, based on the device operating status obtained in step S110, the current command receiving status of the streaming media playback device is determined, thereby clarifying the processing permissions of the streaming media playback device for the input voice signal, including whether processing the current input voice signal is allowed, and the range of allowed commands. The mapping relationship between ambient noise intensity and the target voiceprint matching threshold, voiceprint identity verification, the correspondence rules between operating status and command receiving status, and the specific process of permission division logic involved in this step will be explained in detail in subsequent embodiments.

[0047] Step S130: When the voiceprint identity verification is successful and the instruction receiving status indicates that the input voice signal can be processed, the input voice signal is semantically parsed to obtain candidate instructions; based on the operating scenario, the current range of valid instructions is determined from the preset scenario semantic library.

[0048] This step is the semantic parsing and valid instruction range determination step of the method. The prerequisite is that the voiceprint identity verification in step S120 is successful and the instruction receiving status indicates that the current input voice signal can be processed. If the voiceprint identity verification fails or the instruction receiving status indicates that the current input voice signal cannot be processed, the processing flow of the input voice signal is directly terminated, the semantic parsing operation is not performed, and no feedback information is output, so as to reduce the possibility of invalid input voice signals occupying the computing power of the streaming media playback device and disturbing the normal use of the user.

[0049] Specifically, the input voice signal that has passed the aforementioned verification is semantically parsed to extract the user's control intent carried by the input voice signal and convert it into candidate instructions that the streaming media playback device can recognize and execute. This step involves a semantic parsing model and a contextual association chain based on historical interactions to parse the input voice signal. Simultaneously, in this step, based on the streaming media playback device's operating scenario obtained in step S110, corresponding instruction rules are matched from a preset scenario semantic library to determine the valid instruction range under the current operating scenario, clarifying the set of instructions allowed to be executed under the current scenario and the invalid instructions that need to be intercepted. Limiting the valid range of instructions by matching the operating scenario can reduce misjudgment and misexecution of instruction intent across scenarios, thereby further reducing the risk of accidental triggering. The preset scenario semantic library and the valid instruction range involved in this step will be described in detail in subsequent embodiments.

[0050] Step S140: When the candidate instruction is within the range of valid instructions, obtain the confidence score of the candidate instruction and determine the target confidence threshold based on the running scenario.

[0051] This step is a dynamic security verification step of the method. The prerequisite condition is that the candidate instruction in step S130 belongs to the current valid instruction range. If the candidate instruction does not belong to the current valid instruction range, the processing flow of the input voice signal is directly terminated, the invalid instruction is intercepted, and it does not enter the subsequent processing stage.

[0052] Specifically, candidate commands that pass scenario compliance verification undergo confidence level verification to balance the response sensitivity of commands and the need for preventing false triggering in different scenarios. The confidence level score quantifies the degree to which a candidate command represents the user's genuine operational intent, and is based on dimensions such as the clarity of the input voice signal, the matching degree between the candidate command and the current operating scenario, and the logical consistency of historical interaction contexts. The target confidence level threshold is the minimum confidence requirement for a candidate command to proceed to subsequent steps. It is dynamically determined based on the current operating scenario of the streaming media playback device and can be dynamically adjusted according to the level of interference in the scenario (usually manifested as ambient noise intensity) to balance command response sensitivity and operational safety. The calculation method of the confidence level score and the specific rules for determining the target confidence level threshold involved in this step will be explained in detail in subsequent embodiments.

[0053] Step S150: When the confidence score is greater than or equal to the target confidence threshold, control the streaming media playback device to execute the operation corresponding to the candidate instruction.

[0054] This step is the final execution step of the method, with the prerequisite that the confidence score of the candidate instruction in step S140 is greater than or equal to the target confidence threshold. If the confidence score of the candidate instruction is lower than the target confidence threshold, the processing flow of the input voice signal is directly terminated, no operation is performed, and no feedback information is output to avoid disturbing the user's normal use. Specifically, only candidate instructions that pass the aforementioned three-layer filtering verification will be executed by the streaming media playback device. Unlike the processing flow of existing wake-up-free technologies, this method effectively reduces the risk of false triggering of wake-up-free voice interaction by eliminating wake-up word dependence and simplifying the interaction process through multi-layer progressive verification. At the same time, it reduces unnecessary computing power consumption of the device and takes into account interaction accuracy, device operating efficiency, and user privacy and security.

[0055] In some implementations, determining the target voiceprint matching threshold based on ambient noise intensity includes: determining the target voiceprint matching threshold based on ambient noise intensity through a first mapping relationship; wherein the first mapping relationship includes a negative correlation between ambient noise intensity and the target voiceprint matching threshold.

[0056] This implementation method is used to clarify the dynamic determination rules of the target voiceprint matching threshold in order to adapt to the voiceprint recognition needs under different acoustic environments, balance the accuracy and fault tolerance of voiceprint verification, and reduce the probability of authorized user voiceprints being rejected in high noise environments and unauthorized voiceprints passing verification in low noise environments.

[0057] Specifically, the first mapping relationship is the correspondence rule between ambient noise intensity and the target voiceprint matching threshold. This rule can be stored locally on the device in any form, such as a functional relationship or a piecewise mapping table, and can be iteratively updated with device firmware updates. Ambient noise intensity and the target voiceprint matching threshold are negatively correlated, meaning that within a preset range, the higher the ambient noise intensity, the lower the corresponding target voiceprint matching threshold; conversely, the lower the ambient noise intensity, the higher the corresponding target voiceprint matching threshold. The negative correlation mechanism is employed because: in high-noise environments, the input speech signal is highly susceptible to interference, affecting the stability of voiceprint feature extraction; appropriately lowering the target voiceprint matching threshold can reduce false rejections of authorized user voiceprints. In low-noise environments, the speech signal is less susceptible to interference, resulting in higher accuracy of voiceprint feature extraction; increasing the target voiceprint matching threshold can enhance the strictness of identity verification and reduce the probability of unauthorized voice falsely passing verification.

[0058] In some examples, the first mapping relationship is implemented using a segmented mapping table. The specific segmentation rules are as follows: when the ambient noise intensity is <40dB (quiet scene), the target voiceprint matching threshold is set to 0.85; when the ambient noise intensity is in the 40dB~60dB range (moderate interference scene), the target voiceprint matching threshold is set to 0.75; and when the ambient noise intensity is ≥60dB (noisy scene), the target voiceprint matching threshold is set to 0.7. In other examples, the first mapping relationship is implemented using a negative correlation function.

[0059] In some implementations, reference Figure 2 As shown, voiceprint identity verification of the input speech signal based on the target voiceprint matching threshold includes: Step S210: Extracting voiceprint features of the input speech signal using a convolutional neural network (CNN) model deployed locally on the streaming media playback device; Step S220: Performing similarity matching between the voiceprint features and a preset voiceprint library deployed locally on the streaming media playback device to obtain a voiceprint matching score; Step S230: Determining that the voiceprint identity verification is passed when the voiceprint matching score is greater than or equal to the target voiceprint matching threshold.

[0060] This implementation method clarifies the implementation process of voiceprint identity verification, identifies authorized users, and blocks unauthorized users. It completes the legality verification of user identity locally on the streaming media playback device, reducing the risk of privacy leakage.

[0061] Specifically, the Convolutional Neural Network (CNN) model is a lightweight CNN model deployed locally on the streaming media playback device, achieving real-time extraction of voiceprint features without relying on cloud computing power. This model is pre-trained using a voiceprint dataset and can extract unique voiceprint features from the input speech signal. The preset voiceprint library is an encrypted storage database deployed locally on the streaming media playback device, pre-loaded with voiceprint feature data of authorized users. Users can add, delete, and modify the list of authorized users in the device settings interface; all read and write operations of the voiceprint library are completed locally on the device. Similarity matching is used to match the real-time extracted input speech features with the voiceprint features of the user. The voiceprint features of the audio signal are compared with the voiceprint features of each authorized user stored in the preset voiceprint database to calculate the similarity score, which is in the range of 0-1. The higher the score, the higher the similarity between the input speech and the authorized user's voiceprint. The voiceprint identity verification rule is as follows: the voiceprint identity verification is considered successful only when the highest calculated voiceprint matching score is greater than or equal to the target voiceprint matching threshold, and the input speech signal is identified as coming from an authorized user of the device. If the highest voiceprint matching score is lower than the target voiceprint matching threshold, or there is no matching authorized user feature in the voiceprint database, the voiceprint identity verification is considered unsuccessful, and the processing flow of the input speech signal is terminated.

[0062] In some examples, the voiceprint features of each authorized user in the preset voiceprint database are generated by extracting 3-5 sets of speech samples with different speaking speeds and intonations to improve the accuracy of voiceprint matching in different usage scenarios. In other examples, when the voiceprint matching score is within the critical range of ±0.05 of the target voiceprint matching threshold, the device can trigger secondary voiceprint feature extraction and matching to reduce the possibility of misjudgment caused by a single matching error. In still other examples, the CNN model can extract the dynamic features of the speech speed and intonation of the input speech signal to assist in verification, supporting users to enable "voiceprint + face" dual authentication to resist attacks that simulate the voiceprints of authorized users.

[0063] In some implementations, the current command receiving state of the streaming media playback device is determined based on its operating state, including: when the operating state is active, determining that the command receiving state indicates that processing any input voice signal is allowed; when the operating state is dormant, determining that the command receiving state indicates that comparison between the input voice signal and a preset set of voice signals is allowed, and when the input voice signal belongs to the preset set of voice signals, processing of the input voice signal is allowed; when the operating state is silent, determining that the command receiving state indicates that processing of any input voice signal is refused.

[0064] This implementation method is used to clarify the instruction receiving status corresponding to different operating states of streaming media playback devices, indicate the processing method for input voice commands under different operating states, and reduce unnecessary computing power consumption and the risk of false triggering when not in use.

[0065] Specifically, the active state refers to the working state of the streaming media playback device with the screen lit and a valid application with audio / video playback or human-computer interaction capabilities running in the foreground, including but not limited to scenarios such as movie playback, PPT presentation, tutorial playback, and in-app interaction. In this state, the instruction reception status is determined to allow processing of any input voice signal. The sleep state refers to the standby state of the streaming media playback device with the screen lit but no valid application running in the foreground, including but not limited to scenarios such as the system desktop and screensaver interface. In this state, the instruction reception status is determined to only compare the input voice signal with a preset voice signal set. Only if the input voice signal belongs to the set is processing allowed. The preset voice signal set is a pre-configured set of basic instructions, including but not limited to basic operation instructions such as "power on", "power off", "open XX application", and "adjust volume". The silent state refers to the working state of the streaming media playback device with the screen off or the user manually turning on the do-not-disturb mode. In this state, the instruction reception status is determined to refuse to process any input voice signal.

[0066] In some examples, the preset voice signal set allows administrators to customize it in the device settings interface. Users can add or delete command types that are allowed to be processed in sleep mode according to their own usage habits.

[0067] In some implementations, reference Figure 3 As shown, semantic parsing of the input speech signal to obtain candidate instructions includes: Step S310: Obtaining historical interaction instructions within a preset time window; Step S320: Constructing a context association chain containing the running scenario, historical interaction instructions, and input speech signal using a semantic parsing model based on the Transformer architecture; Step S330: Performing semantic parsing of the input speech signal based on the context association chain to obtain candidate instructions.

[0068] This implementation method is used to clarify semantic parsing by constructing a contextual association chain of scene-history-current instruction, thereby achieving coherent semantic parsing based on multi-turn interactions and reducing intention misjudgment caused by single-turn instruction matching.

[0069] Specifically, the preset time window is the effective retention time of pre-defined historical interaction commands. Only valid interaction commands within this time window will be included in the construction scope of the context association chain. The lightweight semantic parsing model based on the Transformer architecture is deployed locally on the streaming media playback device. This model is pre-trained with voice interaction corpus through vertical scenarios (i.e., the corresponding usage scenarios of the streaming media playback device), which can effectively extract the intent features and scenario-related features in the input voice signal. The context association chain connects the operation object and operation logic of the historical interaction command with the semantic features of the current input voice signal to form a semantic logic link. Through this context association chain, the logical continuity between the current command and the historical interaction can be clearly defined, reducing the occurrence of isolated semantic parsing out of context. It is understandable that in related technologies, the semantic parsing model can directly process the input voice signal. It converts the input voice signal into a compact token through a built-in acoustic encoder, and can further directly extract acoustic features such as voice clarity and quality, without having to convert the input voice signal into text before inputting it into the semantic parsing model.

[0070] In some examples, in a TV video playback scenario, if a user first issues a "pause" command and then issues a "continue" command within 10 seconds, the semantic parsing model can clearly identify the target of the "continue" command as the currently playing video content through the contextual association chain, and directly parse it to obtain the candidate command "continue playing the current video". If the user issues a "open document" command, the speech parsing model determines that the scenario does not match and filters it directly.

[0071] In some implementations, when the input speech signal is identified as an ambiguous instruction, the semantic parsing model combines the contextual association chain to perform intent reasoning and completion on the input speech signal to obtain candidate instructions.

[0072] This implementation method clarifies the ambiguity resolution of vague instructions by automatically completing the intent through contextual association chains or by actively asking the user for intent, thus avoiding blind execution that could lead to misoperation.

[0073] Specifically, fuzzy commands refer to voice commands that lack a clear target, scope, or context, making it impossible to directly discern the user's true intent. This includes, but is not limited to, general commands like zoom in, zoom out, continue, and switch, which lack a specific target. One implementation method employs intent reasoning completion, where the voice parsing model automatically supplements the target and scope of the fuzzy command based on the current context and the target of historical interaction commands, thus completing a full candidate command. Another implementation method involves outputting a prompt message through a streaming media playback device and opening a voice monitoring window. Upon receiving a user's response voice signal, the response content is semantically parsed to determine candidate commands. If no response is received within a timeout period, the interaction process is terminated.

[0074] In some examples, such as in a projector PPT presentation scenario, the blur command "zoom in" is automatically completed to "zoom in on the text details of the current PPT page"; in a TV movie playback scenario, the blur command "zoom in" is automatically completed to "zoom in on a portion of the currently playing screen"; and in a makeup tutorial scenario, the blur command "switch" is automatically completed to "switch to a similar makeup tutorial".

[0075] In other examples, the TV is in sleep mode and receives the ambiguous instruction "turn on". The semantic parsing model cannot determine the operation object through the context association chain. The device outputs a voice prompt message "What application do you want to open?" and based on the user's reply voice "open video application", it parses the candidate instruction "open video playback application".

[0076] In some implementations, the preset scenario semantic library includes a set of valid instructions corresponding to the running scenario and invalid instruction filtering rules; the set of valid instructions is used to determine whether a candidate instruction belongs to the range of valid instructions, and the invalid instruction filtering rules are used to verify and intercept candidate instructions that do not belong to the range of valid instructions; wherein, the scenario semantic library supports user-defined extended instructions.

[0077] This implementation method is used to clarify the composition and matching rules of the scene semantic library. By defining the effective instruction set and invalid filtering rules for different scenes, it clarifies the range of effective instructions in different scenes, intercepts invalid instructions across scenes, and reduces the probability of misjudgment of intent.

[0078] Specifically, the preset scenario semantic library is deployed locally on the streaming media playback device and stored in categories according to device type and operating scenario. Each operating scenario corresponds to an independent set of valid instructions and invalid instruction filtering rules. The set of valid instructions is a standardized set of instructions that are allowed to be executed and match the user's needs in the current operating scenario. The invalid instruction filtering rules clearly define the rules for blocking instructions that are prohibited from being executed in the current scenario. For candidate instructions, it is first determined whether they belong to the range of valid instructions based on the set of valid instructions, and then a second interception verification is performed based on the invalid instruction filtering rules. In addition, users can add and adjust the set of valid instructions and the invalid instruction filtering rules in the settings interface of the streaming media playback device.

[0079] In some examples, the preset scene semantic library is configured according to device type and operating scenario as follows:

[0080] Smart TV - Video Playback Scene: Valid command set includes volume adjustment, fast forward / rewind, pause / continue, episode selection, and channel switching; invalid command filtering rules block cross-scene commands such as height adjustment, keystone correction, and text enlargement.

[0081] Smart TV - Application Interaction Scenarios: Valid command set includes opening / closing applications, searching content, and switching accounts; invalid command filtering rules include blocking cross-scenario commands such as replaying tutorials and highlighting key points;

[0082] Bestie Phone - Music Playback Scenario: Valid command set includes volume adjustment, previous / next track, pause / continue; invalid command filtering rules block cross-scenario commands such as channel switching, power off, and page turning.

[0083] Bestie Phone - Makeup / Outfit Tutorial Scenario: Valid command set includes pause / replay, zoom in on details, switch tutorials, and adjust screen brightness; invalid command filtering rules include blocking cross-scenario commands such as selection, annotation, and adjusting trapezoidal shape.

[0084] Projector-Conference PPT Scenario: Valid command set includes page turning, zooming in / out, highlighting key points, and exiting the presentation; invalid command filtering rules include blocking cross-scenario commands such as skipping to the next track, switching audio channels, and adjusting height.

[0085] Projector - Video Playback Scenarios: Valid command set includes volume adjustment, fast forward / rewind, pause / continue; invalid command filtering rules block cross-scenario commands such as annotation, page turning, and opening applications.

[0086] In other examples, scene recognition is based on a comprehensive judgment of screen content and foreground applications, with a 0.5-second delay buffer for the matching of the preset scene semantic library to resist malicious interference from high-frequency scene switching.

[0087] In some implementations, obtaining a confidence score for a candidate instruction includes: extracting the speech clarity of the input speech signal through a semantic parsing model; and obtaining a confidence score for the candidate instruction based on the speech clarity, the matching degree between the candidate instruction and the operating scenario, and the logical consistency of the contextual association chain.

[0088] This implementation method clarifies the calculation method of confidence score, and comprehensively quantifies the confidence of candidate instructions through multi-dimensional parameters, providing a basis for judgment in subsequent dynamic threshold verification.

[0089] Specifically, speech clarity refers to the signal-to-noise ratio and the integrity of the human voice frequency band of the input speech signal, used to characterize the recognizability of the speech signal, and is extracted through the aforementioned semantic parsing model; the higher the degree of matching between the candidate instruction and the set of valid instructions in the current scenario, the more the candidate instruction meets the user's needs in the current scenario, and the higher the confidence level; the logical consistency of the contextual association chain refers to the logical coherence between the operation logic of the candidate instruction, the operation object and historical interaction instructions, and the current running scenario. The higher the logical consistency of the contextual association chain, the more the candidate instruction meets the user's continuous interaction intent, and the higher the confidence level; the semantic parsing model calculates a confidence score in the range of 0-1 based on the weighted calculation of the aforementioned three parameters. The higher the score, the higher the credibility of the candidate instruction as the user's true intent.

[0090] In some examples, the confidence score is calculated with speech clarity accounting for 40%, scene matching accounting for 35%, and contextual logical consistency accounting for 25%.

[0091] In some implementations, determining the target confidence threshold based on the operating scenario includes: determining the target confidence threshold based on the ambient noise intensity through a second mapping relationship; wherein the second mapping relationship includes a positive correlation between the ambient noise intensity and the target confidence threshold.

[0092] This implementation method is used to define the dynamic determination rules for the target confidence threshold in order to adapt to the verification requirements of different interference scenarios and take into account both command response rate and false triggering.

[0093] Specifically, the ambient noise intensity characterizes the degree of scene interference in the current operating scenario; the second mapping relationship is the correspondence rule between ambient noise intensity and the target confidence threshold, which can take the form of a functional relationship or a piecewise mapping table. The second mapping relationship is positively correlated, meaning that the higher the ambient noise intensity, the higher the target confidence threshold, and vice versa. This is because in high-interference scenarios (high ambient noise intensity), the input speech signal is greatly affected by environmental interference, and semantic parsing is prone to deviation. Appropriately increasing the confidence threshold can improve the verification strictness and filter out low-confidence misidentified commands. In low-interference scenarios, the input speech signal quality is high, and the semantic parsing accuracy is high. Appropriately decreasing the confidence threshold can ensure that the user's valid commands are responded to in a timely manner and prevent a decrease in responsiveness due to over-verification.

[0094] In some examples, the second mapping relationship is implemented using a segmented mapping table. The segmentation rules are as follows: when the ambient noise intensity is <40dB (low-interference scenarios, such as living alone or late at night), the target confidence threshold is set to 0.7; when the ambient noise intensity is in the 40dB~60dB range (medium-interference scenarios, such as meetings), the target confidence threshold is set to 0.8; and when the ambient noise intensity is ≥60dB (high-interference scenarios, such as family gatherings or outdoors), the target confidence threshold is set to 0.85. In other examples, the dynamic determination rule of the confidence threshold is embedded in the core algorithm of the streaming media playback device and stored encrypted. At the same time, a security range of 0.7-0.85 is set to prevent external modifications to system parameters from causing the confidence threshold to be too low, thus leading to false triggering.

[0095] In some embodiments, before controlling the streaming media playback device to execute the operation corresponding to the candidate instruction, the method further includes: acquiring at least one type of detection data from the streaming media playback device; and when confirming that the detection data meets the preset execution conditions corresponding to the candidate instruction, controlling the streaming media playback device to execute the operation corresponding to the candidate instruction; wherein, the detection data meeting the preset execution conditions corresponding to the candidate instruction includes at least one of the following: when the candidate instruction is a valid instruction related to distance, the user viewing distance contained in the detection data is within a preset distance range; when the candidate instruction is a valid instruction related to screen state, the real-time screen state contained in the detection data matches the preset screen state set corresponding to the candidate instruction; when the candidate instruction is a valid instruction related to posture matching, the user posture data contained in the detection data matches the preset posture set corresponding to the candidate instruction.

[0096] This implementation method clarifies three specific ways to verify candidate instructions based on detection data, and further realizes multiple verification of candidate instructions. The detection data is usually obtained through sensors or corresponding acquisition algorithms / models for the required data.

[0097] Specifically, candidate commands are valid commands related to distance, meaning they require the user to operate within the device's effective viewing distance. The execution of these commands presupposes the user is in a normal usage position in front of the device to prevent accidental operation triggered by invalid voice commands from a distance. Detection data is collected through the infrared distance sensor, TOF sensor, or camera built into the streaming media playback device to calculate the real-time viewing distance between the user and the device screen. Candidate commands are also valid commands related to screen state, meaning they require the device's real-time screen to meet specific conditions before execution. The execution of these commands presupposes that the screen requires a corresponding adjustment. To prevent meaningless misoperations, the detection data is obtained through the image acquisition module and projection image analysis model of the streaming media playback device, and is used to characterize the real-time display status of the device's image. Candidate instructions are valid instructions that are related to posture matching. They refer to mechanical adjustment candidate instructions that require the user's operation posture and the device's posture status to meet specific conditions before they can be executed. The premise for the execution of such instructions is that the user has a corresponding operation intention, in order to prevent the device's mechanical actions triggered by accidental voice. The detection data is obtained through the posture sensor, human presence sensor, and touch sensor built into the streaming media playback device, and is used to characterize the user's real-time posture and the device's posture status.

[0098] In some examples, for smart TV scenarios, candidate commands related to effective distance, such as volume adjustment, channel changing, and episode selection, have a preset distance range of 1-3 meters. The command is only executed when the infrared sensor detects that the user's viewing distance is within this effective range. For smart projector scenarios, candidate commands related to image state, such as keystone correction, are only executed when keystone distortion is detected in the projected image; otherwise, the user is prompted that the image is already square and no correction is needed. For candidate commands related to image state, such as obstacle avoidance, the command is only executed when foreign objects such as paintings or ornaments are detected in the projection area, and the image size is automatically adjusted to avoid them. For mobile smart screens (like the "best friend" screen), candidate commands related to posture matching, such as height adjustment, are only executed when the stand's touch sensor detects the user's hand approaching the corresponding area. For candidate commands related to posture matching, such as brightness adjustment, the command is only executed when the device's eye-tracking module detects the user looking at the brightness adjustment controls on the screen. In other examples, two or more execution conditions can be combined for joint verification. For example, for the candidate instruction of switching tutorials on the best friend's phone, the user's viewing distance (within the effective range of 1-2 meters) and the user's posture (the user is looking at the screen tutorial playback area) can be verified at the same time.

[0099] In some implementations, controlling the streaming media playback device to execute the operation corresponding to the candidate instruction includes: determining whether the candidate instruction belongs to a preset high-risk instruction set; when the candidate instruction belongs to the high-risk instruction set, triggering a risk verification process for the candidate instruction, the risk verification process including: controlling the streaming media playback device to output an inquiry prompt message, and upon receiving a confirmation signal, controlling the streaming media playback device to execute the operation corresponding to the candidate instruction.

[0100] This implementation method is used to clarify the secondary verification rules for high-risk instructions, set up an additional verification step for high-risk operations, and reduce the probability of high-risk instructions being executed incorrectly due to misidentification.

[0101] Specifically, the preset high-risk instruction set is a pre-configured set of instructions involving core device permissions, potentially affecting user rights or device operating status, stored locally on the streaming media playback device, and allowing administrators to customize the set. In this embodiment, it is first determined whether a candidate instruction belongs to the preset high-risk instruction set. If not, the operation corresponding to the candidate instruction is executed directly. If it does, a mandatory risk verification process is triggered, and the corresponding instruction is executed only if the verification passes. If the verification fails or no confirmation information is received within the timeout period, the current interaction process is terminated directly, and no operation is performed. Further, in conjunction with the aforementioned embodiments, the risk verification process includes two optional methods: one is short-term instruction confirmation: the streaming media playback device outputs a voice prompt through the speaker to confirm with the user whether to execute the high-risk operation, and simultaneously opens a voice monitoring window. Verification is only considered successful when a clear confirmation instruction from the user is received in the window; the other is behavior association verification: sensor detection data of the streaming media playback device is obtained to verify whether the current scenario meets the preset execution conditions of the high-risk instruction. Verification is only considered successful when the sensor detection data meets the preset execution conditions.

[0102] In some examples, the preset set of high-risk commands includes commands such as powering off, deleting files, switching accounts, restoring factory settings, and switching input sources; users can add commands such as turning child lock on / off and uninstalling applications to the set of high-risk commands in the device settings interface.

[0103] In other examples, for high-risk commands involving switching input sources, the device employs a combination of behavioral correlation verification and short-term command confirmation. First, it detects a device connection signal at the corresponding input source interface via the interface sensor, then outputs a prompt message "Confirm switching to HDMI2 input source?". Upon receiving the user's confirmation command, it executes the input source switching operation. For high-risk commands involving powering off, the device uses a short-term command confirmation method, outputting a prompt message "Confirm power off?". If the user's confirmation command is received within 5 seconds, the device executes the power off operation; if no response is received within the timeout period, the operation is canceled.

[0104] Next, the voice interaction method provided in the embodiments of this application will be further explained in conjunction with specific application scenarios.

[0105] (1) Television and film broadcasting scenarios

[0106] In this embodiment, the streaming media playback device is a smart TV, and the authorized user's voiceprint has been pre-recorded in the device's local voiceprint library; the current scenario is that the user is watching a movie in the living room, the ambient noise level is 50dB (moderate interference), the TV is in an active state, and the running scenario is a movie playback scenario.

[0107] Interaction Process: The user issues the voice command "turn up the volume"; the TV determines the target voiceprint matching threshold (e.g., 0.75) based on the ambient noise intensity, performs voiceprint identity verification on the input voice, obtains a voiceprint matching score (e.g., 0.8), and the verification passes; the TV determines the command receiving status to be allowed to process any command based on the active state; it performs semantic parsing on the input voice to obtain the candidate command "turn up the volume", determines the effective command range based on the movie / TV playback scenario, and confirms that the candidate command belongs to the effective range; it determines the target confidence threshold (e.g., 0.8) based on the current scenario, calculates the confidence score of the candidate command (e.g., 0.92), which meets the confidence threshold requirement; the TV executes the volume increase operation.

[0108] Accidental Interception: During the interaction, if the keyword "pause" appears in the TV playback content, the device will perform voiceprint verification on the voiceprint. If the voiceprint is confirmed to be from an unauthorized user, the process will be terminated directly without performing any operation.

[0109] (2) Makeup tutorial scenario for mobile smart screen (best friend's phone)

[0110] In this embodiment, the streaming media playback device is a mobile smart screen (Bestie Phone), and the authorized user's voiceprint has been pre-recorded in the device's local voiceprint library; the current scenario is that the user is using the device to watch a makeup tutorial, the ambient noise level is 35dB (low interference), the Bestie Phone is in an active state, and the running scenario is a tutorial playback scenario.

[0111] Interaction Process: The user issues the voice command "zoom in on eye makeup details"; the device determines the target voiceprint matching threshold (e.g., 0.85) based on the ambient noise intensity, performs voiceprint identity verification on the input voice, obtains a voiceprint matching score (e.g., 0.9), and passes the verification; the device determines the command receiving status to be allowed to process any command based on the active state; it performs semantic parsing on the input voice to obtain the candidate command "zoom in on the current eye makeup area", determines the effective command range based on the tutorial playback scenario, and confirms that the candidate command is within the effective range; it determines the target confidence threshold (e.g., 0.7) based on the current scenario, calculates the confidence of the candidate command (e.g., 0.88), and meets the confidence threshold requirement; the device obtains the detection data from the eye-tracking sensor, confirms that the user is looking at the eye makeup area on the screen, and meets the preset execution conditions for posture matching corresponding to the candidate command; and executes the partial zoom-in operation.

[0112] Accidental Interception: During the interaction, if a family member of a user who has not registered their voiceprint gives the voice command "Close Tutorial", the voiceprint identity verification of the friend's phone will obtain a voiceprint matching score (e.g., 0.6), which is lower than the target voiceprint matching threshold, and the verification will fail, thus filtering the command.

[0113] (3) Smart projector conference PPT scenario

[0114] In this embodiment, the streaming media playback device is a smart projector, and the authorized user's voiceprint has been pre-recorded in the device's local voiceprint library; the current scenario is a conference room using the device to display a PPT, the ambient noise level is 65dB (high interference), the device is in an active state, and the running scenario is a conference PPT presentation scenario.

[0115] Interaction process: The user issues a voice command "turn to the next page"; the device determines the target voiceprint matching threshold (e.g., 0.7) based on the ambient noise intensity, performs voiceprint identity verification on the input voice, obtains a voiceprint matching score (e.g., 0.72), and the verification passes; the device determines the command receiving status as allowing processing of any command based on the active state; semantic parsing of the input voice yields the candidate command "PPT down page", the valid command range is determined based on the PPT presentation scenario, and the candidate command is confirmed to be within the valid range; the target confidence threshold (e.g., 0.85) is determined based on the current scenario, and the confidence of the candidate command (e.g., 0.86) is calculated, meeting the confidence threshold requirement; the device executes the page-turning operation.

[0116] High-risk command verification: During the interaction, the user issues a voice command "switch to HDMI2". The device confirms that the command belongs to the preset high-risk command set and triggers the risk verification process. The device first detects that there is a device connected to the HDMI2 interface through the interface sensor, which meets the preset execution conditions. Then, it outputs a voice question "Confirm switch to HDMI2?". After receiving the user's confirmation reply, it performs the input source switching operation.

[0117] The various embodiments or implementation methods described in this specification are presented in a progressive manner. Each embodiment focuses on the differences from other embodiments, and the same or similar parts between the embodiments can be referred to each other.

[0118] In the various embodiments of the specification, some or all of the steps and their optional implementations can be arbitrarily combined with some or all of the steps in other embodiments, or arbitrarily combined with the optional implementations in other embodiments.

[0119] refer to Figure 4 As shown in the figure, this application embodiment also provides a voice interaction device applied to a streaming media playback device, including a data acquisition module 410, a verification module 420, and an instruction execution module 430. The verification module 420 includes a first verification unit 421, a second verification unit 422, and a third verification unit 423.

[0120] The data acquisition module 410 is configured to acquire the input voice signal and acquire the current ambient noise intensity, operating status and operating scenario of the streaming media playback device;

[0121] The first verification unit 421 is configured to determine the target voiceprint matching threshold based on the ambient noise intensity, and to perform voiceprint identity verification on the input voice signal based on the target voiceprint matching threshold; and to determine the current instruction receiving status of the streaming media playback device based on the operating status.

[0122] The second verification unit 422 is configured to perform semantic parsing on the input voice signal to obtain candidate instructions when the voiceprint identity verification is passed and the instruction receiving status indicates that the input voice signal can be processed; and to determine the current range of valid instructions from the preset scenario semantic library based on the operating scenario.

[0123] The third verification unit 423 is configured to obtain the confidence score of the candidate instruction when the candidate instruction belongs to the range of valid instructions, and determine the target confidence threshold based on the running scenario.

[0124] The instruction execution module 430 is configured to control the streaming media playback device to execute the operation corresponding to the candidate instruction when the confidence score is greater than or equal to the target confidence threshold.

[0125] In some embodiments, the first verification unit 421 is specifically used to determine the target voiceprint matching threshold based on the ambient noise intensity through a first mapping relationship; wherein, the first mapping relationship includes a negative correlation between the ambient noise intensity and the target voiceprint matching threshold.

[0126] In some embodiments, the first verification unit 421 is specifically used to extract the voiceprint features of the input speech signal through a convolutional neural network (CNN) model deployed locally on the streaming media playback device; perform similarity matching between the voiceprint features and a preset voiceprint library deployed locally on the streaming media playback device to obtain a voiceprint matching score; and determine that the voiceprint identity verification is passed when the voiceprint matching score is greater than or equal to the target voiceprint matching threshold.

[0127] In some embodiments, the first verification unit 421 is specifically used to determine, when the running state is active, that the instruction receiving state indicates that processing any input voice signal is allowed; and when the running state is dormant, to determine that the instruction receiving state indicates that comparison between the input voice signal and a preset voice signal set is allowed, and when the input voice signal belongs to the preset voice signal set, processing of the input voice signal is allowed.

[0128] In some embodiments, the second verification unit 422 is specifically used to obtain historical interaction instructions within a preset time window; construct a context association chain containing the running scenario, historical interaction instructions, and input speech signals using a semantic parsing model based on the Transformer architecture; and perform semantic parsing on the input speech signals based on the context association chain to obtain candidate instructions.

[0129] In some embodiments, the second verification unit 422 is specifically used to perform intent reasoning and completion on the input speech signal by combining the semantic parsing model with the context association chain when the input speech signal is identified as an ambiguous instruction, so as to obtain candidate instructions.

[0130] In some embodiments, the preset scenario semantic library in the second verification unit 422 includes a set of valid instructions corresponding to the running scenario and invalid instruction filtering rules; the set of valid instructions is used to determine whether a candidate instruction belongs to the range of valid instructions, and the invalid instruction filtering rules are used to verify and intercept candidate instructions that do not belong to the range of valid instructions; wherein, the scenario semantic library supports user-defined extended instructions.

[0131] In some embodiments, the third verification unit 423 is specifically used to extract the speech clarity of the input speech signal through a semantic parsing model; and to obtain a confidence score of the candidate instruction based on the speech clarity, the matching degree between the candidate instruction and the running scenario, and the logical consistency of the context association chain.

[0132] In some embodiments, the third verification unit 423 is specifically used to determine the target confidence threshold based on the ambient noise intensity through a second mapping relationship; wherein, the second mapping relationship includes a positive correlation between the ambient noise intensity and the target confidence threshold.

[0133] In some embodiments, the verification module 420 further includes a fourth verification unit, which is specifically used to acquire at least one type of detection data from the streaming media playback device. When it is confirmed that the detection data meets the preset execution conditions corresponding to the candidate instruction, the fourth verification unit controls the streaming media playback device to execute the operation corresponding to the candidate instruction. The detection data meeting the preset execution conditions corresponding to the candidate instruction includes at least one of the following: when the candidate instruction is a valid instruction related to distance, the user viewing distance contained in the detection data is within a preset distance range; when the candidate instruction is a valid instruction related to screen state, the real-time screen state contained in the detection data matches the preset screen state set corresponding to the candidate instruction; when the candidate instruction is a valid instruction related to posture matching, the user posture data contained in the detection data matches the preset posture set corresponding to the candidate instruction.

[0134] In some embodiments, the instruction execution module 430 is specifically used to determine whether a candidate instruction belongs to a preset high-risk instruction set; when a candidate instruction belongs to a high-risk instruction set, a risk verification process for the candidate instruction is triggered, the risk verification process includes: controlling the streaming media playback device to output an inquiry prompt message, and when a confirmation signal is received, controlling the streaming media playback device to execute the operation corresponding to the candidate instruction.

[0135] It should be understood that the interaction process and specific implementation details of each component of the voice interaction device have been described in detail in the foregoing method embodiments, and will not be repeated here. Through the voice interaction device, it is possible to achieve the technical effects of reducing the risk of accidental triggering, reducing unnecessary computing power consumption, balancing the convenience of wake-up-free interaction, the accuracy of command response, and operational security.

[0136] This application also provides an electronic device, including a processor, which is configured to invoke instructions to cause the electronic device to execute the steps of the voice interaction method provided in any of the foregoing embodiments.

[0137] This application also provides a storage medium, including an executable program stored thereon, which, when executed by a processor, implements the voice interaction method provided in any of the foregoing embodiments.

[0138] For ease of understanding, the following focuses on explaining the terminology used in this embodiment:

[0139] In this application embodiment, the processor is a circuit with signal processing capabilities. In one implementation, the processor can be a circuit with instruction read and execute capabilities, such as a Central Processing Unit (CPU), a microprocessor, a Graphics Processing Unit (GPU) (which can be understood as a type of microprocessor), or a Digital Signal Processor (DSP). In another implementation, the processor can implement certain functions through the logical relationships of hardware circuits. The logical relationships of the aforementioned hardware circuits are fixed or reconstructable. For example, the processor is a hardware circuit implemented by an Application-Specific Integrated Circuit (ASIC) or a Programmable Logic Device (PLD), such as an FPGA. In a reconstructable hardware circuit, the processor loads a configuration document and implements a cyclical process of hardware circuit configuration, which can be understood as the processor loading instructions to implement the functions of some or all of the above units or modules in a cyclical process. In addition, it can also be a hardware circuit designed for artificial intelligence, which can be understood as an ASIC, such as a Neural Network Processing Unit (NPU), a Tensor Processing Unit (TPU), a Deep Learning Processing Unit (DPU), etc.

[0140] The computer-readable storage medium provided in this embodiment can execute the voice interaction method of the above embodiment. Its implementation principle and technical effect are similar to those of the above embodiment, and will not be repeated here.

[0141] The aforementioned computer-readable storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. The readable storage medium can be any available medium accessible to a general-purpose or special-purpose computer.

[0142] An exemplary readable storage medium is coupled to a processor, enabling the processor to read information from and write information to the readable storage medium. Of course, the readable storage medium can also be a component of the processor. The processor and the readable storage medium can reside in application-specific integrated circuits (ASICs). Alternatively, the processor and the readable storage medium can exist as discrete components in an electronic device or a host device.

[0143] Those skilled in the art will understand that all or part of the steps of the above-described method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When executed, the program performs the steps of the above-described method embodiments; and the aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks.

[0144] In the description of this specification, references to "one embodiment," "some embodiments," "illustrative embodiment," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with an embodiment or example is included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0145] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application 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 or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.

Claims

1. A voice interaction method, characterized in that, Applied to a streaming media playback device, the method includes: The system acquires the input voice signal and obtains the current ambient noise level, operating status, and operating scenario of the streaming media playback device. The target voiceprint matching threshold is determined based on the ambient noise intensity, and the voiceprint identity is verified on the input voice signal based on the target voiceprint matching threshold; the current instruction receiving status of the streaming media playback device is determined based on the operating status. When the voiceprint identity verification is successful and the instruction receiving status indicates that the input voice signal can be processed, the input voice signal is semantically parsed to obtain candidate instructions; based on the operating scenario, the current range of valid instructions is determined from a preset scenario semantic library. When the candidate instruction falls within the range of valid instructions, obtain the confidence score of the candidate instruction, and determine the target confidence threshold based on the running scenario; When the confidence score is greater than or equal to the target confidence threshold, the streaming media playback device is controlled to execute the operation corresponding to the candidate instruction.

2. The voice interaction method according to claim 1, characterized in that, Determining the target voiceprint matching threshold based on the ambient noise intensity includes: Based on the ambient noise intensity, the target voiceprint matching threshold is determined through a first mapping relationship; wherein, the first mapping relationship includes a negative correlation between the ambient noise intensity and the target voiceprint matching threshold.

3. The voice interaction method according to claim 1, characterized in that, The voiceprint identity verification of the input speech signal based on the target voiceprint matching threshold includes: The voiceprint features of the input speech signal are extracted using a convolutional neural network (CNN) model deployed locally on the streaming media playback device. The voiceprint features are matched with a preset voiceprint library deployed locally on the streaming media playback device to obtain a voiceprint matching score. When the voiceprint matching score is greater than or equal to the target voiceprint matching threshold, the voiceprint identity verification is deemed successful.

4. The voice interaction method according to claim 1, characterized in that, Determining the current instruction receiving status of the streaming media playback device based on the operating status includes: When the running state is active, it is determined that the instruction receiving state indicates that processing any input voice signal is allowed; When the running state is in sleep state, it is determined that the instruction receiving state indicates that the input voice signal is allowed to be compared with a preset voice signal set. When the input voice signal belongs to the preset voice signal set, the processing of the input voice signal is allowed. When the running state is silent, it is determined that the instruction receiving state indicates that any input voice signal is refused to be processed.

5. The voice interaction method according to claim 1, characterized in that, The step of semantically parsing the input speech signal to obtain candidate instructions includes: Retrieve historical interaction commands within a preset time window; Using a semantic parsing model based on the Transformer architecture, a context association chain is constructed that includes the running scenario, the historical interaction commands, and the input voice signal; The input speech signal is semantically parsed based on the context association chain to obtain the candidate instruction.

6. The voice interaction method according to claim 5, characterized in that, When the input speech signal is identified as an ambiguous instruction, the semantic parsing model combines the context association chain to perform intent reasoning and completion on the input speech signal to obtain the candidate instruction.

7. The voice interaction method according to claim 1, characterized in that, The preset scenario semantic library includes a set of valid instructions corresponding to the running scenario and invalid instruction filtering rules; the set of valid instructions is used to determine whether the candidate instruction belongs to the range of valid instructions, and the invalid instruction filtering rules are used to verify and intercept the candidate instruction that does not belong to the range of valid instructions; The scenario semantic library supports user-defined extension instructions.

8. The voice interaction method according to claim 5, characterized in that, The process of obtaining the confidence score of the candidate instruction includes: The semantic parsing model is used to extract the speech clarity of the input speech signal; and based on the speech clarity, the matching degree between the candidate instruction and the running scenario, and the logical consistency of the context association chain, the confidence score of the candidate instruction is obtained.

9. The voice interaction method according to claim 1, characterized in that, Determining the target confidence threshold based on the operating scenario includes: Based on the environmental noise intensity, the target confidence threshold is determined through a second mapping relationship; wherein, the second mapping relationship includes a positive correlation between the environmental noise intensity and the target confidence threshold.

10. The voice interaction method according to claim 1, characterized in that, Before controlling the streaming media playback device to execute the operation corresponding to the candidate instruction, the method further includes: Acquire at least one detection data from the streaming media playback device, and when it is confirmed that the detection data meets the preset execution conditions corresponding to the candidate instruction, control the streaming media playback device to execute the operation corresponding to the candidate instruction; Wherein, the detection data satisfies the preset execution conditions corresponding to the candidate instruction, including at least one of the following: When the candidate instruction is a valid instruction that is related to distance, the user viewing distance contained in the detection data is within a preset distance range; When the candidate instruction is a valid instruction that is related to the screen state, the real-time screen state contained in the detection data matches the preset screen state set corresponding to the candidate instruction; When the candidate instruction is a valid instruction that is related to posture matching, the user posture data contained in the detection data matches the preset posture set corresponding to the candidate instruction.

11. The voice interaction method according to claim 1, characterized in that, The operation of controlling the streaming media playback device to execute the candidate instruction includes: Determine whether the candidate instruction belongs to a preset high-risk instruction set; When the candidate instruction belongs to the high-risk instruction set, a risk verification process for the candidate instruction is triggered, the risk verification process including: The system controls the streaming media playback device to output an inquiry prompt message. Upon receiving a confirmation signal, the system controls the streaming media playback device to execute the operation corresponding to the candidate instruction.

12. A voice interaction device, applied to a streaming media playback device, characterized in that, It includes a data acquisition module, a verification module, and an instruction execution module. The verification module includes a first verification unit, a second verification unit, and a third verification unit. The data acquisition module is configured to acquire the input voice signal and acquire the current ambient noise intensity, operating status, and operating scenario of the streaming media playback device; The first verification unit is configured to determine a target voiceprint matching threshold based on the ambient noise intensity, and to perform voiceprint identity verification on the input voice signal based on the target voiceprint matching threshold; Based on the operating state, determine the current instruction receiving state of the streaming media playback device; The second verification unit is configured to perform semantic parsing on the input voice signal to obtain candidate instructions when the voiceprint identity verification passes and the instruction receiving status indicates that processing of the input voice signal is permitted; Based on the aforementioned operating scenario, the current range of valid instructions is determined from a preset scenario semantic library; The third verification unit is configured to obtain the confidence score of the candidate instruction when the candidate instruction belongs to the range of valid instructions, and determine the target confidence threshold based on the running scenario; The instruction execution module is configured to control the streaming media playback device to execute the operation corresponding to the candidate instruction when the confidence score is greater than or equal to the target confidence threshold.

13. An electronic device, characterized in that, Includes a processor, the processor being configured to invoke instructions to cause the electronic device to execute the voice interaction method as described in any one of claims 1-11.

Images