An AI voice intelligent interaction system for decompressing toys

Abstract

The application relates to the technical field of intelligent interaction, and discloses an AI voice intelligent interaction system for decompression toys; the system is used for solving the problem that the traditional method fails to jointly identify the stress behavior of the toy, the prosodic features of the voice, the semantic content and the historical interaction state; the system firstly uniformly collects the stress of the toy, the voice of the user and the conversation state information, and associates and merges the stress, the prosodic features of the voice, the semantic expression changes and the interaction historical state, determines the decompression stage and the emotional risk of the user; the system schedules the voice reply category, the tone strength, the tactile feedback, the motion feedback and the switching timing constraint, and updates the state changes and the strategy bearing relationship, so that the decompression process, the soothing effectiveness and the accompanying interaction stability can be improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent interactive technology, specifically to an AI voice intelligent interactive system for decompression toys. Background Technology

[0002] With the development of speech recognition, natural language processing, and affective computing technologies, voice-activated companion toys and devices are gradually evolving from fixed voice playback to intelligent interaction. For example, the published invention patent application CN107030691B discloses a data processing method and device for a caregiver robot, which can select the interaction mode based on the emotional state of the interactive object. Another example is the published invention patent application CN107480122B, which discloses an artificial intelligence interaction method and device that can generate interactive responses by combining semantic recognition results and emotional characteristics. These solutions, to a certain extent, improve the interactivity and responsiveness of companion devices.

[0003] However, existing solutions only consider interactive selection based on emotion classification results or generating response content based on semantic content and emotional features, without considering the closed-loop control of the actual scenario of the stress-relief toy. When users are in a state of intense negative emotional fluctuation, they will exhibit changes in force behaviors such as squeezing, patting, and kneading; changes in speech rhythm such as speech speed, volume, pauses, and crying; and semantic changes such as complaining, venting, and repetitive expressions. Furthermore, existing solutions can only identify single pieces of information from the toy and do not conduct joint analysis of the toy's force behaviors, speech rhythm features, semantic content, and historical interaction states. Consequently, it is difficult to accurately determine the user's stress relief stage and emotional risk level. Due to the lack of joint perception and stage judgment mentioned above, the content of voice responses, tone intensity, tactile or motor feedback methods, and the timing of switching soothing strategies are not targeted. Problems such as inappropriate response timing, mismatched soothing strategies, inconsistent feedback intensity, and interrupted venting result in difficulty in maintaining the stress relief effect and poor effectiveness and continuity of the companionship interaction. Therefore, it is necessary to build an AI voice intelligent interaction system for stress relief toys, establish a multimodal closed-loop recognition and soothing control mechanism for venting scenarios, determine the user's stress relief stage and emotional risk level, and dynamically adjust the interaction output method to improve the accuracy of responses, the effectiveness of soothing, and the continuity of interaction. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides an AI voice intelligent interaction system for decompression toys, which solves the problem in traditional methods of failing to jointly identify the toy's force behavior, voice prosody features, semantic content, and historical interaction status.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] An AI voice-activated intelligent interaction system for decompression toys includes:

[0007] Interactive data acquisition module: used to collect force information from toys, user voice information, and session status information;

[0008] State merging module: used to associate and organize force changes, speech rhythm changes, semantic expression changes, and interaction history states;

[0009] Stage Determination Module: Used to determine the current decompression stage and emotional risk level based on the merged interaction status;

[0010] The constraint scheduling module is used to control the content of voice responses, tone intensity, tactile feedback methods, action feedback methods, and the timing of switching soothing strategies based on the decompression stage and the level of emotional risk.

[0011] Write-back update module: Used to record and update state changes and strategy inheritance relationships after interaction.

[0012] Preferably, the interactive data acquisition module includes:

[0013] The inputs of the force sensing layer, the sound pickup unit, the posture detection unit, and the conversation management unit are aligned according to a unified time base;

[0014] Threshold filtering, event segmentation, segment merging, and itemized recording are performed on force fluctuations and speech activities;

[0015] Within the initial observation window, the session establishment is determined based on the force events, speech segments and their overlapping relationships, while pre-set handling is performed for abnormal noise, drop disturbances, channel jitter and time stamp out-of-sync.

[0016] Preferably, the process of threshold filtering, event segmentation, segment merging, and itemized recording of force fluctuations and speech activity includes:

[0017] Forced wavebands that are continuously above the effective force threshold and whose adjacent fluctuation intervals are within a preset range are grouped into the same force event.

[0018] Speech segments with adjacent vocal segments spaced within a preset range are grouped into the same expression segment, and force acquisition entries and speech acquisition entries are generated respectively.

[0019] Preferably, the state merging module includes:

[0020] The state merging module performs temporal merging of force acquisition items, voice acquisition items, conversation state items and historical state items according to a unified merging window, forming force change items, voice prosody items, semantic expression items and interaction history association items.

[0021] Generate a composite entry with a time slice identifier, modal integrity flag, and conflict status flag;

[0022] Write the summary of the compound entry into the current session's acceptor table.

[0023] Preferably, the summary of the composite entry is written to the current session's acceptance table, including:

[0024] Retain summaries of the most recent short window composite items, summaries of the most recent medium window trend items, and summaries of the most recent feedback-related items;

[0025] Record corresponding markers when there are only force events, only voice events, missing historical states, or when semantic entries conflict with prosodic entries, for use in subsequent window construction and stage determination.

[0026] Preferably, the stage determination module includes:

[0027] Based on the short window merging status and the medium window trend status, and combined with current session information and historical correlation information, stage judgment and risk judgment are made.

[0028] It switches between states according to a preset state machine, including contact establishment, catharsis establishment, high arousal catharsis, fallback transition, stabilization and soothing, and recovery end.

[0029] When a composite entry is missing, the modality is incomplete, or there is a conflict of information, a conservative decision is maintained, and the decision entry is sent to the constraint scheduling module.

[0030] Preferably, based on the short window merging status and medium window trend status, combined with current session information and historical correlation information, stage judgment and risk judgment are performed, including:

[0031] The current stage is determined based on the main receiving behavior, volume level, speech rate level, pause density, main semantic category, reassurance and rejection state, expression of urgent concern, and the direction of change of the previous state;

[0032] The risk level is determined based on the duration of high-intensity stress, the density of cathartic semantics, the duration of silence, and the counterstimulation markers, with current session information taking precedence over historical preference information.

[0033] Preferably, the constraint scheduling module includes:

[0034] The output constraint set is generated based on the stage judgment items and historical association information. An executable strategy is selected from the preset strategy family according to the content category boundary, tone level boundary, tactile mode boundary, action mode boundary, insertion timing condition and waiting time condition.

[0035] When the switching gating conditions are met, the strategy is switched and the process is scheduled between the candidate output state, the execution output state, and the observation freeze state.

[0036] Preferably, the write-back update module includes:

[0037] Based on the execution record entries and the collected entries after feedback, a feedback observation window is established to determine the effects of force changes, rhythm changes, semantic changes and succession relationships, and to generate in-session update structure and historical update structure.

[0038] Adjust the acceptability of strategy families, tone levels, haptic patterns, and action patterns according to progressive rules;

[0039] Write the updated results to the current session's acceptance table and the long-term preference table;

[0040] When the anti-stimulation condition is met, a policy rollback signal is sent to the constraint scheduling module.

[0041] Compared with existing technologies, the present invention provides an AI voice intelligent interaction system for decompression toys, which has the following beneficial effects:

[0042] 1. This invention, through unified collection, association, merging, stage determination, constraint scheduling, and write-back updates of toy force information, user voice information, semantic expression information, and historical interaction states, enables collaborative recognition of changes in force behaviors such as squeezing, patting, and kneading, as well as rhythmic changes such as speech rate, volume, pauses, and crying, and semantic changes such as complaining, venting, and repetitive expressions during the user's catharsis process. This allows for continuous judgment of the current stress relief stage and emotional risk level. Based on the stress relief stage and emotional risk level, the invention controls the type of voice response, tone intensity, tactile feedback method, action feedback method, and strategy switching timing. It also updates and corrects the strategy continuity relationship based on post-feedback state changes, avoiding mismatches in technical response timing, soothing strategies, inconsistent feedback intensity, and interruptions to venting. This achieves sustained stress relief, effective soothing, and stable companionship interaction, thereby improving the sustainability of stress relief, the effectiveness of soothing, and the stability of companionship interaction.

[0043] 2. This invention continuously observes the user's force, voice, semantics, and strategy continuity after each feedback, and writes this information into a conversation continuity table and a long-term preference table. This determines whether the current interaction is slow, short-lived, or escalates again. Based on the judgment results, subsequent soothing strategies are maintained, downgraded, rolled back, or reselected. This reduces the problems of disconnect between interactions, frequent occurrence of similar strategies, and distortion of historical preferences in existing technologies, and improves the continuity of strategy adjustment, the accuracy of individual adaptation, and the traceability of multi-round interactions. Attached Figure Description

[0044] Figure 1 This is a schematic diagram of the structure of an AI voice intelligent interaction system for decompression toys according to the present invention. Detailed Implementation

[0045] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0046] Example 1: Figure 1 An AI voice-based intelligent interaction system for decompression toys is presented, including:

[0047] Interactive data acquisition module: used to collect toy force information, user voice information, and session status information, specifically implemented as follows:

[0048] The interactive acquisition module is located in the main control board of the decompression toy and is connected to the force sensing layer, the sound pickup unit, the posture detection unit, and the conversation management unit. It is used to collect, filter, segment, merge, and itemize the force information of the toy, the user's voice information, and the conversation status information according to a unified time standard after the user comes into contact with the decompression toy, so as to generate basic interactive data for the status merging module to call. The interactive acquisition module does not directly output the raw continuous waveform, but performs event processing inside the module first, so that information from different sources has a unified event boundary, a unified time base, and a unified data format before entering the subsequent processing link.

[0049] After power-on, the interactive acquisition module enters a sampling state, continuously receiving raw inputs from the force acquisition channel, sound acquisition channel, posture detection channel, and system clock channel, and using the system clock as a unified reference for time alignment. The force acquisition channel outputs force change information for different areas of the toy body. The force acquisition area can be selected to cover the chest and abdomen area, and extended to the head, back, and lateral limb areas. The chest and abdomen area is chosen as the core acquisition area because it is easier to form continuous contact during decompression actions such as hugging, squeezing, and kneading, and can more stably reflect the intensity and duration of the release. The head, back, and lateral limb areas are used to supplement the distribution characteristics of patting, lateral kneading, and hugging actions. Each force acquisition unit outputs at least instantaneous force values ​​and duration markers. The sound acquisition channel can use a near-field pickup unit. The system is used to collect user voices near the toy and output the start and end times of voice activities, segment duration, and volume offset values ​​relative to background noise. The posture detection channel can use an acceleration detection unit or a posture change detection unit to identify abnormal interaction states such as falling, violent shaking, and moving. The session management unit outputs at least the current session number, session start time, most recent interaction time, current silence duration, and cache integrity flag. Depending on the operating load, it can also output resource usage levels, which are divided into three levels: low, medium, and high, corresponding to normal acquisition mode, limited acquisition mode, and degraded acquisition mode, respectively. The low level indicates that the main control board has complete acquisition and processing capabilities, the medium level indicates that the main control board is suitable for maintaining core sampling and basic segmentation, and the high level indicates that the main control board prioritizes maintaining session continuity and suppresses unnecessary field writing.

[0050] Before establishing a valid interaction, the interaction acquisition module sets an initial observation window to distinguish between genuine decompression contact and non-targeted interactions such as accidental touches, moving, and placement. The initial observation window can be 5 to 12 seconds, with 8 seconds being preferred. This setting is because if the window is too short, the user's initial tentative holding and short vocalizations are not easily captured completely, which can easily lead to misjudgment or omission. If the window is too long, the system's response to the establishment of genuine interactions is too slow, which is not conducive to the immediate acceptance of decompression scenarios. Within the initial observation window, the interaction acquisition module separately counts valid force events, valid voice segments, and the overlap between the two types of events. At least 3 valid force events or the cumulative duration of valid voice segments must appear in any core area or extended area. When the duration exceeds 2 seconds, or when there is a time overlap of more than 1 second between the force event and the voice event, the interaction is considered established, and the system switches from the waiting-to-sample state to the session establishment state. The effective voice cumulative duration is set to 2 seconds because sounds shorter than 2 seconds are more likely to correspond to short exclamations, accidental environmental sounds, or occasional speech, making it difficult to stably represent continuous interaction. The overlap duration between the force event and the voice event is set to 1 second because this duration can exclude false triggers caused by the superposition of accidental touches and environmental speech, while also covering the common duration of complete short sentences uttered during hugging or squeezing. If none of the above conditions are met, the interaction acquisition module remains in the waiting-to-sample state, only continuing to perform basic listening and background baseline updates, without triggering session establishment.

[0051] Valid force events are determined using both force and duration thresholds. The effective force threshold can range from 0.8 N to 1.5 N, with 1.0 N being preferred. The lower limit of 0.8 N is chosen because forces below this value more often correspond to light touches, clothing friction, natural deformation of toys, or picking / putting actions. The upper limit of 1.5 N is chosen because this value covers light holding and initial compression, making it less likely to miss genuine decompression contact. The duration threshold can be no less than 0.2 seconds; peak fluctuations shorter than this duration are not counted as valid events to exclude transient impacts and sensor noise. Valid voice events are determined using… The background baseline is used for determination by adding amplitude offset. The background baseline can be obtained by statistically analyzing the ambient sound over the most recent 2 to 3 seconds under the sampling state. The effective speech threshold can be the average background noise plus 6 dB. 6 dB is used as the offset threshold because this difference can distinguish between background noise fluctuations in the same environment and the effective sound emitted by the user in the near field of the toy, while avoiding the threshold being too high and missing low-frequency expressions. For high-noise usage environments, without changing the determination logic, the offset threshold can also be the average background noise plus 8 to 10 dB, but under normal usage environments, 6 dB is still the preferred value.

[0052] The basic interactive data generated by the interactive acquisition module is recorded using an entry-based data structure. Force acquisition entries include at least the following fields: timestamp, region number, average force value, peak force value, duration, and event validity marker. The timestamp is used for unified sorting with voice acquisition entries and session state entries; the region number identifies the force source region; the average and peak force values ​​characterize the sustained and instantaneous intensity of the action, respectively; the duration distinguishes between long-term holding and short-term patting; and the event validity marker indicates whether the entry meets the minimum conditions for entering the subsequent processing chain. Voice acquisition entries include at least the following fields: timestamp, start time, end time, segment duration, relative volume level, and near-field marker. The relative volume level describes the intensity change relative to the background baseline and can be selected as light. Four levels of volume are available: normal, enhanced, and intense. Near-field markers distinguish user voices near the toy from far-field sounds from the external environment. High-confidence near-field speech segments can be selected from those with a relative volume level above normal and a duration meeting the minimum effective speech duration. Session status entries must include at least the session number, current status, cumulative interaction duration, silence duration, and cache integrity flag. Cumulative interaction duration indicates the duration of the current session since its inception; silence duration indicates the duration of no interaction after the most recent effective interaction event; and cache integrity flag indicates whether the current session's cache data is continuous and complete. When the resource occupancy level field is enabled, a resource occupancy level can be appended for subsequent modules to use simplified processing paths when resources are limited.

[0053] The interactive acquisition module internally employs a segmentation-merging processing method, rather than directly statistically analyzing the raw signal. For the force channel, continuous force fluctuations are first screened according to the effective force threshold, and then merged according to the event interval. Force wavebands that are continuously higher than the effective force threshold and whose adjacent fluctuation intervals do not exceed 0.15 seconds can be merged into the same force event. The reason for setting 0.15 seconds as the force merging interval is that fluctuations shorter than this interval mostly belong to the natural fluctuations in the same squeezing or kneading process. If they are split into multiple events, it is easy to cause the force event density to be too high. For the voice channel, continuous sound streams are first segmented according to the effective force threshold. The system detects speech activity using pitch thresholds and then merges it according to segment intervals. Speech segments with an interval of no more than 0.3 seconds between adjacent vocal segments can be merged into the same expression segment. The reason for setting the speech merging interval at 0.3 seconds is that users' emotional expressions are often accompanied by short pauses, inhalations, and intermittent vocalizations. If the merging interval is too short, it is easy to divide a complete expression into multiple segments, affecting subsequent prosodic organization and semantic continuity. After the segmentation and merging are completed, the interaction acquisition module caches the force events, speech segments, and conversation states in parallel along a unified timeline, so that information from different sources within the same time period forms a basic interaction set that can be compared.

[0054] To support the windowed processing of the subsequent state merging module, the interactive acquisition module also established force event density parameters, voice event density parameters, and session activity parameters. Force event density represents the number of valid force events per unit time, for example, it can be counted every 5 seconds. Voice event density represents the cumulative duration of valid voice segments per unit time, for example, it can be counted every 5 seconds. The session activity parameter represents the joint existence ratio of force events and voice events in the last 10 seconds, with a value from 0 to 1, where 0 indicates no joint existence events in the last 10 seconds, and 1 indicates a high probability of joint existence events in the last 10 seconds. There is a persistent overlap between force and speech activity; the statistical period for force event density and speech event density is set to 5 seconds because this duration is sufficient to reflect the concentration of short-term user actions and short-phrase expressions, while not weakening the ability to recognize changes in the interaction rhythm due to an excessively long period; the statistical duration for session activity parameters is set to 10 seconds because this duration can cover the common transition process from tentative contact to stable expression or from expression to pause; after the basic statistics are completed, the above parameters, along with the entry summary, are written into the session management cache for use in the next sampling period and subsequent modules;

[0055] The interactive acquisition module forms a complete acquisition closed loop; the input end continuously receives raw signals from multiple channels, the processing end completes threshold filtering, event segmentation, fragment merging, field generation, and window statistics based on a unified time base, the output end forms a set of basic interactive items and a set of basic statistical parameters, and the write-back end writes the current window statistical results and cache integrity status into the session management cache area; through this processing link, the interactive acquisition module completes the unification of time standards, event standards, and session standards, enabling subsequent modules to directly perform correlation processing according to the unified standards;

[0056] When a sudden increase in environmental noise, a toy falling, sensor unit jitter, or a brief loss of synchronization in the session buffer occurs during sampling, the interactive acquisition module handles the situation according to preset anomaly rules. A sudden increase in environmental noise can be defined as a background noise increase exceeding 10 decibels within 2 seconds. The observation duration is set to 2 seconds to exclude transient fluctuations caused by single collision sounds, coughs, or brief door knocks. The change threshold is set to 10 decibels because this magnitude usually indicates a significant change in the acquisition environment, rather than general background fluctuations. When a sudden increase in environmental noise occurs, the voice acquisition channel enters a restricted mode, retaining only near-field high-confidence voice segments and temporarily not outputting low-confidence voice entries with unclear boundaries. The toy falling status can be reset simultaneously through global acceleration transients and the force applied to the entire area. The system will determine whether such events are non-user-initiated decompression actions and will not include them in the valid force entries. Sensor unit jitter can be defined as at least five consecutive amplitude abrupt changes within one second on a single channel, with no synchronous force changes in adjacent channels. The reason for using one second and five abrupt changes as the criteria is that normal holding, squeezing, or patting is usually accompanied by continuous contact in adjacent areas or time, while high-frequency isolated jumps on a single channel are closer to sensor jitter or poor contact. Channels identified as jittering will be temporarily downgraded within the current window, retaining only the original record and not participating in the valid event statistics. When the session cache briefly loses synchronization, a time base reset will be triggered. Data acquisition will continue until the reset is complete, but session creation and interaction entry will not be triggered to avoid event mismatching due to cached time stamp errors.

[0057] Regarding time constraints, the force sampling period can be set to no more than 50 milliseconds, the voice activity detection refresh period can be set to 200 to 500 milliseconds, the conversation state refresh period can be set to 1 second, and the total latency for single-cycle acquisition and processing can be set to no more than 300 milliseconds. The reason for setting the force sampling period to no more than 50 milliseconds is to capture short-term peak changes during slapping, rapid squeezing, and kneading. The reason for setting the voice activity detection refresh period to 200 to 500 milliseconds is to ensure timely identification of the start and end of speech without increasing the processing burden due to excessively frequent refreshes. The reason for controlling the total latency for a single cycle to within 300 milliseconds is to ensure that subsequent state merging and phase... The system determines whether to complete processing while the user is still in the current action and expression state, avoiding significant interaction delays. Regarding resource constraints, the interaction acquisition module can be configured to save only high-frequency interaction entries from the last 120 seconds and low-frequency session summaries from the last 30 minutes. The reason for saving 120 seconds as the high-frequency entry duration is that this duration is usually sufficient to cover the continuous process from a high-intensity outburst to initial subsidence. The reason for saving 30 minutes as the low-frequency session summary duration is that this duration is mainly used for cross-session transitions within the same usage period, rather than for long-term preference history storage. This approach retains the recent information required for the current session while avoiding the storage pressure caused by continuously saving high-frequency raw data.

[0058] State merging module: Used to associate and organize force changes, phonetic prosody changes, semantic expression changes, and interaction history states. Specifically, it is implemented as follows:

[0059] The state merging module, based on the unified time benchmark established by the interactive acquisition module, correlates and organizes force acquisition entries, voice acquisition entries, conversation state entries, and historical state entries. To ensure that information from different sources is merged under the same criteria, the state merging module establishes a unified merging window. This unified merging window can be divided into two levels: a short window and a medium window. The short window is used to organize immediate changes, with a duration of 2 to 5 seconds, preferably 3 seconds. The medium window is used to organize continuous trends, with a duration of 15 to 30 seconds, preferably 20 seconds. The reason for choosing the above range for the short window is that this duration can cover the minimum interaction of a complete squeeze combined with a short expression. The reason for using the medium window as the above range is that its duration can cover a continuous trend segment of emotional rise, pause, or fall; the short window and the medium window together constitute the time scope of the state merging module; historical state entries should include at least the duration of the last 3 conversations, the sequence of the main stages of the last 3 conversations, the frequency of interactions in the last 7 days, the acceptance record of past reassurance methods, and the protection mode trigger history markers, etc.; when processing the interaction information in the current window, the state merging module includes both the previous entries in the current conversation and the historical state entries in the merging scope, so that the state in the current time slice reflects both immediate changes and the interaction context and recent characteristics;

[0060] The state merging module organizes the force acquisition entries by short window, forming force change entries. Each force change entry includes at least the short window start and end time, main force area, number of events, average force value, peak force value, sustained pressure ratio, area switching frequency, and behavior type label. The main force area represents the area with the highest force contribution within the current short window; the sustained pressure ratio represents the proportion of force states above the medium pressure threshold within the current short window; and the area switching frequency represents the speed at which the force center switches between different areas. Behavior type labels can be selected from five categories: holding, squeezing, patting, kneading, and mixed. When the duration of medium pressure in the chest and abdomen area exceeds 50% of the short window duration and there are fewer than two peak jumps, it can be categorized as holding. When force events above the medium pressure threshold occur consecutively in the same area with an interval of less than 0.5 seconds, it can be categorized as holding. The following conditions are considered: Squeezing; when the peak value is high, the duration is short, and the area switching frequency is fast, it can be classified as patting; when moderate force events alternate between the left and right sides and the rhythm is relatively stable, it can be classified as kneading; other states can be classified as mixed. The reason for setting the duration ratio to over 50% is to ensure that the corresponding force state dominates within the current short window; the reason for setting the peak value jump to less than 2 times is that the holding state usually exhibits continuous stability rather than frequent impacts; the reason for setting the event interval to less than 0.5 seconds is that continuous squeezing usually exhibits relatively close repetition; the medium pressure threshold can be set to 2.0 Newtons to 3.5 Newtons, preferably 2.5 Newtons; the reason for setting the medium pressure threshold to the above range is that this range can distinguish between ordinary holding and more obvious cathartic force, and will not trigger the strong behavior type classification too early due to the threshold being too low.

[0061] The state merging module further organizes the prosodic variations of the voice acquisition entries to form speech prosodic entries. Speech prosodic entries include at least the following fields: average volume level, direction of volume change, speech rate level, pause density, prolongation marker, discontinuous vocalization marker, crying voice marker, and vocal continuity marker. The average volume level can be selected from four levels: soft, normal, raised, and intense; the speech rate level can be selected from four levels: slow, normal, slightly fast, and rapid. The direction of volume change indicates the rise, fall, or maintenance of volume between adjacent short or medium windows; the vocal continuity marker indicates whether there is sustained expression within the current window; and the pause density is recorded as the number of significant pauses per 10 seconds, used to characterize whether the expression tends to be rapid, fragmented, or declining. Long vowel markers are used to identify prolonged final vowels or elongated sighs; intermittent vocalization markers are used to identify multiple interruptions in the middle of a speech; and crying voice markers are used to identify vocalization states with obvious spitting, increased nasality, or vocal trembling. Speech prosody entries can be generated by short windows to describe immediate prosodic features, or by medium windows to describe continuous trend changes. If the average volume increases from normal to elevated, the speech rate increases from normal to slightly faster, and the pause density decreases within the same medium window, the prosodic change direction of that medium window is recorded as intensification. If the average volume decreases, the pause density increases, and the crying voice marker persists, it is recorded as decline accompanied by fragile expression. Through the above processing, speech information is transformed from a simple audible or silent state into a prosodic change state that can be called upon for stage determination.

[0062] The state merging module employs a limited category merging method for semantic expression changes. Input for semantic expression changes can come from speech-to-text transcription results or from keyword entries generated from speech-to-text. The state merging module merges the semantic content within the current window into a limited semantic category, ensuring that the semantic state can stably enter subsequent processing steps. Semantic expression entries must include at least the following fields: timestamp, main semantic category, repetition marker, reinforcement marker, negation marker, help-seeking marker, and end-of-expression marker. The main semantic category can be selected as emotional venting, event narration, request for comfort, rejection of comfort, self-denial, or blame from others. Repeated complaints, calming confirmations, and ending expressions; repetition expression markers are used to indicate that semantic content of the same or similar categories appears repeatedly within the middle window; reinforcement expression markers are used to record words of intensity such as "very," "always," and "fundamentally"; negation expression markers are used to record exclusionary expressions such as "don't," "useless," and "don't mention"; help-seeking expression markers are used to record content with a clear request for help; and ending expression markers are used to record content indicating that the interaction is about to end. By performing limited merging of semantic categories, the content state within the same time slice can establish a correspondence with changes in force and rhythm, avoiding the original text being too discrete and weakening the consistency of subsequent judgments;

[0063] The state merging module also incorporates previous feedback records and historical state entries from the current session into the current time slice, forming an interaction history association entry. This interaction history association entry includes at least the following fields: previous feedback type, direction of force change after feedback, direction of volume change after feedback, historical preference label, number of times it takes effect, and the most recent counter-stimulation marker. The previous feedback type indicates whether the previous round used a confirmation, cathartic, rhythmic, or other feedback method. The direction of force change and the direction of volume change after feedback record the decrease, increase, or maintenance of the user's force and voice volume after the previous round of feedback, respectively. The historical preference label indicates the user's preference for different soothing methods, such as a preference for gentle vibration, a preference for short voice responses, or a preference for longer silent acceptance. The number of times it takes effect records the number of times the corresponding preference or feedback method has been positively validated in the historical session. The most recent counter-stimulation marker records the most recent instance of a negative change caused by a mismatch in feedback methods. By introducing interaction history association entries, the current merging result can simultaneously reflect the connection between the current input state and the previous output state.

[0064] The merged state structure output by the state merging module adopts a composite entry format. Each composite entry is generated around the same time slice and includes at least the time slice number, force change field, prosodic change field, semantic change field, historical association field, and modal integrity flag. The time slice number is used to identify the short window or medium window sequence number to which the current composite entry belongs, and the modal integrity flag is used to indicate whether all types of information are complete in the current time slice. When the force change field is valid but the prosodic change field is invalid, the modal integrity flag can be recorded as a partially valid force type. When the prosodic change field is valid but the force change field is invalid, it can be recorded as a partially valid speech type. When both the force change field and the prosodic change field are valid, it can be recorded as a jointly valid type. If the semantic change field is missing but both the force change field and the prosodic change field are valid, it can still be recorded as a jointly valid type, with an additional explanation of the missing semantics. The reason for setting the modal integrity flag is that subsequent stage judgments need to distinguish whether the current state is formed based on complete interactive information or based on a single modality, so as to adopt normal judgment or conservative judgment respectively.

[0065] The state merging module sets thresholds for the number of force jumps, continuous expression, and repeated complaints to unify the density of changes within the constraint window. The force jump threshold can be set to 3 to 6 times every 3 seconds, with 4 times being preferred. The reason for setting the statistical duration to 3 seconds is that this duration matches the preferred short window diameter, facilitating direct judgment of jump density from the short window results. The reason for setting the threshold range to 3 to 6 times is that exceeding this range usually corresponds to rapid venting, while below this range it is more likely to manifest as ordinary hugging or slow kneading. The continuous expression threshold can be set to a single voice segment lasting longer than 4 seconds. The reason for using 4 seconds as the threshold is that exceeding this duration usually indicates that the user has moved from short, tentative phrases to sustained or prolonged expression. The state of continued venting; the threshold for repeated complaints can be set to the same main semantic category appearing no less than 3 times within a 20-second window; the reason for setting the statistical window to 20 seconds is that this duration is consistent with the preferred window size, which can cover one content retention process; the reason for setting the number of occurrences to no less than 3 times is that when this frequency is reached continuously, it can more stably represent that the user's current content is in a repeated retention state, rather than a normal event description; the above parameters together constitute the merging parameter system of the state merging module, where the window parameter is used to determine the merging time range, the behavior classification parameter is used to classify the type of force-induced behavior, the expression density parameter is used to characterize the persistence of speech and semantics, and the running constraint parameter is used to limit the merging output cycle and cache size;

[0066] The state merging module employs a dual-path approach: forward output and write-back update. Forward output sends composite entries to the stage determination module, while write-back update writes the composite entry summary into the current session's inheritance table for use when constructing the next intermediate window. The current session's inheritance table retains at least the summaries of the most recent short window composite entries, the summaries of the most recent intermediate window trend entries, and the summary of the most recent feedback-related entries, enabling the next round of merging to reference the previous state within the session, rather than starting the process anew each time. Through forward output and write-back update, the state merging module continuously provides new state foundations to subsequent modules and maintains the continuous inheritance relationships within the current session.

[0067] When boundary conditions occur, the state merging module processes according to preset rules; when there is only a force event but no effective speech, merging is not stopped, but instead, a no-speech-expression state entry is formed, and the speech missing mark is retained; when there is only speech but no force event, a no-force-expression state entry is formed, and the force missing mark is retained; when historical states are missing, the current session priority merging mode is switched, without waiting for historical data to be recovered; when semantic entries and prosodic entries conflict in direction, for example, the semantic content is calm but the volume is significantly increased, the conflict state mark is retained; the conflict state mark is used to indicate that there is a directional inconsistency between different modalities, and subsequent stage judgments can be based on this mark to take conservative processing, rather than directly drawing stage conclusions based on a single modality; if the number of entries in the short window is insufficient to form a stable merge, for example, the force collection entry only appears once and the speech entry duration is too short, then the current short window only forms a partial merged entry, and is not directly upgraded to the medium window trend basis;

[0068] Regarding time constraints, the state merging module can output short-window merged states once per second and correct medium-window trend states once every 5 seconds, with a total merging latency not exceeding 500 milliseconds. The reason for using 1 second as the short-window output cycle is that, although 3 seconds is preferred, a shorter output cycle is needed to maintain state continuity when using a sliding update method. The reason for using 5 seconds as the medium-window correction cycle is that it allows for phased correction of the trend states within a 20-second medium window without excessively increasing the processing load. The reason for controlling the total merging latency within 500 milliseconds is that the phase determination module needs to respond promptly to user input. The merged result is received while the previous actions and expressions are still continuous. In terms of resource constraints, each session can keep a maximum of 40 medium-window entries and 120 short-window entries online, and the remaining entries are transferred to the summary cache. The reason for setting the number of medium-window entries to 40 is that, based on the preferred 20-second medium window, it can cover the main trend changes in a longer interactive session. The reason for setting the number of short-window entries to 120 is that, based on the preferred 3-second short window combined with a 1-second sliding output cycle, it can cover a more complete history of high-frequency changes. The merged composite entries, along with modal integrity markers and conflict status markers, are output to the stage determination module.

[0069] Stage Determination Module: Used to determine the current decompression stage and emotional risk level based on the merged interaction status. Specifically, it is implemented as follows:

[0070] The stage determination module uses the short-window merged state and medium-window trend state output by the state merging module as a basis to jointly determine the current decompression stage and emotional risk level. The stage determination adopts a processing method that prioritizes the current conversation and is supplemented by historical state, using changes in force, rhythm, semantics, and historical correlation information within the current time slice as the direct basis for judgment. The current decompression stage can be divided into the contact establishment stage, catharsis establishment stage, high arousal catharsis stage, decline transition stage, stabilization and soothing stage, and recovery and end stage. The emotional risk level can be divided into low risk, concern risk, enhanced soothing risk, and protective intervention risk.

[0071] The contact establishment phase indicates that the user has established continuous contact with the toy, but catharsis has not yet fully developed. This phase can be established when the main force-bearing behavior within the middle window is mainly hugging or gentle squeezing, the vocal rhythm is mainly normal volume and short sentences, the repetitive complaint threshold has not been reached, and the previous state was not in the high-arousal catharsis phase. The reason for using the previous state not being in the high-arousal catharsis phase as a criterion is that if the previous state was already in the high-arousal catharsis phase, the current brief decline is more suitable as a transitional state, rather than being reclassified as the initial contact. The catharsis establishment phase indicates that the expression of negative emotions begins to concentrate, and both actions and vocalizations show continuous changes. This phase can be established when squeezing or slapping behaviors are frequent within the middle window. The high-arousal catharsis stage is defined as follows: The volume level increases from soft or normal to elevated, the speech rate increases from normal to slightly faster, and semantic entries related to emotional catharsis or event narration appear consecutively at least twice. The reason for setting the consecutive occurrence count at least twice is that a single occurrence is more likely to correspond to sporadic expression, while two or more consecutive occurrences indicate that the user has entered a stable expression state. The high-arousal catharsis stage indicates that the user has entered a high-intensity catharsis state. This stage is defined as follows: Within two consecutive short windows, there are more than three high-intensity stress events, the voice volume increases or becomes intense, the speech rate becomes slightly faster or more rapid, and at least one of the semantic entries—repeated complaints, self-denial, or strong refusal to be comforted—appears. The observation scope... The reason for using two consecutive short windows is to avoid misjudgment caused by occasional force exertion within a single short window; the reason for setting the cumulative number of times to 3 or more is that continuously reaching this frequency better indicates that the high-intensity venting has formed a sustained state; the fallback transition phase is used to indicate that the high arousal state has begun to weaken but still needs to be maintained; this phase can be established when the peak force decreases by 20% to 40% compared to the previous medium window, preferably more than 30%, the volume decreases by at least one level, the pause density increases, and the user still maintains contact or intermittent expression; the reason for setting the peak force decrease within the above range is that it is difficult to stably indicate a fallback trend when it is below 20%, while it is more likely to indicate a significant fallback when it is above 40%. The process of moving away from a high-intensity emotional state and stabilizing the user's speech is as follows: The stabilization and calming phase indicates that the user is able to tolerate a gentler, more soothing output. This phase can be established when there are no high-intensity stress events within a continuous medium window, the speech rate returns to normal or slow, calming confirmation or soothing acceptance semantic items appear, and the anti-stimulation flag is not triggered. The recovery and ending phase indicates that the current interaction should naturally come to an end. This phase can be established when the silence duration reaches 20 to 60 seconds, preferably 35 seconds, or when the end expression flag is triggered. The reason for taking the silence duration within the above range is that if it is less than 20 seconds, it is easy to misjudge the pause as the end, and if it is more than 60 seconds, the system will maintain an unnecessary activity state for a long time.

[0072] The emotional risk level is determined synchronously with the stage assessment. Low risk corresponds to the contact establishment stage and a partially stable reassurance stage; concern risk corresponds to the catharsis establishment stage and a partially declining transition stage; enhanced reassurance risk corresponds to the high arousal catharsis stage; and protective intervention risk corresponds to a high arousal catharsis state that lasts for 10 to 30 seconds, preferably more than 20 seconds, with persistent reassurance rejection and the simultaneous appearance of urgent concern expressions. The reason for setting the duration within the above range is that instantaneous escalation does not necessarily require entering protective restrictions, while a continuous period without decline is more suitable for triggering stronger protective control. Urgent concern expressions can be expressions of seeking help, expressions of self-harm tendencies, or any of the following semantic items: continuous high-intensity denial, self-deprecation, or dangerous tendencies. The reason for this setting is that protective intervention risk is based on the simultaneous existence of duration, refusal to reassure, and dangerous semantics, rather than being directly triggered by a single abnormal expression.

[0073] To ensure clear decision boundaries, the stage decision module employs a state machine to organize stage switching. The initial state of the state machine is the pending decision state. Upon receiving the first jointly valid composite entry, it can enter the contact establishment stage. If the force density, speech intensity, and cathartic semantic density all increase simultaneously within two consecutive short windows, it switches from the contact establishment stage to the cathartic establishment stage. If, within the cathartic establishment stage, any two of the three conditions—high-intensity force, rapid rhythm, and high-density cathartic semantics—are consistently met for at least 5 seconds, it switches to the high-arousal cathartic stage and simultaneously increases the risk level. The reason for setting the duration to at least 5 seconds is that short... At this duration, it may still be considered a short-term escalation, insufficient to support risk escalation; if at least two of the following three conditions occur after the high-arousal catharsis phase: decreased force, decreased volume, and increased pauses, then switch to the fallback transition phase; if the fallback transition phase lasts for a medium window and the high-intensity catharsis condition is not triggered again, then enter the stabilization and calming phase; if the silence threshold is reached during the stabilization and calming phase, or the end expression marker is triggered, then enter the recovery and end phase; if a valid squeeze and valid speech overlap event reappears during the recovery and end phase, then return to the contact establishment phase; through state machine expression, the phase switching conditions are fixed on the reviewable branch conditions, rather than relying on abstract judgments;

[0074] The phase determination module sets phase identification parameters, phase switching parameters, risk maintenance parameters, and runtime sequence parameters. Phase identification parameters include at least a high-intensity force event threshold, a silence duration threshold, and a threshold for the consecutive occurrences of cathartic semantics. The high-intensity force event threshold can be between 3.5 Newtons and 7.0 Newtons, preferably 4.5 Newtons. The reason for choosing this range for the high-intensity force threshold is that it can reliably distinguish between general hugging and cathartic force, and forms a hierarchical relationship with the light pressure threshold and medium pressure threshold in the previous module. Phase switching parameters include at least a phase switching stabilization time and a fallback confirmation time. The phase switching stabilization time can be between 3 seconds and 8 seconds, preferably 5 seconds. The reason for choosing this range for the phase switching stabilization time is that it can suppress back-and-forth switching caused by short-term fluctuations without delaying the actual phase change due to excessive waiting time. The fallback confirmation time can be between 10 seconds and 20 seconds, preferably 15 seconds. The confirmation time is set within the above range to distinguish between a true fallback and a short pause, preventing users from quickly rebounding after a brief period of silence in a highly aroused state. The risk maintenance parameter includes at least a risk escalation maintenance time; this time can range from 5 to 12 seconds, with 8 seconds being preferred. The reason for setting the risk escalation maintenance time within this range is that a sudden escalation does not necessarily require escalation to protective control, while maintaining the status quo for this duration without fallback better indicates a persistent risk. The runtime sequence parameters include at least a stage judgment refresh cycle, a risk level verification cycle, and a result delivery time limit; the stage judgment refresh cycle can be no higher than 1 second, the risk level verification cycle can be 1 to 2 seconds, and the result delivery time limit after the stage switching decision is completed can be no more than 500 milliseconds. This setting ensures that stage and risk results are promptly transmitted to subsequent modules while the user's current actions and expressions remain continuous.

[0075] The results output by the stage determination module are recorded using structured determination entries. Each determination entry includes at least the following fields: current stage, current risk, entry time, duration, switching reason, and credibility level. The current stage field indicates the stage type after this round of determination; the current risk field indicates the corresponding risk level; the entry time field records the time when the stage or risk first occurs; the duration field records the duration of the state; the switching reason field records the dominant condition that triggers the stage change, such as sustained high-intensity force, decreased volume, increased pauses, or the appearance of an ending expression; and the credibility level field indicates whether the current determination comes from a jointly valid entry or a partially valid entry. Credibility levels are divided into two levels: high credibility and limited credibility. Jointly valid entries correspond to high credibility, while partially valid entries correspond to limited credibility. The reason for setting the credibility levels in this way is that the subsequent constraint scheduling module needs to select normal scheduling or conservative scheduling based on the reliability of the determination results.

[0076] When anomalies and boundary conditions occur, the stage determination module processes them according to preset rules. If the composite entry in the current window only contains force changes without speech and semantic changes, the stage determination maintains a conservative judgment between the contact establishment stage and the catharsis establishment stage, and does not directly enter the high-arousal catharsis stage. The reason for this setting is that although simple force behavior can indicate changes in contact intensity, it is not enough to directly push up the risk level without prosodic and semantic support. If there are only speech and semantic entries without force entries, the upper limit of the risk level can be taken as the risk of concern, unless the expression of urgent concern continues to exist. If the force, prosodic and semantic state reflected by the current composite entry is inconsistent with the common succession pattern corresponding to the historical preference label, the current session determination takes priority, and the historical preference is only used as an auxiliary reference and does not cover the current state. If the composite entry is missing for a short time in the current update cycle, the previous stage remains unchanged, and the next short window update result is waited for to avoid stage jump caused by missing data in a single cycle. If the modal integrity mark is partially valid and conflict state marks exist at the same time, the original stage is maintained first or only a small adjustment is made, and cross-stage switching is not directly performed.

[0077] The time limit for distributing the results after the phase switching decision is completed can be no more than 500 milliseconds. The reason for controlling the time limit within the above range is that the constraint scheduling module needs to receive the phase and risk results while the current interaction state is still continuous. The output content should include at least the current phase field, the current risk field, the switching reason field, and the trust level field, and be sent to the constraint scheduling module.

[0078] The constraint scheduling module controls the content of voice responses, tone intensity, tactile feedback methods, action feedback methods, and the timing of switching soothing strategies based on the decompression stage and emotional risk level. Specifically, it implements this as follows:

[0079] The constraint scheduling module generates an output constraint set based on the stage judgment results and historical correlation information, and selects an executable strategy within the range defined by the constraint set. The parameters in the constraint scheduling module can be divided into output intensity parameters, insertion timing parameters, strategy switching parameters, state machine timing parameters, and running delay parameters. The output intensity parameters are used to limit the stimulation levels of speech, touch, and action; the insertion timing parameters are used to limit the time window for feedback; the strategy switching parameters are used to limit the frequency of strategy family changes; the state machine timing parameters are used to control the transition rhythm between candidate output, executed output, and observation freeze; and the running delay parameters are used to limit the timeliness of constraint generation and execution.

[0080] After receiving the judgment entries, the constraint scheduling module establishes an output constraint set. The output constraint set includes at least the following fields: content constraint field, tone constraint field, tactile constraint field, action constraint field, switching timing field, and waiting time field. The content constraint field is used to limit the allowed response categories for the current stage. The contact establishment stage corresponds to confirmation and short sentence continuation content; the cathartic establishment stage corresponds to continued expression and brief companionship content; and the high-arousal cathartic stage corresponds to low-stimulus confirmation, soothing rhythm guidance, and protective companionship content, while blocking follow-up questions, explanations, judgments, and repetitions. The content is divided into three categories: the transition phase corresponds to slowing down the rhythm, guiding breathing, and brief soothing content; the stabilizing and soothing phase corresponds to calming confirmation and low-density tidying content; and the recovery and ending phase corresponds to ending and concluding content. The content constraint field does not directly generate specific words, but rather limits the range of content categories that can be entered in the current phase. The reason for this setting is that although all phases belong to companionship interaction, the requirements for content density, content nature, and expression are significantly different between the high arousal catharsis phase and the recovery and ending phase. If the content boundaries are not established first, it is easy for the response content to be mismatched with the current state.

[0081] The tone constraint field is used to limit the speaking speed, volume, and emotional fluctuation intensity. Speaking speed can be divided into three levels: slow, normal, and slightly fast; volume can be divided into three levels: soft, normal, and raised; and emotional fluctuation intensity can be divided into three levels: low, medium, and high. During the high-arousal catharsis phase, only speaking speeds below normal are allowed, volume cannot exceed the upper limit of normal volume, and emotional fluctuation intensity is limited to a low level. During the catharsis establishment phase, a neutral and stable tone is allowed, outputting within the normal speaking speed and volume range, but without entering high fluctuation intensity. During the fallback transition phase, a tone of... Slow speech rate and low-fluctuation tone; a gentle and mild tone is allowed during the stabilization and soothing phase, but an excited tone and overly emotional rendering are prohibited; short sentences, low density, and low-fluctuation tone are preferred during the recovery and end phase; the reason for limiting the intensity of emotional fluctuations in the high arousal and cathartic phase to a low level is that the user is in a significantly agitated state at this stage, and if the emotional fluctuations in the voice are too strong, it is easy to create secondary stimulation; the reason for retaining a gentle and mild tone during the stabilization and soothing phase is that the user has a certain capacity to receive soothing information at this stage, but it is still not suitable to enter an overly active mode of expression;

[0082] The tactile constraint field is used to limit the vibration frequency, vibration amplitude, and duration. The vibration frequency can be divided into three levels: low frequency, medium frequency, and high frequency. The vibration amplitude can be divided into three levels: level 1, level 2, and level 3. During the high arousal catharsis phase, only low frequency and rhythmic medium frequency vibrations are allowed, while high frequency vibrations are prohibited. During this phase, the vibration amplitude is only allowed to be level 1 or rhythmic level 2, and level 3 vibration is not allowed. During the stabilization and soothing phase, the vibration amplitude can be adjusted between level 1 and level 2. During the recovery and end phase, the tactile feedback density is reduced first. The duration of a single vibration can be between 0.5 seconds and 3 seconds. The reason for setting the duration of a single vibration within the above range is that a duration shorter than 0.5 seconds is not conducive to forming a perceptible accompanying rhythm, while a duration longer than 3 seconds is prone to overlapping with the user's current action and causing additional interference. In the rhythmic accompanying state, the tactile output can be in the form of a 1-second vibration followed by a 1-second pause cycle. The reason for setting the cycle in this form is that it can provide a stable rhythmic cues without creating a continuous sense of pressure, and it is easy to gradually synchronize with the user's breathing or rhythm.

[0083] The action constraint field is used to limit the action type, action amplitude, and action duration. The action type can be selected from low-stimulation actions such as head turning, chest rise and fall, hugging and tightening, and ear swaying. The action amplitude can be 3 to 12 degrees, preferably not exceeding 8 degrees. The action duration can be 0.8 to 4 seconds. The reason for setting the action amplitude within the above range is that it is not easy for users to perceive it stably when it is less than 3 degrees, and it is easy to create obvious swaying or a sense of disturbance when it is greater than 12 degrees. The reason for setting the upper limit of 8 degrees is that the action feedback of the stress relief toy should be mainly a slight sense of companionship, rather than exaggerated performance. Sudden large swings are prohibited in the high arousal catharsis stage, and only low-stimulation actions such as slow chest rise and fall and small turning are allowed. Slow hugging and tightening actions can be added in the fall transition stage. The action frequency is reduced in the stabilization and soothing stage to avoid disturbing the user again when the user tends to be calm. In principle, no new active actions are triggered in the recovery end stage, and only necessary low-amplitude contraction actions are retained. If the current risk is a protective intervention risk, then swinging actions that exceed the upper limit of the action amplitude of the current stage are blocked.

[0084] The switching timing field is used to limit when voice can be inserted and when the soothing strategy can be switched; the constraint scheduling module prioritizes inserting voice during the phonation pause window or the force trough window; the phonation pause window can be 0.4 to 1.2 seconds of continuous no effective phonation, preferably 0.6 seconds; the reason for setting the lower limit of the pause window to 0.4 seconds is that intervals shorter than this length are mostly still within the same sentence expression and are not suitable for inserting feedback; the reason for setting the upper limit to 1.2 seconds is that after this length, the user is more likely to enter a new expression cycle or a silent state; the force trough window can be set when the force value in the most recent second is lower than the current middle window. The duration is set at 60% of the average value and lasts for more than 0.4 seconds. The reason for using the most recent 1 second as the observation range is that this duration can cover a short-term force drop process. The reason for using 60% of the current average value of the middle window as the trough condition is that this ratio can distinguish between general small fluctuations and obvious force drops. The reason for setting the duration to more than 0.4 seconds is to ensure that the voice insertion window has basic stability, rather than a false window formed by instantaneous jitter. Voice feedback only enters the candidate execution state when at least one type of window is established, in order to avoid interruption or forced interruption during continuous user speech or continuous high-intensity squeezing.

[0085] The waiting time field controls the minimum interval between the end of the previous feedback and the triggering of the next feedback. The minimum interval can be 2 to 8 seconds. The reason for setting the lower limit to 2 seconds is that if the duration is shorter than this, multiple rounds of feedback are likely to overlap, and the user has not yet completed the perception and response to the previous feedback. The minimum interval for the high arousal catharsis stage is preferably 4 to 6 seconds. The reason for selecting this range is that the user is still in the process of continuous catharsis during this stage, and too much feedback is likely to cause interruption. The minimum interval for the stabilization and soothing stage is preferably 6 to 8 seconds. The reason for selecting this range is that the feedback density should be reduced after the state stabilizes to avoid the toy making excessive noise or moving excessively and causing reverse disturbance. The waiting time field and the switching timing field work together to make the entry of voice, touch and movement simultaneously constrained by the content boundary and time rhythm.

[0086] The constraint scheduling module establishes a set of strategy families; the strategy families include at least the presence confirmation family, catharsis support family, destimulation support family, rhythm synchronization family, fallback guidance family, and termination convergence family, etc.; the presence confirmation family is used to maintain the presence of companionship with minimal stimulation in the contact establishment state or high arousal state; the catharsis support family is used to support the user's expression without interrupting the narration during the catharsis establishment phase; the destimulation support family is used to compress content density, reduce voice stimulation, and maintain companionship under the risk of strengthening soothing or protecting intervention risks; the rhythm synchronization family is used to guide the user from a rapid rhythm to a more gentle rhythm through slow speech, light vibration, or low-amplitude movements; the fallback guidance family is used to... The fallback trend is maintained during the fallback transition phase; the termination convergence family is used to complete the interactive convergence during the recovery termination phase; the constraint scheduling module does not directly generate specific sentences, but selects executable policies from the allowed policy families within the scope jointly defined by the current stage, risk level, trust level, and historical association fields, and then maps the policy to voice category, tone level, vibration mode, and action mode, and sends it to the corresponding execution unit; the policy number, voice category number, tone level, vibration mode, and action mode in the output data come from the preset output set corresponding to the currently allowed policy families, rather than being generated freely on the spot during scheduling;

[0087] Strategy switching must meet switching gating conditions. These conditions include at least the following: the new phase continuously reaches a stable switching time, the previous strategy observation result shows ineffectiveness or counter-stimulation, and the number of consecutive switching does not exceed the upper limit. The upper limit for the number of consecutive switching can be 2 to 4 times, with 3 times being preferred. The reason for setting the upper limit to the above range is that fewer than 2 times is not conducive to making necessary corrections to short-term ineffective feedback, while more than 4 times will cause frequent fluctuations in the output direction. The reason for setting the preferred value to 3 times is that this number can suppress strategy jitter while preserving adjustment space. If the number of consecutive switching exceeds the upper limit, the strategy family will not continue to be changed frequently, but will instead enter the de-stimulation companion family to stabilize the output direction. Strategy switching is also affected by the confidence level field. When the stage confidence level is restricted confidence, the constraint scheduling module adopts conservative output, only allowing low-stimulation existence confirmation strategies, and not entering high-density voice soothing strategies or multi-round guidance strategies. The reason for this setting is that if the current judgment is based on incomplete modality formation, it is not advisable to immediately adopt strong guidance or high-density output.

[0088] The constraint scheduling module uses an output state machine to organize the scheduling process. The initial state of the output state machine is a silent waiting state. When the current stage allows voice output and the waiting time is met, it enters the candidate output state. In the candidate output state, it checks whether the phonation pause window or the force wave trough window is valid. If at least one type of window is valid, it enters the execution output state. If the window condition is not met, it maintains the candidate output state until timeout. The candidate output timeout time can be 1 to 3 seconds, preferably 2 seconds. The reason for setting the timeout time within the above range is that if it is less than 1 second, a stable insertion window may not have been reached yet, and if it is more than 3 seconds, the current feedback opportunity will fail and the overall rhythm will be slowed down. If the candidate output timeout is still not valid, the current voice output is abandoned, and only tactile or light movement is retained. The system responds to avoid scheduling blockage caused by waiting for voice insertion; after the output state is completed, it enters an observation freeze state; the observation freeze state can last from 4 to 10 seconds, preferably 6 seconds; the reason for choosing the freeze duration within the above range is to allow observation space for the user's natural response after feedback, while restricting the scheduling side from immediately switching to a new high-density voice strategy; 6 seconds can provide sufficient time for the user to make the first round of behavior and vocal response to the current feedback without excessively slowing down the interaction; this freeze duration is used by the scheduling side to restrict high-density switching, and does not replace the effect observation cycle in the write-back update module; during the freeze period, low-density haptic reception is allowed according to user changes, but it does not enter a new high-density voice strategy, nor does it frequently switch strategy families;

[0089] The input data structure of the constraint scheduling module includes at least the following fields: stage field, risk field, confidence level field, historical preference field, and previous policy effect field. The stage field and risk field are derived from the stage determination module. The confidence level field is used to constrain the scheduling strength. The historical preference field is used to select a more suitable policy family. The previous policy effect field is used to determine whether to maintain the current policy or trigger a switch. The output data structure includes at least the following fields: policy number, voice category number, tone level, vibration mode, action mode, execution start time, execution end time, and observation freeze duration. The execution start time and execution end time are used to limit the time boundary of the current round's output. The observation freeze duration is used to limit the duration for which the scheduling side enters frozen observation after the execution ends.

[0090] When anomalies and boundary conditions occur, the constraint scheduling module handles them according to preset rules. If the voice execution unit malfunctions, the voice category is automatically masked, allowing only tactile and motion output. If the vibration unit malfunctions, the priority of low-density presence confirmation voices is increased. If the motion unit is stuck, the motion mode is set to empty and written to the fault field. If the current risk is a protective intervention risk, interrogative voices, lively tones, rapid vibrations, and swinging movements exceeding the upper limit of the current stage's motion amplitude are forcibly masked. If the current stage's confidence level is restricted confidence, and the previous strategy effect field is invalid or anti-stimulation, cross-family switching is not performed, and convergence to the presence confirmation family or destimulation companion family is prioritized. If the input fields are missing but basic stage and risk information is still available, the minimum constraint set is allowed to be generated in conservative mode to avoid the toy completely losing its responsiveness due to missing local fields.

[0091] Regarding time constraints, the delay from stage entry to output constraint generation can be no more than 300 milliseconds, and the delay from constraint generation to execution unit reception can be no more than 200 milliseconds; regarding resource constraints, the constraint scheduling module can be selected to save only the current policy family and the last 3 execution records, without saving the original text of large-scale content templates.

[0092] The write-back update module is used to record and update the state changes and strategy succession relationships after the interaction. Specifically, it is implemented as follows:

[0093] The write-back update module records, judges, and updates the state changes and policy continuity relationships after feedback based on the execution record entries and the collected entries after feedback. The execution record entries include at least the policy number, tone level, vibration mode, action mode, execution start time, execution end time, and freeze observation duration. The collected entries after feedback include at least the force events, voice segments, semantic entries, and conversation activity parameters within the observation segment. The parameters in the write-back update module can be divided into feedback observation parameters, effect judgment parameters, preference update parameters, cache retention parameters, and write-back timing parameters. The feedback observation parameters are used to limit the observation range after feedback, the effect judgment parameters are used to judge the continuity effect of the current policy, the preference update parameters are used to control the preference adjustment rhythm within and across sessions, the cache retention parameters are used to limit the storage range of high-value historical entries, and the write-back timing parameters are used to limit the timeliness of update result generation and feedback callback.

[0094] The write-back update module establishes a feedback observation window, which can be 4 to 10 seconds, with 6 seconds being preferred. This range is chosen to allow sufficient time after feedback to capture the user's initial force, vocalization, and semantic responses, while avoiding excessively long observation times that could slow down subsequent strategy adjustments. This observation window is used to assess the feedback effect and does not replace the freeze duration control in the constraint scheduling module. The write-back update module extracts force change, prosody change, semantic change, and sequence relationship fields from the observation window. The force change field records the peak force compared to before and after feedback. Whether the window is lowered, whether the density of force events is reduced, and whether the main behavior type changes from patting or squeezing to holding; the prosody change field is used to record whether the volume level decreases, the speech rate slows down, and the pause density increases; the semantic change field is used to record whether there are expressions of reassurance and acceptance, expressions of calming and confirmation, expressions of rejection of reassurance, or expressions of ending; the continuity field is used to record whether the current strategy execution forms a continuous continuity with the previous stage, whether it triggers a counterstimuli, and whether it is necessary to maintain, downgrade, or switch the strategy family; continuous continuity can be taken as no abnormal jump occurs in the current stage, and the direction of state change after feedback is consistent with the direction of target adjustment;

[0095] The write-back update module determines the effectiveness based on changes observed in the observation window. If at least two of the following four conditions are met: peak force decreases by at least 20%, volume decreases by at least one level, pause density increases, and reassurance / refusal expression does not increase, the current strategy is considered to have effectively continued. The reason for requiring at least two conditions is that a single change may be caused by occasional fluctuations, while two or more changes in the same direction better indicate a genuine easing trend after feedback. If only one condition is met, it is considered to have only moderately continued. The reason for considering only one condition as moderately continued is that a single improvement is not sufficient. Sufficient to stably support the establishment of the effect, but can be used as a reference for maintaining the current strategy or making slight adjustments; if any of the following conditions are met, such as an increase in high-intensity stress events, an increase in expressions of refusal to soothe, or an increase in risk level, it is recorded as a counterstimulation; the reason for using 20% ​​as the stress reduction threshold is that slight fluctuations are not enough to stably indicate that the feedback is effective, while a decrease of more than 20% can better reflect a short-term easing trend; a decrease in volume of at least one level, an increase in pause density, and no increase in expressions of refusal to soothe are used together to constrain misjudgments caused by a single change in stress, so that the judgment result takes into account the three directions of action, rhythm, and semantics simultaneously;

[0096] The update results generated by the write-back update module are divided into in-session update structure and historical update structure. The in-session update structure includes at least the following fields: current policy effect flag, post-feedback state direction, number of consecutive effective times, number of consecutive ineffective times, recent counter-stimulation flag, and next waiting time. The current policy effect flag is used to identify whether the current feedback is effective, moderate, or counter-stimulating. The post-feedback state direction is used to indicate whether the current interaction is generally declining, maintained, or intensified. The number of consecutive effective times and consecutive ineffective times are used to provide a basis for maintaining or switching the policy in the next round. The recent counter-stimulation flag is used to indicate the most recent negative feedback in the current session. The next waiting time is used by the constraint scheduling module to adjust the minimum feedback interval in the next round. The historical update structure includes at least the following fields: policy family acceptance, tone level acceptance, tactile pattern acceptance, action pattern acceptance, and stage matching preference level. Policy family acceptance, tone level acceptance, tactile pattern acceptance, and action pattern acceptance can all be divided into three levels: low, medium, and high. The stage matching preference level is used to indicate the historical adaptability of a certain type of output method in a specific stage.

[0097] Preference updates employ a gradual write-back approach rather than a single, large-scale rewrite. If the same strategy family forms effective acceptance in the same phase of at least two different conversations, the acceptance level of that strategy family in the corresponding phase is increased from low to medium, or from medium to high. If the same strategy family triggers counterstimuli twice consecutively, its acceptance level is downgraded. Acceptance updates for tone level, tactile patterns, and action patterns are executed according to the same directional rules. If the upgrade or downgrade conditions are not met, the original acceptance level remains unchanged. The reason for setting the upgrade or downgrade threshold to two times is that a single result may be affected by occasional environmental noise, short-term emotional fluctuations, or differences in execution conditions, while at least two consecutive pieces of evidence in the same direction better reflect the true fit. Through gradual write-back, the write-back update module can accumulate cross-conversation fit information and avoid excessive long-term preference shifts caused by a single abnormal interaction.

[0098] The write-back update module also maintains the current session continuity relationship. After each round of feedback, the write-back update module writes the current stage, current risk, current strategy, current effect, and observation results into the current session continuity table. The current session continuity table includes at least the stage change direction, risk change direction, strategy effectiveness, and counterstimulation records of the most recent rounds, which are used by the state merging module and the stage determination module in subsequent judgments. The state merging module prioritizes the summary entries and continuity direction in the continuity table, while the stage determination module prioritizes the effect markers and information on re-entering the high-arousal cathartic state from the fallback transition state. In this way, when a user becomes agitated again after the fallback transition state, the system can identify that the current state belongs to the re-agitation of the previous cathartic process, rather than mistakenly believing that a new session has begun.

[0099] The write-back update module uses an update state machine to organize its internal update process. The update state machine initially enters a waiting observation state, and then enters an observation and acquisition state upon receiving an execution record entry. If the observation window expires and the feedback changes meet the valid acceptance rules, it enters a positive update state. In the positive update state, the write-back update module increases the priority of the current strategy family and can appropriately extend the next waiting time to reduce unnecessary high-frequency feedback. If the observation window expires and only the acceptance general rules are met, it enters a lightweight update state. In the lightweight update state, it only maintains the current strategy priority and shortens the next verification cycle to continue observing whether further switching is needed. If the observation window expires and the anti-stimulation rule is met, it enters an anti-stimulation update state. In the anti-stimulation update state, an anti-stimulation flag is written, the priority of the current strategy family is reduced, and a strategy rollback signal is fed back to the constraint scheduling module. The strategy rollback signal is used to require the constraint scheduling module to prioritize switching to a similar strategy with lower stimulation or a reduced stimulation companion strategy in the next round, rather than laterally switching to a new high-density speech strategy. The reason for this setting is that when the current strategy has already shown a negative effect, continuing to increase the output intensity or frequently changing high-stimulation strategies is more likely to aggravate the mismatch risk.

[0100] When anomalies and boundary conditions occur, the write-back update module handles them according to preset rules; if no new valid collection entries are added in the observation window, no strategy effect conclusion is drawn, only insufficient observation is recorded, and the previous priority remains unchanged; if the session suddenly ends in the observation window, session termination write-back is executed first, only the current session inheritance table is updated without modifying long-term preferences; if communication interruption or local execution unit failure occurs after feedback, the effect is judged only on the still valid output channels, and the failure consequences are not directly attributed to the strategy itself; if input fields are missing but at least one of the core fields of execution record, force change, or rhythm change is retained, then it is allowed to be updated in the current session. In limited mode, a minimum update is performed to prevent the current round of feedback from being completely lost. If the historical cache space is insufficient, priority is given to retaining high-risk session summaries and valid acceptance entries from the most recent 14 days, while compressing low-risk short session summaries. The reason for choosing a retention period of 14 days is that this duration can cover users' recent interaction habits without weakening current preference changes due to an excessively long retention period. The long-term preference table can be set to retain only the 5 most recent high-reliability policy records for each stage. The reason for retaining 5 records is that fewer records are insufficient to reflect recent preference changes, while more records will dilute the role of recent high-reliability acceptance results in the current policy selection.

[0101] Regarding time constraints, the total time from the end of the feedback observation window to the completion of the update result write-back can be no more than 500 milliseconds, and the time for the update signal to be fed back to the constraint scheduling module can be no more than 200 milliseconds. Regarding resource constraints, the write-back update module can be selected to retain only the update status and instance of the current active session, the current session inheritance table and the long-term preference table summary, without saving large-scale original speech and high-frequency original force waveforms.

[0102] In this embodiment, when a user holds and squeezes a stress-relieving toy to vent their emotions verbally, the interaction acquisition module simultaneously collects force information, voice information, and conversation state information. The state merging module, according to a unified time window, correlates and organizes changes in force, voice rhythm, semantic expression, and historical interaction states. The stage determination module determines the user's current stress-relieving stage based on the organization results and simultaneously determines the emotional risk level. The constraint scheduling module then controls the voice response category, tone intensity, tactile feedback method, action feedback method, and strategy switching timing based on the stress-relieving stage and emotional risk level. The write-back update module continues to observe changes in the user's force, vocalization, and expression after feedback, judges whether the current strategy is effective, and updates the conversation continuity and historical preferences, thereby forming a continuous closed-loop interaction.

[0103] It should be noted that this invention can be deployed on the device itself to realize embedded applications, or it can run on a PC or other terminal with a user interface, thereby meeting various hardware environments and usage requirements.

[0104] The above embodiments can be implemented in whole or in part by software, hardware, firmware, or any other combination. When implemented in software, the above embodiments can be implemented in whole or in part by a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, the processes or functions of the embodiments of this application are implemented in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted wirelessly or wiredly from one website, computer, server, or data center to another website, computer, server, or data center. Wired methods include optical fiber, twisted pair, coaxial cable, etc. Wireless methods include infrared, microwave, etc. Available media include any available media that can be accessed by a computer or data storage devices such as servers and data centers that contain one or more sets of available media. Available media can be magnetic media (floppy disks, hard disks, magnetic tapes), optical media (DVDs), or semiconductor media. Semiconductor media can be solid-state drives.

[0105] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0106] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An AI voice intelligent interaction system for decompression toys, characterized in that, include: Interactive data acquisition module: used to collect force information from toys, user voice information, and session status information; State merging module: used to associate and organize force changes, speech rhythm changes, semantic expression changes, and interaction history states; Stage Determination Module: Used to determine the current decompression stage and emotional risk level based on the merged interaction status; The constraint scheduling module is used to control the content of voice responses, tone intensity, tactile feedback methods, action feedback methods, and the timing of switching soothing strategies based on the decompression stage and the level of emotional risk. Write-back update module: Used to record and update state changes and strategy inheritance relationships after interaction.

2. The AI ​​voice intelligent interaction system for decompression toys according to claim 1, characterized in that, The interactive data acquisition module includes: The inputs of the force sensing layer, the sound pickup unit, the posture detection unit, and the conversation management unit are aligned according to a unified time base; Threshold filtering, event segmentation, segment merging, and itemized recording are performed on force fluctuations and speech activities; Within the initial observation window, the session establishment is determined based on the force events, speech segments and their overlapping relationships, while pre-set handling is performed for abnormal noise, drop disturbances, channel jitter and time stamp out-of-sync.

3. The AI ​​voice intelligent interaction system for decompression toys according to claim 2, characterized in that, Threshold filtering, event segmentation, fragment merging, and itemized recording of force fluctuations and speech activity are performed, including: Forced wavebands that are continuously above the effective force threshold and whose adjacent fluctuation intervals are within a preset range are grouped into the same force event. Speech segments with adjacent vocal segments spaced within a preset range are grouped into the same expression segment, and force acquisition entries and speech acquisition entries are generated respectively.

4. The AI ​​voice intelligent interaction system for decompression toys according to claim 1, characterized in that, The state merging module includes: The state merging module performs temporal merging of force acquisition items, voice acquisition items, conversation state items and historical state items according to a unified merging window, forming force change items, voice prosody items, semantic expression items and interaction history association items. Generate a composite entry with a time slice identifier, modal integrity flag, and conflict status flag; Write the summary of the compound entry into the current session's acceptor table.

5. The AI ​​voice intelligent interaction system for decompression toys according to claim 4, characterized in that, Write the summary of the composite entry into the current session's acceptor table, including: Retain summaries of the most recent short window composite items, summaries of the most recent medium window trend items, and summaries of the most recent feedback-related items; Record corresponding markers when there are only force events, only voice events, missing historical states, or when semantic entries conflict with prosodic entries, for use in subsequent window construction and stage determination.

6. The AI ​​voice intelligent interaction system for decompression toys according to claim 1, characterized in that, The stage determination module includes: Based on the short window merging status and the medium window trend status, and combined with current session information and historical correlation information, stage judgment and risk judgment are made. It switches between states according to a preset state machine, including contact establishment, catharsis establishment, high arousal catharsis, fallback transition, stabilization and soothing, and recovery end. When a composite entry is missing, the modality is incomplete, or there is a conflict of information, a conservative decision is maintained, and the decision entry is sent to the constraint scheduling module.

7. The AI ​​voice intelligent interaction system for decompression toys according to claim 6, characterized in that, Based on the short window merging status and medium window trend status, combined with current session information and historical correlation information, stage judgment and risk assessment are performed, including: The current stage is determined based on the main receiving behavior, volume level, speech rate level, pause density, main semantic category, reassurance and rejection state, expression of urgent concern, and the direction of change of the previous state; The risk level is determined based on the duration of high-intensity stress, the density of cathartic semantics, the duration of silence, and the counterstimulation markers, with current session information taking precedence over historical preference information.

8. The AI ​​voice intelligent interaction system for decompression toys according to claim 1, characterized in that, The constraint scheduling module includes: The output constraint set is generated based on the stage judgment items and historical association information. An executable strategy is selected from the preset strategy family according to the content category boundary, tone level boundary, tactile mode boundary, action mode boundary, insertion timing condition and waiting time condition. When the switching gating conditions are met, the strategy is switched and the process is scheduled between the candidate output state, the execution output state, and the observation freeze state.

9. The AI ​​voice intelligent interaction system for decompression toys according to claim 1, characterized in that, The write-back update module includes: Based on the execution record entries and the collected entries after feedback, a feedback observation window is established to determine the effects of force changes, rhythm changes, semantic changes and succession relationships, and to generate in-session update structure and historical update structure. Adjust the acceptability of strategy families, tone levels, haptic patterns, and action patterns according to progressive rules; Write the updated results to the current session's acceptance table and the long-term preference table; When the anti-stimulation condition is met, a policy rollback signal is sent to the constraint scheduling module.

Images