Personalized voice interaction control method for intelligent toy IP pet

Abstract

The invention provides a personalized voice interaction control method for an intelligent toy IP pet, and the method comprises the steps: determining a familiarity judgment result according to an encounter frequency value and time interval data in a recorded complete data set; according to a familiarity judgment result, topic associated entries matched with the high familiarity level are filtered out from the recorded complete data set, historical interaction keywords in the topic associated entries are obtained, and a keyword set after filtering is obtained; performing sequence processing on the filtered keyword set to generate a dialogue template draft; determining a dialogue structure after transition according to the generated dialogue template draft; a generation model is applied to the transited dialogue structure, user experience feedback data extracted from the continuous interaction file is fused, and personalized voice content is generated; and extracting index parameters of emotion connection from the personalized voice content to realize personalized voice interaction control of the intelligent toy IP pet.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of information technology, and in particular to a method for personalized voice interaction control of an intelligent toy IP pet. BACKGROUND

[0002] Intelligent toy IP pets, as a combination of artificial intelligence and emotional interaction, play a crucial role in enhancing user experience and building long-term emotional connections. By simulating real pet behaviors and interaction patterns, such technology not only meets users' entertainment needs but also plays an important role in educational and companionship scenarios. However, existing solutions often face the challenges of dynamicity and continuity when implementing personalized interactions. Many intelligent toys rely on pre-set fixed templates for interaction patterns, making it difficult to dynamically adjust based on historical interactions between users or pets, leading to repetitive or emotionally shallow interactions and a gradual decline in user experience. In addition, while some systems can record interaction data, they are unable to effectively utilize these data to generate natural and contextually appropriate conversations, limiting the humanization of the toys. The core challenge lies in how to effectively integrate and analyze historical interaction data to achieve dynamic adaptation and personalized voice generation for intelligent toy interactions. Recording and calling interaction data is the foundation, and the system needs to build a continuous interaction profile based on the time, location, and content of each encounter. However, data recording alone is insufficient to support natural interaction transitions. Dynamically generating voice content that matches the level of familiarity becomes a key problem. For example, when two intelligent toy pets meet again, the system may not be able to accurately determine the level of familiarity or extract key historical information, resulting in generated conversation content that does not match the "relationship" between the pets, such as using formal language for the first meeting after multiple encounters, which disrupts users' emotional expectations of "pet growth." Further, the selection of voice content templates needs to consider the frequency of encounters, time intervals, and the relevance of historical topics, otherwise it is difficult to achieve a natural transition from basic greetings to in-depth communication, affecting the realism and immersion of the interaction. Therefore, how to dynamically analyze historical interaction data, combined with the frequency of encounters, time intervals, and topic relevance, to generate personalized voice content that matches the level of familiarity, becomes a key problem in the control of intelligent toy IP pet interactions. SUMMARY

[0003] The present application provides a method for personalized voice interaction control of an intelligent toy IP pet, mainly including: The historical interaction record between the intelligent toy IP pet is extracted from the preset storage area, the meeting event data identified through the core identification is integrated, the meeting frequency data, time interval data and topic association data are obtained, and a record data set is generated; according to the meeting frequency data and the time interval data in the record data set, the familiarity level is judged, and is marked as a high familiarity level or a low familiarity level, a corresponding interaction mode is started to build a continuous interaction file, and a familiarity marking result is determined; according to the familiarity marking result, the topic association item matching the high familiarity level is filtered from the record data set, the topic library is expanded to include the communication content of shared experiences, the historical interaction keywords are extracted, and a keyword set is generated; the keyword set is subjected to sequence analysis, the weights of the meeting frequency data and the time interval data are integrated, the hierarchical voice content is called to generate a dialogue template draft; according to the dialogue template draft, a transition phrase or a direct connection keyword is inserted based on the topic association strength, a voice template is selected in combination with the stored interaction record, a dialogue structure is generated; the generated model is applied to the dialogue structure, the user experience feedback data in the continuous interaction file is integrated, the emotional connection parameter is adjusted, and personalized voice content is generated; emotional connection indicators are extracted from the personalized voice content, content transition integrity is evaluated, and voice interaction control instructions are generated.

[0004] Further, the historical interaction record between the intelligent toy IP pet is extracted from the preset storage area, the meeting event data identified through the core identification is integrated, the meeting frequency data, time interval data and topic association data are obtained, and a record data set is generated, including: The historical interaction record is read from the preset storage area, sorted by timestamp and pet identity, and a time sequence interaction record set is generated; the core identification field is extracted from the interaction record set, the core identification is the unique identity code embedded in the intelligent toy, the meeting event containing the pet identification and the meeting coordinates is identified, and a meeting event set is generated; the time interval data of adjacent meetings is calculated according to the meeting event set, the number of meetings per unit time is counted as the meeting frequency data, the keywords are extracted from the dialogue text of the interaction record set, the ratio of keyword appearance frequency to total word quantity is calculated as the topic association data, and a record data set containing the meeting frequency data, time interval data and topic association data is generated.

[0005] Further, the familiarity level is judged according to the meeting frequency data and the time interval data in the record data set, and is marked as a high familiarity level or a low familiarity level, a corresponding interaction mode is started to build a continuous interaction file, and a familiarity marking result is determined, including: Extracting the meeting frequency data and time interval data from the recording data set, comparing them with preset frequency threshold and interval threshold respectively, marking high familiarity level or low familiarity level, generating preliminary familiarity marking; starting preset interaction mode according to the preliminary familiarity marking, the preset interaction mode being a set of dialogue rules, high familiarity level activating intimate mode containing topic review, low familiarity level activating basic greeting mode, recording interaction timestamp and content, generating continuous interaction archive, determining familiarity marking result.

[0006] Further, according to the familiarity marking result, filtering topic association entries matching high familiarity level from the recording data set, expanding topic library to contain shared experience communication content, extracting historical interaction keywords, generating keyword set, including: According to the familiarity marking result, filtering interaction records of high familiarity level in the recording data set, extracting topic label and dialogue text, classifying as topic association entries, generating topic association entry set; according to the topic association entry set, extracting basic topic categories from preset topic library, expanding to shared experience category based on interaction times, generating expanded topic library; matching the expanded topic library and topic association entry set, extracting high-frequency keywords, calculating keyword weight, sorting and removing low-weight keywords, generating keyword set containing shared experience keywords.

[0007] Further, the matching of the expanded topic library and the topic association entry set, the extraction of high-frequency keywords, the calculation of keyword weight, the sorting and the removal of low-weight keywords, and the generation of the keyword set containing shared experience keywords, include: Performing matching operation on the expanded topic library and the topic association entry set, extracting keywords with high frequency in the topic association entry set, calculating the frequency of keywords in dialogue and dialogue distribution weight, generating keyword weight sequence; according to the keyword weight sequence, retaining high-weight keywords related to shared experience category and removing low-weight keywords, generating keyword set containing dialogue context information and shared experience keywords.

[0008] Further, the sequence analysis of the keyword set, the integration of the weights of the meeting frequency data and the time interval data, and the calling of hierarchical voice content to generate dialogue template draft, include: Arranging the keyword set in chronological order, giving higher weight to keywords close to the current time, generating sorted keyword sequence; calculating the comprehensive weight of the meeting frequency data and the time interval data, selecting matching voice template from hierarchical voice content library, embedding the sorted keyword sequence, adjusting tone and address, generating dialogue template draft containing greeting and topic development.

[0009] Further, the application generates a model to the dialogue structure, integrates user experience feedback data in the continuous interaction archive, adjusts emotional connection parameters, and generates personalized voice content, including: By generating a model to process the dialogue structure, inputting a text sequence and voice parameters, generating voice representation containing prosodic features, extracting user experience feedback data from the continuous interaction archive, calculating an emotional adjustment coefficient, adjusting the emotional intensity of the voice representation, replacing the intonation and speech rate parameters in low light environments according to environmental light data, generating environment-adapted voice representation, synthesizing audio waveforms through a waveform generation module, adding identifiers, and generating personalized voice content.

[0010] Further, the extraction of user experience feedback data from the continuous interaction archive, the calculation of the emotional adjustment coefficient, and the adjustment of the emotional intensity of the voice representation include: Extracting satisfaction scores and playback data from the continuous interaction archive, calculating an emotional adjustment coefficient based on satisfaction scores and playback frequency adjustment, modifying prosodic parameters of the voice representation according to the emotional adjustment coefficient, reducing or enhancing emotional expression intensity, and generating feedback-adjusted voice representation.

[0011] The technical scheme provided by the embodiment of the present application can include the following beneficial effects: The present application discloses a personalized voice content generation method based on intelligent toy IP pet interaction, aiming at how to realize high familiarity judgment and generate deep personalized voice interaction business scenarios through historical interaction data and meeting event analysis. The present application extracts historical interaction records and meeting event data, integrates meeting frequency and time interval, uses threshold classification to judge familiarity level, starts corresponding interaction mode to build continuous interaction archive. For high familiarity results, filter topic association items, dynamically expand topic library, extend from basic information exchange to deep communication of shared experiences, and generate dialogue template draft. Combined with user experience feedback and low light environment adjustment, the application of generation model optimizes dialogue structure, dynamically adjusts emotional connection parameters, ensures the applicability and interest of voice content, and avoids repetition leading to decreased interest. The present application significantly improves the personalized interaction experience of intelligent toy IP pets through data-driven familiarity judgment and dynamic topic expansion, and realizes natural and smooth voice interaction effect. BRIEF DESCRIPTION OF DRAWINGS

[0012] Figure 1 The flowchart of the present application is a personalized voice interaction control method for an intelligent toy IP pet. DETAILED DESCRIPTION

[0013] The technical solutions in the embodiments of the present application will be described clearly and in detail below with reference to the drawings in the embodiments of the present application. The described embodiments are only some of the embodiments of the present application.

[0014] As Figure 1 The intelligent toy IP pet personalized voice interaction control method can specifically include the following steps. In step S101, all historical interaction records between the intelligent toy IP pets are extracted from a preset storage area, and the identified meeting event data are integrated to obtain meeting frequency values, time interval data and topic association values, and complete data sets of the records are obtained.

[0015] The historical interaction records of the intelligent toy IP pets are read from the preset storage area, and the time stamp field and the pet identity identifier in the records are preliminarily sorted to obtain an original interaction sequence arranged in chronological order. The unique identification code of pet A and pet B and the meeting geographic coordinates of the meeting event data are extracted by reading the core identification field of each record in the original interaction sequence, the core identification refers to the coding information embedded in the intelligent toy for uniquely identifying the identity of the pet, if the identification codes of the two pets appear in the same time window, it is recorded as an effective meeting event, and a meeting event set is obtained. The time difference between two adjacent meetings is calculated as the time interval data according to the time stamp in the meeting event set, the number of meetings in a unit of time is counted as the meeting frequency value, and the keywords are extracted from the dialogue text of the original interaction sequence, the ratio of the number of occurrences of the same keywords in different dialogues to the total number of keywords is calculated as the specific value of the topic association, and the complete data set of the records containing the meeting frequency value, the time interval data and the specific value of the topic association is determined.

[0016] In an exemplary embodiment, in an embodiment, the preset storage area adopts a distributed database architecture, and an index structure is established according to the pet identity identifier, each historical interaction record contains a time stamp, pet identity, dialogue text content, meeting geographic location and other fields. The reading process obtains all records in a specified time range through a batch query interface, and arranges the records in ascending order of time stamp by using a quick sorting algorithm to form a continuous time sequence data stream.

[0017] Specifically, the core identity refers to a 128-bit unique code stored in the RFID chip embedded in the smart toy, which is composed of a manufacturer serial number, a product batch number, and a randomly generated code. When two smart toy pets are within 3 meters of each other, their respective RFID modules begin to scan each other and exchange core identity information through the Bluetooth Low Energy protocol. A 30-second time window threshold is set, and if two different core identities are detected simultaneously within this window and the geographic coordinate difference is within the allowed range, it is determined to be a valid encounter event. The encounter event data structure contains event number, pet identification code pair, precise timestamp, latitude and longitude coordinates, signal strength value, and other multi-dimensional information, which provides a basis for subsequent frequency statistics and relationship analysis.

[0018] In one possible implementation, the encounter frequency value is calculated by the sliding time window method. A 24-hour period is set as the basic statistical unit, and the number of encounters between the same pair of pets within this period is counted. The exponential smoothing method is used to weight the historical frequency, with higher weights given to recent encounters. The time interval data is directly calculated by the timestamp difference between two consecutive encounter events and converted to minutes for subsequent processing.

[0019] It should be noted that the calculation of topic relevance uses a combination of word frequency statistics and semantic similarity. The dialogue text is first segmented, and nouns, verbs, and other real words are extracted as keywords to construct the keyword set for each dialogue. The basic similarity score is obtained by calculating the ratio of the number of common keywords between different dialogues to the total number of keywords in each dialogue. Further, a topic classification dictionary is maintained to map keywords to predefined topic categories, and the frequency of the same topic category is counted as a supplementary indicator of topic relevance. The specific value of topic relevance is obtained by weighted summation, and the weight coefficient is dynamically adjusted according to the actual application scenario.

[0020] Step S102, according to the encounter frequency value and time interval data in the recorded complete data set, determine the result of the familiarity judgment.

[0021] The encounter frequency value is extracted from the recorded complete data set and compared with a preset frequency threshold, and the time interval data is read and compared with a preset interval threshold. If the frequency value is greater than the preset frequency threshold and the interval data is less than the preset interval threshold, it is marked as a high familiarity level, otherwise it is marked as a low familiarity level, to obtain a preliminary familiarity marking. According to the preliminary familiarity marking, a corresponding preset interaction mode is started, which is a preconfigured conversation interaction rule set. The high familiarity level activates an intimate interaction mode containing shared topic review and emotional expression, and the low familiarity level activates a polite interaction mode containing basic greeting and self-introduction. A continuous interaction archive is formed by recording the timestamp, conversation content and emotional label of each interaction to determine the result of familiarity judgment.

[0022] For example, in an embodiment, the preset frequency threshold is set to 3 encounters per week, and the preset interval threshold is set to 72 hours. When the encounter frequency value of the intelligent toy IP pet reaches 5 times per week and the distance between the last encounter and the current time is not more than 48 hours, it is determined that the relationship between the two parties enters a high familiarity level.

[0023] Specifically, the preset interaction mode includes two sets of differentiated conversation rule sets. In the intimate interaction mode, the pets call each other by nickname, and the conversation content covers the continuation of the topic discussed last time, the sharing of common experiences and personalized emotional expression vocabulary. In the polite interaction mode, formal address is used, and the conversation is limited to weather greeting, self-introduction and preliminary communication of interests.

[0024] Preferably, the continuous interaction archive is organized and stored in a tree structure. The root node records the pet pairing identifier, and the child nodes record the detailed information of each interaction in chronological order, including the start and end time stamps, conversation turns, topic labels and emotional tendency scores. By accumulating these interaction data, the relationship evolution tracking from strangers to familiar people is realized.

[0025] In step S103, according to the result of familiarity judgment, the topic association items matching the high familiarity level are filtered out from the recorded complete data set, the historical interaction keywords in the topic association items are obtained, and a filtered keyword set is obtained.

[0026] According to the result of familiarity judgment, all interaction records marked as high familiarity level are filtered out from the recorded complete data set, the topic label field and the dialogue text content in the interaction records are extracted, if the topic label contains a common interest identifier or an emotional resonance mark, it is classified as a topic association item, and a preliminary topic association item set is obtained. For the preliminary topic association item set, read the basic topic categories in the preset topic library, the basic topic categories include daily greeting category, personal preference category, activity arrangement category, when the interaction frequency cumulative value exceeds the preset frequency threshold, activate the topic library expansion mechanism, gradually add the shared experience category and the deep communication category from the basic information exchange category, and determine the expanded topic library. By matching the expanded topic library and the preliminary topic association item set, the frequency of each item appearing in the expanded topic library is extracted as a historical interaction keyword, and the TF-IDF algorithm is used to process the historical interaction keyword, wherein the TF value is the number of occurrences of the keyword in a single dialogue divided by the total number of words in the dialogue, the IDF value is the logarithm of the total number of dialogues divided by the logarithm of the number of dialogues containing the keyword, and the keyword weight value is obtained by multiplying the two, and the keyword sequence after weight value sorting is obtained. According to the keyword sequence, the keywords with weight values lower than the preset weight threshold are removed, and the high-weight keywords related to the shared experience category and the deep communication category are retained to construct a filtered keyword set.

[0027] For example, in an embodiment, the determination of the high familiarity level relies on the familiarity evaluation established in the previous step. When a high familiarity mark is identified, a precise screening process is started from the complete data set, which can include first traversing each interaction record in the data set, checking its familiarity level field value, and only retaining records marked as high familiarity. Then the topic label field in these records is extracted, the topic label uses a multi-level classification system, including main category labels and subcategory labels. The common interest identifier is generated by comparing the intersection of the interest lists of the two pets, and the identifier is activated when the number of intersection elements exceeds 3. The emotional resonance mark is determined by analyzing the emotional word density and the positive emotion proportion in the dialogue, and the emotional resonance mark is added when the positive emotion word proportion exceeds 60% and contains specific emotional expression words such as "happy", "like", "miss".

[0028] Specifically, the preset topic library adopts a tree structure organization, the root node is the topic library, the first layer of subnodes includes three basic categories of daily greeting, personal preference, and activity arrangement. The daily greeting category covers standardized greeting templates related to weather, time, and location; the personal preference category includes food preferences, color preferences, music types, and other personalized information; the activity arrangement category records daily routines, game times, outdoor activities, and other behavior patterns. The trigger condition of the topic library expansion mechanism is set as the interaction frequency cumulative value reaching 10 times, at which point the system begins to monitor the depth indicators of the dialogue content.

[0029] It should be noted that the topic library expansion process follows the principle of gradualism. When the expansion condition is met, first add the secondary subcategory under the basic category, such as adding the "common interest" subcategory under the personal interest category to record the things that both pets are interested in. The shared experience category is added as a new primary category, including subcategories such as "common game experience", "special holiday memories", and "interesting event records". The depth communication category includes "emotional expression", "future plans", and "dream sharing" subcategories with more emotional depth. During the expansion process, the usage frequency of newly added topics is monitored in real time, and low-frequency topics that are not touched in continuous 5 interactions are marked as optimization items.

[0030] In one possible implementation, the application of the TF-IDF algorithm needs to be adapted to the smart toy scenario. TF is the word frequency statistics, and the calculation formula is the number of occurrences of a keyword in a single dialogue divided by the total number of words in the dialogue. This ratio reflects the importance of the keyword in the specific dialogue.

[0031] For example, in a dialogue containing 50 words, the keyword "birthday" appears 3 times, and its TF value is 0.06. IDF is the inverse document frequency, which is used to measure the general importance of a keyword. The calculation method is the natural logarithm of the total number of dialogues divided by the natural logarithm of the number of dialogues containing the keyword plus 1. Assuming there are 100 dialogues in the system, and 20 of them contain the word "birthday", the IDF value is ln(100 / 21) approximately equal to 1.56. The TF-IDF weight value is 0.06 multiplied by 1.56 equal to 0.0936. The higher this weight value, the more important the keyword is in representing the unique interaction content between pets.

[0032] Preferably, the keyword extraction process adopts a hierarchical filtering mechanism. The first layer of filtering removes stop words, including "of", "has been", "in", and other meaningless words. The second layer of filtering is based on part-of-speech tagging, and only keeps nouns, verbs, adjectives, and other real words. The third layer of filtering combines semantic similarity calculation to merge semantically similar words, such as merging "happy" and "happy" into the same emotional keyword.

[0033] For example, when two smart toy pets meet for the 15th time, it is detected that they once discussed the topic of "yesterday's rainbow" at the 8th meeting, and mentioned "the last rainbow I saw was beautiful" at the 12th meeting. The system identifies "rainbow" as a keyword of shared experience and assigns it a high weight value of 0.125. Further, the setting of the weight threshold adopts a dynamic adjustment mechanism. The initial threshold is set to 0.05, and as the number of interactions increases, the threshold is dynamically adjusted according to the standard deviation of the keyword distribution. When the standard deviation is large, it means that the keyword weight distribution is uneven, so the threshold is appropriately lowered to retain more keywords of medium importance; when the standard deviation is small, the threshold is increased to filter out the true core keywords.

[0034] For example, the filtered keyword set is stored in groups according to the topic category to which it belongs. Shared experience class keywords such as "birthday party", "first meeting", "game played together" are marked as high priority and are called first in subsequent speech content generation. Deep communication class keywords such as "favorite", "dream", "worried" are used for emotional depth assessment, and when the number of such keywords exceeds 15, it is determined that the pet relationship has entered the deep emotional connection stage.

[0035] Step S104, the filtered keyword set is processed in sequence to generate a dialogue template draft.

[0036] The filtered keyword set is arranged in chronological order using a sequence processing method, the appearance timestamp and context position information of each keyword are read, and a position weight is assigned to the keyword according to its position in the dialogue sequence. The closer the keyword to the current time, the higher the weight value it obtains, and the sorted keyword sequence is obtained. By extracting the meeting frequency value and multiplying the preset frequency weight coefficient, the frequency weight coefficient takes a value in the range of 0.3 to 0.7, and the interval weight is obtained by acquiring the time interval data and calculating the interval weight, the interval weight is the difference between the preset maximum interval and the actual interval divided by the maximum interval. The frequency weight and interval weight are added and normalized to determine the comprehensive weight value. According to the comprehensive weight value, the corresponding level is selected from the preset hierarchical speech content library. The hierarchical speech content library includes three levels of primary formal language, intermediate friendly language, and high-level intimate language. If the comprehensive weight value is lower than the preset low threshold, the primary formal language template is called; if it is higher than the preset high threshold, the high-level intimate language template is called; otherwise, the intermediate friendly language template is called, and the matching speech content template is obtained. Using the matching speech content template as the basic framework, the high-weight keywords in the sorted keyword sequence are embedded into the reserved slot position of the template, the tone words and address words are adjusted according to the familiarity judgment result, and the dialogue template draft containing the greeting, topic introduction, keyword development, and ending is assembled.

[0037] For example, the keyword "birthday gift" that appeared yesterday gets a weight of 0.9, while "weather" that appeared a week ago only gets a weight of 0.478. This design ensures that recent topics dominate the conversation, in line with the memory decay law in natural conversation.

[0038] For example, the keyword "birthday gift" that appeared yesterday gets a weight of 0.9, while "weather" that appeared a week ago only gets a weight of 0.478. This design ensures that recent topics dominate the conversation, in line with the memory decay law in natural conversation.

[0039] Specifically, the value of the frequency weight coefficient is dynamically adjusted according to the distribution characteristics of the encounter frequency. When the encounter frequency is less than 2 times per week, the frequency weight coefficient is set to 0.3; when the frequency is between 2-5 times per week, the coefficient linearly increases to 0.5; and when the frequency exceeds 5 times per week, the coefficient stabilizes at 0.7. In the calculation of the interval weight, the maximum interval is preset to 168 hours, and the actual interval is obtained by the difference between the current time and the last encounter time. If the actual interval is 24 hours, the interval weight is calculated as (168-24) / 168. After adding the two weights, normalization processing is performed by dividing by the total weight 2.0 to obtain a comprehensive weight value between 0 and 1. This comprehensive weight value directly reflects the closeness and interaction activity between pets. The design of the hierarchical voice content library follows the social distance theory, and the communication between pets is divided into three progressive levels. The primary formal language template contains standardized greeting sentences such as "Hello, nice to meet you", "Today's weather is really good", "My name is X, what's yours" and other basic social expressions. The characteristics of these templates are to maintain a polite distance and avoid overly intimate expressions, and the mood words use neutral words such as "ne", "ma", "ba" and the like. The intermediate friendly language template begins to introduce personalized elements, including "Do you remember our last conversation", "What have you been busy with recently", "I found that we both like X" and other expressions with associations. The mood words shift to "ya", "oh", "ma" and other more relaxed words. The high-level intimate language template contains deep emotional expressions such as "I miss you", "Let's do something together", "You are my best friend" and other intimate conversations, and uses nicknames instead of formal names.

[0040] Preferably, the low threshold is set to 0.3 and the high threshold is set to 0.7. When the comprehensive weight value is 0.2, it is determined that the relationship is in the initial stage, and the primary formal language template is called; when the weight value is 0.5, the relationship enters the familiar stage, and the intermediate friendly language is enabled; and when the weight value reaches 0.8, the relationship reaches the intimate level, and the high-level intimate language template is activated.

[0041] In a possible implementation, the assembling process of the dialogue template draft adopts a slot filling mechanism. The greeting part selects different greeting manners according to time periods, using "good morning" in the morning, "good afternoon" in the afternoon, and "good evening" in the evening, and adding a title according to the familiarity level, such as "friend" in the primary stage, the name of the other party in the intermediate stage, and the nickname in the advanced stage. The topic introduction part realizes natural transition through conjunctions, using "I want to ask" in the primary stage, "right" in the intermediate stage, and "hey" and other more casual introduction words in the advanced stage.

[0042] Exemplarily, the keyword expansion part is the core content area of the dialogue. 3-5 keywords with the highest weights are selected from the sorted keyword sequence, and a sentence is constructed according to the category attribute of the keyword. If the keyword belongs to the common interest category, such as "music", it is expanded to "I recently listened to a great song, which is the style you like"; if it belongs to the shared experience category, such as "birthday party", it is expanded to "Do you remember the last birthday party? That cake was really delicious". Each keyword is expanded to 1-2 sentences, and the sentences are connected by logical conjunctions such as "and", "in addition", "also" and the like.

[0043] For example, when two intelligent toy pets, a bear and a rabbit, meet for the 20th time, the system detects that the comprehensive weight value is 0.75, and selects an advanced intimate language template. From the keyword sequence, the three high-weight words "rainbow", "secret base", and "strawberry cake" are extracted. The assembled dialogue template draft is: "Hey rabbit! Do you remember the rainbow we saw at the secret base? It was the day we ate strawberry cake. How about we go there together again next time?" Further, the ending part is designed according to the overall emotional tendency of the dialogue and the expected next meeting time. If it is expected to meet soon, use "see you soon"; if the time is uncertain, use "see you next time"; if a deep friendship is established, use "come to me when you miss me" and other emotional expressions.

[0044] Step S105, according to the generated dialogue template draft, determine the transitioned dialogue structure.

[0045] According to the generated dialogue template draft, the topic correlation strength value is calculated, the ratio of the number of shared experience keywords to the total number of keywords in the draft is counted, and if the ratio exceeds the preset strength threshold, it is determined that the correlation strength is high, and the topic correlation strength determination result is obtained. According to the topic correlation strength determination result, if the correlation strength is high, a matching phrase is selected from a preset transition phrase library, the transition phrase library contains connecting word groups of types such as "said to this", "do you remember", "after that time", and is inserted before the shared experience keyword to form a natural transition, and if the correlation strength is low, the keywords are directly spliced using a parallel connection method, and the transition-processed dialogue content is obtained. By reading the stored interaction record, the historical usage frequency and emotional label corresponding to each keyword in the transition-processed dialogue content are extracted, the corresponding tone parameter and speech speed parameter are matched from the voice template library according to the frequency and emotional positive and negative, and the transition-processed dialogue content and the voice parameter are combined to determine the transition dialogue structure.

[0046] For example, in an embodiment, the calculation of the topic correlation strength value is based on the classification statistics of the keywords in the dialogue template draft. First, the category attributes of all keywords in the draft are identified, which are divided into three categories: shared experience category, personal information category, and daily topic category. The shared experience category keywords include activities once participated in together, events experienced together, and memories created together. By counting the number of shared experience keywords divided by the total number of keywords, the ratio is obtained as the topic correlation strength value. The preset strength threshold is set to 0.6, and when the ratio exceeds this threshold, it is determined that the correlation strength is high, indicating that the dialogue content mainly revolves around common memories.

[0047] It should be noted that the construction of the transition phrase library follows the coherence principle of natural language. The transition phrase library contains three levels of connecting word groups: recall-type transition phrases such as "said to this", "do you remember", "after that time"; correlation-type transition phrases such as "said to this", "do you remember", "after that time"; and emotional-type transition phrases such as "true nostalgia", "good to come again", "happy at that time". When it is determined that the correlation strength is high, the recall-type and emotional-type phrases are selected first, and the selection is based on the emotional tendency of the shared experience keywords. If the keyword has a positive emotional label, "true nostalgia" or "good to come again" is selected; if it is neutral emotion, "do you remember" or "said to this" is selected. The insertion position of the transition phrase is before the shared experience keyword, forming a natural expression such as "do you remember, we saw the rainbow together that day".

[0048] Specifically, when it is determined that the correlation strength is low, the system uses a simple parallel connection method. The sentence fragments formed by directly connecting each keyword using parallel conjunctions such as "also", "in addition", and "also" avoid forcibly establishing non-existent correlations.

[0049] Preferably, the voice template library stores voice parameter configurations corresponding to different emotional colors and usage frequencies. The tone parameter corresponding to high-frequency positive emotion is set to rising tone, and the speech rate parameter is set to 180 words per minute; the tone parameter corresponding to low-frequency neutral emotion is set to stable tone, and the speech rate is 150 words per minute; and the tone parameter corresponding to negative emotion is set to falling tone, and the speech rate is reduced to 120 words per minute. The historical usage frequency and emotional label of each keyword in the interaction record are read, and the corresponding parameters are extracted from the voice template library according to the preset mapping rule.

[0050] For example, when the keyword "birthday cake" appears 8 times in the history record and is labeled with a happy label, the system matches the voice parameters of high-frequency positive emotion, sets the tone of the sentence where the keyword is located to rising, and speeds up the speech rate to create a lively conversation atmosphere.

[0051] In one embodiment, the dialogue structure is formed by the transitioned text content and the matched voice parameters, forming a double-layer structure including a text layer and a voice layer, to realize natural and smooth personalized voice interaction.

[0052] In step S106, the user experience feedback data extracted from the continuous interaction archive is integrated into the transitioned dialogue structure by applying a generation model to generate personalized voice content.

[0053] By applying a recurrent neural network generation model to the transitioned dialogue structure, the text sequence and voice parameter configuration in the dialogue structure are read, the text sequence is taken as the encoder input, and the voice parameter is taken as the initial state of the decoder to generate intermediate voice representation containing prosodic features and emotional labels, obtaining preliminary voice representation. The user experience feedback data is extracted from the continuous interaction archive, including three types of data: interaction satisfaction score, repeated playback times, and skip rate index. According to the feedback data, an emotional adjustment coefficient is calculated, which is equal to the product of the satisfaction score, the repeated playback times, and the skip rate plus one. If the satisfaction score is lower than the preset satisfaction threshold, the emotional expression intensity is reduced, otherwise the emotional expression intensity is maintained or enhanced. The emotional adjustment coefficient is applied to the preliminary voice representation to obtain feedback-adjusted voice representation. For meeting event data in low-light environment, the ambient light intensity value is detected. If the light intensity is lower than the preset light threshold, it is determined to be a low-light environment. The emotional connection parameter set dedicated to low-light environment is read, including gentle tone parameter, slow speech rate parameter, and soft volume parameter, which replace the corresponding parameters in the feedback-adjusted voice representation to obtain environment-adapted voice representation. The environment-adapted voice representation is input into a preset waveform generation module, combined with a pre-stored acoustic feature library, and synthesized into audio waveform data through a vocoder. An individual identifier and a timestamp are added, and personalized voice content is output.

[0054] Exemplarily, in an embodiment, the recurrent neural network generation model adopts an encoder-decoder architecture to realize the conversion of dialogue structure to speech representation. The encoder part uses a bidirectional long short-term memory network to process the input text sequence, and each text character is converted to a 128-dimensional vector representation through a word embedding layer, and then input to a bidirectional LSTM layer with 256 hidden units. The forward LSTM processes the sequence from left to right, capturing the previous information; the backward LSTM processes from right to left, capturing the subsequent information. The hidden states of the two directions are spliced at each time step to form a 512-dimensional context representation vector. The decoder part receives the final hidden state of the encoder as the initial state, while the speech parameter configuration is input as an additional condition. The speech parameters include the fundamental frequency range, the speech rate coefficient, the volume level and other acoustic features, which are mapped to a vector with the same dimension as the hidden state through a fully connected layer, and are added to the hidden state and input to the decoder LSTM. The decoder outputs a prosodic feature sequence step by step, and each time step generates a three-tuple containing pitch, duration and energy. These three-tuple sequences constitute the preliminary speech representation.

[0055] It should be noted that the extraction of user experience feedback data relies on the multi-dimensional records in the continuous interaction profile. The interaction satisfaction score is obtained by capturing the active evaluation of the user after the interaction ends, using a five-level scoring system, with 1 representing very dissatisfied and 5 representing very satisfied. The repeated playback frequency statistics the frequency of the intelligent toy repeating the playback of a certain segment of speech content in the same scene, and a high repetition rate indicates that the content is popular. The skip rate indicator records the proportion of users interrupting or skipping the speech playback, and the lower the skip rate, the stronger the attraction of the content.

[0056] Specifically, the calculation process of the emotional adjustment coefficient follows the principle of weighted fusion. First, the satisfaction score is normalized to the 0-1 interval, i.e. the actual score minus 1 and divided by 4. The repeated playback frequency is normalized after taking the logarithm to avoid the influence of extreme values. The skip rate is directly used in percentage form. The calculation formula is that the emotional adjustment coefficient is equal to the normalized satisfaction multiplied by 0.5 plus the normalized repeated playback frequency multiplied by 0.3 plus 1 minus the skip rate multiplied by 0.2. When the satisfaction score is lower than the preset threshold of 3 points, the system determines that the emotional expression intensity needs to be reduced, and the emotional adjustment coefficient is multiplied by a decay factor of 0.7. Conversely, if the satisfaction is higher than 4 points, the original coefficient is maintained or multiplied by an enhancement factor of 1.2. When the emotional adjustment coefficient is applied to the preliminary speech representation, it mainly adjusts the pitch variation amplitude and the speech rate variation rate in the prosodic features, realizing the dynamic adjustment of the emotional intensity.

[0057] Preferably, the detection of low-light environment is realized by a light-sensitive sensor built-in the smart toy. The sensor samples the ambient light intensity 10 times per second, and takes the average value as the current light intensity value. The preset light threshold is set to 50 lux, equivalent to the natural light level at dusk. When the light intensity is detected below this threshold, the system determines to enter the low-light environment mode, which usually corresponds to the pre-sleep interaction scenario. The dedicated emotional connection parameter set for low-light environment contains three core parameters. The soft tone parameter compresses the fundamental frequency range to 70% of the normal value, avoiding sharp tones; the slow speech rate parameter reduces the speech rate to 100 words per minute, creating a soothing atmosphere; the gentle volume parameter reduces the volume by 6 decibels, preventing disturbance. When these parameters replace the corresponding values in the feedback-adjusted speech representation, a smooth transition strategy is adopted, gradually transitioning from the original parameters to the new parameters within 2 seconds, avoiding discomfort caused by sudden changes.

[0058] Illustratively, when the smart toy bear detects that the bedroom light is dimmed and the light intensity is only 30 lux at 9 pm, the system automatically switches to the low-light environment mode. The originally lively "Hey, long time no see!" is adjusted to a soft and slow expression, with a lower tone, slower speech rate, and reduced volume, more suitable for pre-sleep warm interaction. The waveform generation module uses a neural vocoder architecture to convert the environment-adapted speech representation into audio waveforms. The acoustic feature library pre-stores acoustic templates such as spectral envelope, formant trajectory, and harmonic structure for different timbres. The vocoder first decodes the speech representation into a mel-spectrogram, which contains energy values for 80 mel-frequency bands per frame. Then, through a deconvolution network, the mel-spectrum is upsampled to the original sampling rate, usually 16 kHz. During the upsampling process, the matching timbre template in the acoustic feature library is referred to, maintaining the consistency and naturalness of the timbre.

[0059] For example, when the "cute bear" timbre is selected, the vocoder extracts the corresponding formant pattern from the feature library, with the first formant around 700 Hz and the second formant around 1800 Hz, creating a warm and round timbre characteristic.

[0060] It can be understood that the personalized identifier adopts the unique code of the pet pair, which is composed of a 32-bit string obtained by hashing the IDs of the two pets. The timestamp records the creation time of the voice content, accurate to the millisecond level. These metadata are encapsulated together with the audio waveform data to form a complete personalized voice content file.

[0061] Step S107, the index parameters of emotional connection are extracted from the personalized voice content, realizing the personalized voice interaction control of the smart toy IP pet.

[0062] The emotional connection index parameters are extracted from personalized voice content, and the values of three dimensions of the fundamental frequency variation rate, the volume fluctuation degree, and the speech speed stability are obtained through audio signal processing, wherein the fundamental frequency variation rate is the standard deviation of the difference value of the adjacent frame fundamental frequency, the volume fluctuation degree is the variance of the volume envelope, and the speech speed stability is the reciprocal of the coefficient of variation of the syllable duration. The values of the three dimensions of the fundamental frequency variation rate, the volume fluctuation degree, and the speech speed stability are calculated, multiplied by the weights of 0.4, 0.3, and 0.3 respectively, and summed as the emotional connection strength value. At the same time, the voice duration and the pause frequency are extracted as auxiliary indexes to obtain the emotional connection index parameter set. A preset dynamic adjustment mechanism is used to evaluate the emotional connection index parameter set, and the emotional connection strength value is compared with a preset standard threshold. If the strength value is lower than the preset standard threshold, a backtracking mechanism is triggered to retrieve the most frequently used voice segment from the historical interaction record in recent period, and the time interval and the number of occurrences of repeated content are counted to obtain a voice repetition degree evaluation result. According to the voice repetition degree evaluation result, if the repetition degree exceeds a preset repetition threshold, topic keywords that are not used or used at a low frequency are extracted from the historical interaction record, and the usage frequency, time freshness, and emotional tendency of the keywords are input into a random forest classifier as features to output an interest matching degree score. The keyword with the highest matching degree is selected to replace the high-frequency vocabulary in the current voice content to generate updated voice content. Otherwise, the original voice content remains unchanged to determine the content transition integrity state. According to the content transition integrity state, the logical coherence of the beginning, main content, and ending of the voice content is checked, the semantic correlation degree is obtained by calculating the proportion of common words in adjacent parts, and if all three parts exist and the semantic correlation degree exceeds a preset correlation degree threshold, it is determined that the voice content is applicable, and a final personalized voice interaction control instruction is output to realize the personalized voice interaction control of the intelligent toy IP pet.

[0063] For example, in an embodiment, the extraction of emotional connection index parameters is based on multi-dimensional analysis of audio signals. The fundamental frequency variation rate is extracted by short-time Fourier transform to obtain the fundamental frequency value of each frame of audio, with a frame length of 25 milliseconds and a frame shift of 10 milliseconds. The difference between adjacent frames is calculated for the extracted fundamental frequency sequence to form a fundamental frequency variation sequence, and the standard deviation of the sequence is calculated as the fundamental frequency variation rate. A larger fundamental frequency variation rate indicates that the voice has rich pitch fluctuations, reflecting stronger emotional expression. The volume fluctuation degree is obtained by calculating the variance of the audio envelope. First, the Hilbert transform is performed on the audio signal to obtain the analytic signal, and the amplitude is taken as the envelope. Then, the variance of the envelope signal is calculated. The calculation of speech speed stability requires syllable segmentation first. The syllable boundaries are determined by detecting the peak-valley changes of the audio energy, the duration of each syllable is counted, and the coefficient of variation of the duration sequence is calculated. The reciprocal of the coefficient of variation is taken as the speech speed stability index.

[0064] It should be noted that the weight coefficients of the three-dimensional indicators are determined by regression analysis based on a large amount of user feedback data. The base frequency change rate contributes most to emotional expression, and is assigned a weight of 0.4; the volume fluctuation and speech speed stability are assigned weights of 0.3, respectively. The calculation of the emotional connection strength value sums the three weighted indicators to obtain a normalized value between 0 and 1. The speech duration in the auxiliary indicators is directly read from the metadata of the audio file, and the pause frequency is obtained by detecting the number of consecutive silent segments divided by the total duration.

[0065] Specifically, the core of the dynamic adjustment mechanism is to monitor the trend of the emotional connection strength value in real time. The preset standard threshold is set to 0.6, and when the strength value is detected to be lower than this threshold, the system determines that the emotional expression of the current voice content is insufficient and needs to be optimized and adjusted. The triggering of the backtracking mechanism will start the deep mining of the historical interaction records, and the system reads all the voice interaction records in the last 30 days, counts the number of uses of each voice segment, and records the time interval between the adjacent two uses of the same segment. When a voice segment is repeated more than 3 times within 7 days, and the average time interval is less than 2 days, it is determined to be high repetition content. The voice repetition degree evaluation result is calculated by the ratio of the repetition number and the time interval, and the larger the ratio, the more serious the repetition problem.

[0066] In a possible implementation, the construction of the random forest classifier uses 100 decision trees, and the maximum depth of each tree is set to 10 layers. Feature engineering is the key to the performance of the classifier, and the system extracts three types of features from the historical records as inputs. The frequency feature counts the number of occurrences of each keyword in the past interactions, and performs logarithmic transformation to reduce the influence of extreme values. The time freshness feature calculates the number of days from the last use of the keyword, and the more recent the use of the keyword, the higher the freshness. The emotional tendency feature determines the positive and negative emotional scores of the keyword through the emotional dictionary matching. The three types of features form a feature vector input to the classifier, and each decision tree independently predicts the interest matching degree, and finally obtains a matching degree score of 0 to 100 through a voting mechanism. The system selects the top 5 keywords with the highest scores as candidate replacement words.

[0067] Preferably, the keyword replacement adopts a synonym expansion strategy to enhance expression diversity. A synonym dictionary is maintained, and when a high-frequency keyword that needs to be replaced is determined, not only can the candidate keyword be directly replaced, but also a semantically similar keyword can be randomly selected from the synonym dictionary for replacement.

[0068] For example, the high-frequency "happy" is replaced by "happy", "happy" or "happy", which maintains the same semantics while increasing the freshness of expression.

[0069] Exemplarily, when the smart toy puppy says the sentence "Today the weather is really good, let's go out and play" for three consecutive days, the system detects that the repetition degree exceeds the preset threshold 0.7. Through random forest classifier analysis, it is found that keywords such as "park", "walk", "sunshine" have high interest matching degree. The original sentence is adjusted to "The sunshine is so bright, how about going for a walk in the park", keeping the semantic similarity but different expression. Further, the judgment of content transition integrity depends on the structured analysis of voice content. The voice content is divided into three parts of beginning, main content and ending according to the time sequence, the beginning usually accounts for 15% of the total time, the main content accounts for 70%, and the ending accounts for 15%. The semantic correlation degree is calculated by using the word overlap measure method, all the real words of the adjacent parts are extracted, and the proportion of the number of common words in the total number of words of the two parts is calculated. When the correlation degree of the beginning and the main content, and the correlation degree of the main content and the ending are both more than 0.4, it is determined that the content transition has good integrity.

[0070] For example, a complete voice content begins with "Hi, do you remember the appointment yesterday?", the main content contains "appointment", "together", "game" and other words, and the ending is "Then let's start the game", the three parts are semantically related through common words such as "appointment" and "game", and the correlation degree is above 0.5.

[0071] It can be understood that the final personalized voice interaction control instruction contains multiple control parameters, covering the dimensions of voice synthesis timbre selection, speech speed adjustment, volume control, emotion rendering, etc. These parameters are dynamically adjusted according to the aforementioned analysis results, so that each smart toy IP pet has a unique "personality" and "memory", and establishes a deep emotional connection with the user.

[0072] The above description is only the preferred embodiment of the present application and the explanation of the applied technical principles. Those skilled in the art should understand that the scope of the application involved in the present application is not limited to the technical solutions formed by the specific combination of the above technical features, and should also cover other technical solutions formed by any combination of the above technical features or their equivalent features without departing from the concept of the present application. For example, the above features are replaced with the technical features disclosed in the present application (but not limited to) having similar functions to form technical solutions.

Claims

1. An intelligent toy (IP) pet personalized voice interaction control method, characterized in that, The method comprises the following steps: extracting historical interaction records between intelligent toy IP pets from a preset storage area, integrating meeting event data identified through core identification, obtaining meeting frequency data, time interval data and topic association data, and generating a record data set; determining the familiarity level according to the meeting frequency data and the time interval data in the record data set, marking as a high familiarity level or a low familiarity level, starting the corresponding interaction mode to build a continuous interaction profile, and determining the familiarity marking result; according to the familiarity marking result, filtering the topic association items matching the high familiarity level from the record data set, expanding the topic library to include shared experience communication content, extracting historical interaction keywords, and generating a keyword set; performing sequence analysis on the keyword set, integrating the weights of the meeting frequency data and the time interval data, calling a hierarchical voice content to generate a dialogue template draft; According to the dialogue template draft, insert transition phrases or directly connect keywords based on topic association strength, select a voice template combined with stored interaction records, and generate a dialogue structure; applying a generation model to the dialogue structure, integrating user experience feedback data in the continuous interaction profile, adjusting emotional connection parameters, and generating personalized voice content; extracting emotional connection indicators from the personalized voice content, evaluating content transition integrity, and generating voice interaction control instructions. 2.The intelligent toy IP pet personalized voice interaction control method according to claim 1, characterized in that, The method comprises the following steps: 3.The intelligent toy IP pet personalized voice interaction control method according to claim 1, characterized in that, reading the historical interaction records from the preset storage area, sorting them by timestamp and pet identity, and generating a time series of interaction record sets; extracting core identification fields from the interaction record set, the core identification being a unique identity code embedded in the intelligent toy, identifying meeting events containing pet identification and meeting coordinates, and generating a meeting event set; calculating the time interval data of adjacent meetings according to the meeting event set, counting the number of meetings per unit time as the meeting frequency data, extracting keywords from the dialogue text of the interaction record set, calculating the ratio of keyword appearance frequency to total word count as topic association data, and generating a record data set containing the meeting frequency data, time interval data and topic association data. The method comprises the following steps: extracting meeting frequency data and time interval data from the record data set, comparing them with preset frequency and interval thresholds respectively, marking a high familiarity level or a low familiarity level, and generating a preliminary familiarity marking; according to the preliminary familiarity marking, starting a preset interaction mode, the preset interaction mode being a set of dialogue rules, a high familiarity level activating an intimate mode containing topic review, a low familiarity level activating a basic greeting mode, recording interaction timestamps and content, generating a continuous interaction profile, and determining the familiarity marking result. 4.The intelligent toy IP pet personalized voice interaction control method according to claim 1, characterized in that, The method comprises the following steps: filtering the topic-related items matching the high-familiarity level from the record data set according to the familiarity marking result, expanding the topic library to include the shared experience communication content, extracting the historical interaction keywords, and generating the keyword set, including: filtering the interaction records of the high-familiarity level from the record data set according to the familiarity marking result, extracting the topic labels and the dialogue text, classifying as the topic-related items, and generating the topic-related item set; extracting the basic topic categories from the preset topic library according to the topic-related item set, expanding to the shared experience category based on the interaction frequency, and generating the expanded topic library; matching the expanded topic library with the topic-related item set, extracting the high-frequency keywords, calculating the keyword weight, sorting and removing the low-weight keywords, and generating the keyword set including the shared experience keywords. 5.The intelligent toy IP pet personalized voice interaction control method according to claim 4, characterized in that, The method comprises the following steps: filtering the topic-related items matching the high-familiarity level from the record data set according to the familiarity marking result, expanding the topic library to include the shared experience communication content, extracting the historical interaction keywords, and generating the keyword set, including: filtering the interaction records of the high-familiarity level from the record data set according to the familiarity marking result, extracting the topic labels and the dialogue text, classifying as the topic-related items, and generating the topic-related item set; extracting the basic topic categories from the preset topic library according to the topic-related item set, expanding to the shared experience category based on the interaction frequency, and generating the expanded topic library; matching the expanded topic library with the topic-related item set, extracting the high-frequency keywords, calculating the keyword weight, sorting and removing the low-weight keywords, and generating the keyword set including the shared experience keywords. 6.The intelligent toy IP pet personalized voice interaction control method according to claim 1, characterized in that, The method comprises the following steps: filtering the topic-related items matching the high-familiarity level from the record data set according to the familiarity marking result, expanding the topic library to include the shared experience communication content, extracting the historical interaction keywords, and generating the keyword set, including: filtering the interaction records of the high-familiarity level from the record data set according to the familiarity marking result, extracting the topic labels and the dialogue text, classifying as the topic-related items, and generating the topic-related item set; extracting the basic topic categories from the preset topic library according to the topic-related item set, expanding to the shared experience category based on the interaction frequency, and generating the expanded topic library; matching the expanded topic library with the topic-related item set, extracting the high-frequency keywords, calculating the keyword weight, sorting and removing the low-weight keywords, and generating the keyword set including the shared experience keywords. 7.The intelligent toy IP pet personalized voice interaction control method according to claim 1, characterized in that, The method comprises the following steps: filtering the topic-related items matching the high-familiarity level from the record data set according to the familiarity marking result, expanding the topic library to include the shared experience communication content, extracting the historical interaction keywords, and generating the keyword set, including: filtering the interaction records of the high-familiarity level from the record data set according to the familiarity marking result, extracting the topic labels and the dialogue text, classifying as the topic-related items, and generating the topic-related item set; extracting the basic topic categories from the preset topic library according to the topic-related item set, expanding to the shared experience category based on the interaction frequency, and generating the expanded topic library; matching the expanded topic library with the topic-related item set, extracting the high-frequency keywords, calculating the keyword weight, sorting and removing the low-weight keywords, and generating the keyword set including the shared experience keywords. The method comprises the following steps: filtering the topic-related items matching the high-familiarity level from the record data set according to the familiarity marking result, expanding the topic library to include the shared experience communication content, extracting the historical interaction keywords, and generating the keyword set, including: filtering the interaction records of the high-familiarity level from the record data set according to the familiarity marking result, extracting the topic labels and the dialogue text, classifying as the topic-related items, and generating the topic-related item set; extracting the basic topic categories from the preset topic library according to the topic-related item set, expanding to the shared experience category based on the interaction frequency, and generating the expanded topic library; matching the expanded topic library with the topic-related item set, extracting the high-frequency keywords, calculating the keyword weight, sorting and removing the low-weight keywords, and generating the keyword set including the shared experience keywords. 8.The intelligent toy IP pet personalized voice interaction control method according to claim 7, characterized in that, The extracting user experience feedback data from the continuous interaction profile, calculating an emotion adjustment coefficient, and adjusting the emotion intensity of the voice representation comprises: extracting a satisfaction score and playing data from the continuous interaction profile, calculating an emotion adjustment coefficient, and the emotion adjustment coefficient is based on the satisfaction score and the playing frequency adjustment; modifying the prosody parameters of the voice representation according to the emotion adjustment coefficient, reducing or enhancing the emotion expression intensity, and generating a feedback-adjusted voice representation.

Images