Automatic marketing-oriented AI mobile phone voice interaction method

Abstract

The invention relates to the technical field of intelligent voice interaction, in particular to an AI mobile phone voice interaction method oriented to automatic marketing, which comprises the following steps of: providing a comprehensive basis for personalized marketing and improving the marketing pertinence by recognizing a user emotional state and a real-time user intention in parallel and mastering the user state; an emotion state transition network is dynamically constructed and updated, key indexes are extracted, and a user emotion change rule is analyzed, so that an interaction strategy is accurately adjusted; semantic fragments are recombined by combining emotional stability and emotional infection degree indexes, adaptive dynamic marketing verbal skill is generated, the attraction and persuasion of the verbal skill are enhanced, and the marketing effect is improved; the local NPU is utilized to synthesize emotional voices and superpose features, so that natural and vivid interaction experience is provided, and recognition of a user on marketing content is enhanced; through dynamic decision-making verbal skill strategy cross-scene switching, timely adjustment is carried out according to user changes, marketing strategy stiffness is avoided, and the marketing success rate is guaranteed.

Description

Technical Field

[0001] This invention relates to the field of intelligent voice interaction technology, and more particularly to an AI-powered mobile phone voice interaction method for automated marketing. Background Technology

[0002] With the deep integration of artificial intelligence technology into the communications field, automated marketing outbound calling systems based on AI smartphones have become an important means for enterprises to improve customer reach efficiency and reduce operating costs. Existing technical solutions generally adopt pre-recorded or templated voice scripts, combined with speech recognition and keyword matching technologies, to achieve basic information dissemination or simple question-and-answer interactions.

[0003] However, existing methods face bottlenecks in real-world, dynamic marketing dialogue scenarios. First, their emotion recognition and intent understanding are typically processed sequentially or in isolation, failing to capture the deep coupling and dynamic evolution of user emotional states and real-time intents, such as repeated inquiries (intent) arising from user doubt (emotion). Second, the system lacks the ability to model the inherent patterns of user emotional states, failing to quantify their stability and changing trends, resulting in lagging and mechanical adjustments to the script. Simultaneously, in existing technologies, speech synthesis and script strategies are relatively disconnected, lacking a collaborative enhancement mechanism driven by real-time user emotions for the emotional expression in voice broadcasts and the script content generated based on text logic. Therefore, this invention proposes an AI-powered mobile voice interaction method for automated marketing. Summary of the Invention

[0004] To overcome the aforementioned shortcomings of existing technologies and to achieve the above objectives, this invention provides an AI-powered mobile phone voice interaction method for automated marketing, comprising: S1. During the process of initiating an automated marketing call by an AI mobile phone, the user's voice stream is acquired, and the user's voice stream is processed in parallel using a pre-built emotion-intent joint recognition model, namely emotion state recognition and real-time user intent recognition, to generate a sequence of user emotion state vectors and a sequence of user intent recognition results. S2. Based on the user's emotional state vector sequence, dynamically construct and continuously update the emotional state transition network, and extract the emotional stability and emotional contagion indicators in the emotional state transition network in real time. S3. Combining the emotional stability and emotional contagiousness indicators, semantic fragments of the pre-set marketing script templates are recombined to generate emotionally appropriate dynamic marketing scripts. S4. Utilize the local NPU of the AI ​​phone to synthesize emotional TTS for dynamic marketing scripts, simultaneously overlay emotionally enhanced audio features that match the current emotional state of the user, and output the synthesized speech stream for voice interaction; among which, the emotionally enhanced audio features include background sound effects and speech rhythm. S5. During voice interaction, based on the real-time topology changes of the emotion state transition network and the sequence of user intent recognition results, dynamic decisions are made on whether to switch the dialogue strategy across scenarios.

[0005] Further, S1 includes: The user's speech stream is segmented and windowed to extract prosodic contour features and spectral tilt features, forming a speech aggregation feature vector; Arrange the aggregated speech feature vectors of the current frame and its preceding and following adjacent frames in chronological order to form a temporal feature segment; Temporal feature fragments are input into the shared low-level feature encoder of a pre-trained sentiment-intention joint recognition model; The emotion-intention joint recognition model includes a shared low-level feature encoder, an emotion recognition branch network, and an intention recognition branch network. The shared low-level feature encoder extracts common features, which are then processed by the parallel emotion recognition branch network and the intention recognition branch network. The high-order features output by the shared low-level feature encoder are simultaneously input into the emotion recognition branch network and the intention recognition branch network. The emotion recognition branch network outputs the probability distribution of the current center frame belonging to each preset emotion category; the intent recognition branch network outputs the intent classification label of the current center frame.

[0006] Further, S1 includes: A sliding window analysis is performed on the user's voice stream using an analysis window that contains a preset number of consecutive voice frames; Based on the probability distribution of sentiment categories output by the sentiment recognition branch network for all frames within the analysis window, the dominant sentiment category of the analysis window is determined by maximum a posteriori probability estimation. The one-hot encoded vector of the dominant sentiment category is used as the sentiment state vector of the analysis window; the sentiment state vectors of all analysis windows are arranged in chronological order to generate a sequence of user sentiment state vectors. For each analysis window, the dominant intent label for that analysis window is determined by aggregating the intent classification labels of all frames within the analysis window output by the intent recognition branch network. The dominant intent labels in the continuous analysis window are smoothed and corrected to form a sequence of user intent recognition results. Understandably, after generating the initial sequence of dominant intent labels, the label shifts in adjacent analysis windows are checked, and the current dominant intent labels are corrected or maintained based on the statistical probability of label appearance, ultimately outputting a smooth and coherent sequence of user intent recognition results.

[0007] Further, S2 includes: Define the set of nodes in the emotional state transition network as all preset emotional state categories; initialize the weights of all directed edges between nodes in the emotional state transition network to zero; Read the user's emotional state vector sequence sequentially. For two adjacent emotional state vectors in the user's emotional state vector sequence, use the emotional states they represent as the source node and target node in the emotional state transition network, respectively. In the emotional state transition network, a directed edge from the source node to the target node is searched. If it exists, the weight value of the directed edge is increased by a base weight. If it does not exist, the directed edge is created and its weight value is initialized to the base weight. At the same time, the weights of all existing directed edges in the emotional state transition network are multiplied by an exponential decay coefficient less than 1 according to their creation time.

[0008] Further, S2 includes: Based on the updated emotional state transition network weights, a modularity-based community detection algorithm is used to divide the emotional state transition network nodes, and closely connected emotional state transition network nodes are grouped into the same community; the variance of the edge weights between all emotional state transition network nodes in the same community is calculated, and the reciprocal of the variance is normalized to the [0,1] interval, which is used as the emotional stability index at the current moment. We calculate the weights of all directed edges in the emotional state transition network and the conditional probability distribution of each source node transitioning to all its target nodes. Based on the conditional probability distribution, we quantify the state transition entropy of the emotional state transition network using the Shannon entropy formula and linearly map the negative of the state transition entropy value to the [0,1] interval as an indicator of the emotional contagion at the current moment.

[0009] Further, S3 includes: The pre-set marketing script templates are parsed into multiple independent semantic fragments, and each semantic fragment is labeled with its preset emotional tendency tag and its position in the script logic; Construct an attention matching network: use the sentiment state vector of the latest analysis window in the user sentiment state vector sequence as the query vector, and the sentiment tendency label vectors of all semantic segments as the key vectors. Calculate the attention weight between the query vector and each key vector; use this attention weight as the sentiment association strength value of the corresponding semantic segment. The selection strategy for semantic segments is adjusted based on the currently extracted sentiment stability index: a dynamic selection threshold that is negatively correlated with the sentiment stability index is set; semantic segments with sentiment correlation strength values ​​greater than this dynamic selection threshold are included in the candidate semantic segment set; when sentiment stability is high, the dynamic selection threshold is increased accordingly, resulting in a small candidate semantic segment set with high correlation; when sentiment stability is low, the dynamic selection threshold is decreased accordingly, resulting in a larger candidate semantic segment set and allowing semantic segments with moderate correlation to be selected. The decision to insert bridging content is based on the extracted emotional contagion index: if the emotional contagion index is higher than the preset bridging threshold, then a preset bridging statement with emotional resonance content is inserted between the selected semantic segments; the bridging threshold is a decision threshold value used to control whether to insert emotional bridging statements. Its function is to prevent the abuse of empathetic content when users have low willingness to engage in emotional interaction, and to ensure that bridging statements are only triggered when the emotional interaction window is open, that is, when the contagion is high and the regularity is strong, thereby improving the naturalness and efficiency of the interaction; Based on the logical order defined by the positional identifiers of each semantic fragment, the selected semantic fragments and the inserted bridging statements are arranged and spliced ​​to generate grammatically coherent text, which serves as a dynamic marketing script with emotional adaptation.

[0010] Further, S4 includes: Based on the latest emotional state vector in the user's emotional state vector sequence and the currently extracted emotional contagion index, a composite emotional description vector is constructed. In the pre-constructed sound effect feature space, a K-nearest neighbor search is performed, that is, to find the K sound effect feature prototypes with the smallest Euclidean distance to the composite emotional description vector. Here, the sound effect feature prototype is a high-dimensional feature vector that uniquely associates with the original sound effect file after feature extraction, splicing and normalization. The sound effect identifiers corresponding to the K sound effect feature prototypes are used as the basic background sound effect identifier set for this overlay. Calculate the gain coefficient and spatial reverberation parameter of each sound effect in the basic background sound effect identifier set based on the currently extracted emotional stability index; In the local NPU, the backbone speech stream generated by emotional TTS synthesis is mixed with the background sound effect stream adjusted according to the gain coefficient and reverberation parameters in real time through multiple channels to generate a composite audio stream.

[0011] Further, S4 includes: Based on the emotional state vector of the latest analysis window in the user's emotional state vector sequence, and the currently extracted emotional contagiousness index, the target fundamental frequency curve, target syllable duration pattern, and target energy envelope parameters are determined in the predefined prosodic parameter space. During the local NPU's TTS synthesis process, the prosodic parameters inherent in the text-to-speech engine are adjusted frame by frame using a parameter interpolation algorithm to optimize the target fundamental frequency curve, target syllable duration pattern, and target energy envelope parameters.

[0012] Further, S5 includes: During the interaction, the modularity index of the emotional state transition network is periodically calculated, the sequence of changes in the modularity index over time is tracked and recorded, and the modularity change rate of the current period is calculated. Simultaneously, the user intent recognition result sequence is continuously analyzed to identify the change pattern of the dominant intent label in the user intent recognition result sequence, and the number of times the dominant intent label changes within the preset observation window is counted. Based on the rate of change of modularity and the number of times the dominant intent label changes, it is determined whether the network structure has entered an unstable state and whether the user intent has entered a chaotic state. If both network structure instability and user intent confusion are determined simultaneously, a cross-scenario switching decision for the sales script strategy is triggered, stopping the currently used marketing script template and selecting a new template from the pre-set backup sales script strategy library for loading.

[0013] Furthermore, the step of selecting and loading a new template from a pre-set library of backup script strategies includes: The sequence of user intent recognition results within the most recent predetermined time period before triggering the switching decision is converted into an intent state transition graph. In the intent state transition graph, the intent nodes are different intent labels that appear, and the directed edges represent the transition relationship between intents. The edge weight is determined by the number of transitions. For each alternative dialogue template in the alternative dialogue strategy library, obtain its preset typical intent transfer pattern diagram; Calculate the graph structure similarity between the intent state transition relationship graph transformed from the actual user intent recognition result sequence and the typical intent transition pattern graph of each backup dialogue template; The backup script template with the highest graph structure similarity is selected and loaded as the new current template.

[0014] Compared with existing technologies, the beneficial effects of the AI ​​mobile phone voice interaction method for automated marketing provided by this invention are as follows: 1) By processing user voice streams in parallel and simultaneously performing emotion state recognition and real-time user intent recognition, generating user emotion state vector sequences and user intent recognition result sequences, we can comprehensively grasp the user's emotional state and intent during the call, providing a solid foundation for subsequent personalized interaction, making marketing more targeted and improving user acceptance; by dynamically constructing and updating the emotion state transition network based on the user emotion state vector sequence, and extracting emotion stability and emotion contagion indicators in real time, we can deeply analyze the changing patterns and propagation characteristics of user emotions, so as to understand the dynamics of user emotions more meticulously, providing key basis for subsequent script adjustments, and making interaction strategies more in line with changes in user emotions; 2) By combining emotional stability and emotional contagiousness indicators to semantically reconstruct marketing script templates, dynamic marketing scripts with emotional adaptation are generated. The script content can be flexibly adjusted according to the user's current emotional state, enhancing its appeal and persuasiveness, improving marketing effectiveness, and avoiding user aversion due to inappropriate scripts. By utilizing a local NPU for emotional TTS synthesis, overlaying matched emotionally enhanced audio features, and outputting a speech stream, a more natural and vivid voice interaction experience can be provided, making the synthesized speech more emotionally expressive, bridging the gap with users, and increasing user identification with the marketing content. Based on the emotional state transition network and user intent recognition result sequence, dynamic decisions are made regarding whether to switch script strategies across scenarios. Marketing strategies are adjusted promptly according to changes in user emotions and intent, avoiding a decline in marketing effectiveness due to rigid strategies and ensuring a high marketing success rate throughout the entire interaction process. Attached Figure Description

[0015] Figure 1 This is a flowchart of the AI ​​mobile phone voice interaction method for automated marketing proposed in this invention. Detailed Implementation

[0016] 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.

[0017] Reference Figure 1 In this embodiment of the invention, the specific implementation steps of the AI ​​mobile phone voice interaction method for automated marketing include: S1. During the process of initiating an automated marketing call by an AI mobile phone, the user's voice stream is acquired, and the user's voice stream is processed in parallel using a pre-built emotion-intent joint recognition model, namely emotion state recognition and real-time user intent recognition, to generate a sequence of user emotion state vectors and a sequence of user intent recognition results. S2. Based on the user's emotional state vector sequence, dynamically construct and continuously update the emotional state transition network, and extract the emotional stability and emotional contagion indicators in the emotional state transition network in real time. S3. Combining emotional stability and emotional contagiousness indicators, the pre-set marketing script templates are semantically recombined to generate emotionally adapted dynamic marketing scripts. The marketing script template construction process includes collecting audio recordings of successful marketing conversations in the target industry; secondly, domain experts or natural language processing technology deconstruct the audio recordings into reusable basic semantic units, i.e., semantic fragments; based on the marketing function of the semantic fragments (e.g., opening remarks, product introductions, closing, handling rejection) and emotional tone (e.g., professionalism, enthusiasm, reassurance), they are labeled with emotional tendency tags and logical position identifiers; following common marketing logic flows, such as the AIDA model: attention, interest, desire, action, the semantic fragments are organized into multiple basic script templates. These multiple basic script templates constitute the original material library for AI mobile phone to recombine and dynamically generate scripts, i.e., marketing script templates. It should be noted that the audio recordings used in the construction of the marketing script templates are obtained in strict accordance with relevant laws, regulations, industry norms, and ethical guidelines. Specific compliance paths include, but are not limited to, internal accumulation and anonymized data, obtaining explicit user authorization, cooperation with compliant data providers, and the use of synthetic and simulated data. S4. Utilize the local NPU of the AI ​​phone to synthesize emotional TTS for dynamic marketing scripts, simultaneously overlay emotionally enhanced audio features that match the current emotional state of the user, and output the synthesized speech stream for voice interaction; among which, the emotionally enhanced audio features include background sound effects and speech rhythm. S5. During voice interaction, based on the real-time topology changes of the emotion state transition network and the sequence of user intent recognition results, dynamic decisions are made on whether to switch the dialogue strategy across scenarios.

[0018] In this embodiment of the invention, during the process of an AI-initiated automated marketing call, the user's voice stream is acquired, and a pre-built emotion-intent joint recognition model is used to process the user's voice stream in parallel, namely, emotion state recognition and real-time user intent recognition, to generate a user emotion state vector sequence and a user intent recognition result sequence. The detailed implementation steps include: S11. Perform framing and windowing processing on the user's speech stream, extract prosodic contour features and spectral tilt features, and form a speech aggregation feature vector, specifically including: The acquired continuous user speech stream is divided into frames with a fixed duration, such as 25 milliseconds, and there is an overlap between adjacent frames, such as 10 milliseconds. A Hamming window function is applied to each frame of speech signal to reduce spectral leakage. A Fast Fourier Transform is performed on each frame of speech signal after windowing to obtain its spectrum. The prosodic contour features are extracted by calculating the short-time energy and zero-crossing rate of the speech signal in the time domain, extracting the estimated value of its fundamental period or fundamental frequency, and combining all parameters reflecting rhythm and pitch changes into a prosodic contour sub-vector. Extracting spectral tilt features involves calculating the centroid of the spectrum of the speech signal in that frame and analyzing the attenuation slope of the spectrum in the high-frequency part. This slope is characterized by calculating the average energy ratio of the low-frequency band to the high-frequency band, forming a spectral tilt sub-vector. The prosodic contour subvector and the spectral tilt subvector are concatenated to form a speech aggregation feature vector that represents the speech of that frame.

[0019] S12. Arrange the speech aggregation feature vectors of the current frame and its preceding and following adjacent frames in chronological order to form a temporal feature segment. Specifically, this includes: setting a temporal context window centered on the current frame, wherein the temporal context window includes the current frame and its preceding U frames and following V frames, for example, U=V=5; arranging the current frame and its preceding and following adjacent frames in chronological order, and sequentially concatenating their corresponding speech aggregation feature vectors to form a comprehensive feature vector containing temporal information, which constitutes the temporal feature segment.

[0020] S13. Input the temporal feature fragments into the shared low-level feature encoder of the pre-trained sentiment-intention joint recognition model, specifically including: The temporal feature segment is used as the input data stream and fed into the first layer of the emotion-intention joint recognition model. The shared low-level feature encoder is composed of a multi-layer one-dimensional convolutional neural network or a long short-term memory network. Its function is to automatically learn and extract deep common abstract features from the input temporal feature segment. The deep common abstract features serve both the subsequent emotion recognition and intention recognition tasks. The high-order features output by the shared low-level feature encoder are new feature vectors with fixed dimensions, which are fed into the input of the emotion recognition branch network and the intention recognition branch network in parallel. The emotion-intention joint recognition model includes a shared low-level feature encoder, an emotion recognition branch network, and an intention recognition branch network. The shared low-level feature encoder extracts common features, which are then processed by the parallel emotion recognition branch network and the intention recognition branch network. The high-order features output by the shared low-level feature encoder are simultaneously input into the emotion recognition branch network and the intention recognition branch network. The training process of the emotion-intention joint recognition model includes: S131. Obtain the training dataset, wherein each piece of training data includes: a speech, the emotion category label corresponding to each frame of the speech, and the intent label corresponding to each frame of the speech. S132. Construct an emotion-intention joint recognition model, which includes a shared low-level feature encoder, an emotion recognition branch network, and an intention recognition branch network. The shared low-level feature encoder is used to extract high-order features from the input temporal feature segments. The emotion recognition branch network takes the high-order features as input and outputs the emotion category probability distribution. The intention recognition branch network takes the high-order features as input and outputs the intention classification label. S133. During the training iteration, the speech in the training data is converted into temporal feature segments and then input into the emotion-intention joint recognition model. The first loss value between the emotion category probability distribution output by the emotion-intention joint recognition model and the real emotion label is calculated, and the second loss value between the intention classification label output by the model and the real intention label is calculated. S134. The first loss value and the second loss value are weighted and summed according to preset weights to obtain the joint loss value. The parameters of the shared underlying feature encoder, the emotion recognition branch network and the intent recognition branch network are updated through the backpropagation algorithm until the model converges. During training, a gradient blocking layer is introduced between the shared underlying feature encoder and the two branch networks to limit the gradient flow range of backpropagation.

[0021] S14. The emotion recognition branch network outputs the probability distribution of the current center frame belonging to each preset emotion category; the intent recognition branch network outputs the intent classification label of the current center frame.

[0022] S15. Perform sliding window analysis on the user's speech stream using an analysis window containing a preset number of consecutive speech frames. Specifically, define a time analysis window whose length covers L consecutive speech frames, for example, L=50 frames, corresponding to approximately 1.25 seconds of speech. The analysis window slides on the time axis with a certain step size, for example, 25 frames, thereby dividing the entire speech stream into continuous or overlapping analysis windows.

[0023] S16. Based on the probability distribution of emotion categories output by the emotion recognition branch network for all frames within the analysis window, the dominant emotion category of the analysis window is determined by maximum a posteriori probability estimation, including: for each analysis window that is slid to, obtaining the probability value of each frame of speech within the analysis window that belongs to each preset emotion category output by the emotion recognition branch network, such as the emotion categories of happiness, calmness, confusion, etc.; for each emotion category, calculating the average probability of that category across all frames; and selecting the emotion category with the highest average probability as the dominant emotion category of the analysis window.

[0024] S17. Use the one-hot encoded vector of the dominant sentiment category as the sentiment state vector of the analysis window; arrange the sentiment state vectors of all analysis windows in chronological order to generate a sequence of user sentiment state vectors. Specifically, a vector with a length equal to the total number of preset emotion categories is constructed. The vector has a value of 1 only at the index position corresponding to the dominant emotion category, and 0 at the other positions. The one-hot encoded vector is the emotion state vector of the window. The emotion state vectors of all analysis windows are arranged in chronological order to generate a sequence of user emotion state vectors throughout the entire interaction process.

[0025] S18. For each analysis window, aggregate the intent classification labels output by the intent recognition branch network for all frames within the analysis window to determine the dominant intent label for that analysis window, specifically including: For each analysis window, the intent classification label determined by the intent recognition branch network for each frame of speech within the window is collected, such as inquiring about products, asking about prices, expressing refusal, requesting to be transferred to a human agent, etc.; the frequency of each category of intent classification label is counted; the intent category label with the highest frequency is determined as the dominant intent label for that analysis window; if the frequencies are the same, the label that appears later in time can be selected, or it can be determined according to the preset priority rules.

[0026] S19. Smooth and correct the dominant intent labels in the continuous analysis window to form a user intent recognition result sequence. Understandably, after generating the initial dominant intent label sequence, check the label transfer in adjacent analysis windows, and correct or maintain the current dominant intent label based on the statistical probability of the label's appearance, and finally output a smooth and coherent user intent recognition result sequence.

[0027] In this embodiment of the invention, the detailed implementation steps for dynamically constructing and continuously updating an emotional state transition network based on user emotional state vector sequences, and extracting emotional stability and emotional contagiousness indicators from the emotional state transition network in real time include: S21. Define the set of nodes in the emotional state transition network as all preset emotional state categories; initialize the weights of all directed edges between nodes in the emotional state transition network to zero.

[0028] S22. Read the user's emotional state vector sequence in sequence. For two adjacent emotional state vectors in the user's emotional state vector sequence, use the emotional states they represent as the source node and target node in the emotional state transition network, respectively.

[0029] S23. In the emotional state transition network, find the directed edge from the source node to the target node; if it exists, increase the weight value of the directed edge by a base weight; if it does not exist, create the directed edge and initialize its weight value to the base weight; at the same time, for all existing directed edge weights in the emotional state transition network, multiply them by an exponential decay coefficient less than 1 according to their creation time; where the base weight is a fixed increment value added to the weight of the corresponding directed edge each time a user's adjacent emotional state transition is detected, used to quantify the cumulative intensity of emotional transition, and combined with the exponential decay coefficient to realize the dynamic weakening of the influence of historical transitions; Specifically, a timestamp is maintained for each directed edge in the emotional state transition network to record the last time the weight of the directed edge was updated; each time the emotional state transition network is updated, that is, after processing a new pair of emotional state transitions, the time difference between the current time and the timestamp of the edge is calculated for all existing directed edges. Define a decay rate parameter and calculate the decay coefficient based on the time difference. This decay coefficient is the time difference of the decay rate parameter raised to a power of 1. Multiply the existing weight of each directed edge by the corresponding decay coefficient to decay the historical weights, so that the emotional state transition network can focus more on recent emotional transition patterns.

[0030] S24. Based on the updated emotional state transition network weights, the emotional state transition network nodes are divided using a modularity-based community detection algorithm, and closely connected emotional state transition network nodes are grouped into the same community; the variance of the edge weights between all emotional state transition network nodes in the same community is calculated, and the reciprocal of the variance is normalized to the [0,1] interval, which is used as the emotional stability index at the current moment. Specifically, modularity is an indicator of the strength of the community structure in an emotional state transition network. The community detection algorithm based on modularity aims to find a partitioning scheme that maximizes the modularity of the entire emotional state transition network by continuously trying to assign nodes to different communities. The Louvain algorithm optimizes modularity through iterative local node movements and community aggregation. Specifically, S241, local movement, includes: traversing each node in the emotional state transition network and calculating the modularity gain gained by moving it to the community of each neighboring node; moving the node to the community that brings the maximum positive gain; if all movements are ineffective... If the algorithm generates a positive gain, the node remains in the original community. S242, Community Aggregation includes: after completing the local movement of all nodes, aggregating all nodes belonging to the same community into a new supernode, merging the edge weights within the original community into the self-loop weights of the supernode, and merging the edge weights between communities into the edge weights between supernodes, thus forming a new sentiment state transition network; the above two steps are repeated until the modularity of the sentiment state transition network no longer increases, and the community partitioning obtained at this time is the algorithm result; the final community partitioning result is directly used to calculate the variance of the edge weights within the same community, thereby obtaining the sentiment stability index; To calculate the variance of edge weights among all nodes in the same emotional state transition network, specifically: for a given community, identify all node pairs within that community, including all directed edges, collect the weight values ​​of the directed edges in the current emotional state transition network, and calculate the statistical variance of the weight value sequence.

[0031] S25. Calculate the weights of all directed edges in the emotional state transition network and the conditional probability distribution of the transition from each source node to all its target nodes. Based on the conditional probability distribution, quantify the state transition entropy of the emotional state transition network using the Shannon entropy formula, and linearly map the negative of the state transition entropy value to the [0,1] interval as the emotional contagion index at the current moment. Specifically, calculating the conditional probability distribution of each source node's transition to all its target nodes includes: For each emotional state node in the emotional state transition network as the source node, the weights of all directed edges emanating from the source node are counted. The counted weights are then normalized by dividing the weight of each outgoing edge by the sum of the weights of all outgoing edges from the source node. The normalized value represents the conditional probability distribution of the emotional state transition from the source node to each possible target node in the next moment, given the current emotional state transition network state. Based on conditional probability distributions, the state transition entropy of the emotional state transition network is quantified using the Shannon entropy formula. This includes: for each source node, calculating the Shannon entropy of its calculated conditional probability distribution; the Shannon entropy is calculated by multiplying each probability value in the conditional probability distribution by its base-2 logarithm, summing all products, and taking the opposite of the sum; the entropy values ​​of all source nodes are averaged to obtain the state transition entropy of the entire emotional state transition network; a higher state transition entropy indicates more random and unpredictable emotional state transitions, while a lower state transition entropy indicates more regular and stable emotional state transitions.

[0032] The opposite of the state transition entropy value is linearly mapped to the interval [0,1]. The mapping method is as follows: set the theoretical minimum and maximum state transition entropy values, map the actual state transition entropy values ​​proportionally to the interval between 0 and 1, and take the opposite direction so that the index value obtained when the state transition entropy value is low (strong regularity) is high.

[0033] In this embodiment of the invention, the detailed implementation steps for generating emotionally adapted dynamic marketing scripts by combining emotional stability and emotional contagiousness indicators include: S31. Parse the pre-set marketing script template into multiple independent semantic fragments, and label each semantic fragment with its preset emotional tendency label and its position in the script logic.

[0034] S32. Construct an attention matching network: Use the sentiment state vector of the latest analysis window in the user sentiment state vector sequence as the query vector, and the sentiment tendency label vectors of all semantic segments as the key vectors. Calculate the attention weight between the query vector and each key vector; use this attention weight as the sentiment association strength value of the corresponding semantic segment. Specifically, the sentiment tendency label for each semantic segment is represented by a vector with the same dimension as the user's sentiment state vector. For example, for a segment expressing apology, its sentiment tendency label vector may have a higher value in the apology dimension. The latest user sentiment state vector is denoted as Q. The sentiment tendency label vectors of all H semantic segments are denoted as Di = D1, D2, ..., DH, where i is the semantic segment index used to traverse or identify the sentiment tendency label vectors of all semantic segments, H is the number of semantic segments, D1 is the sentiment tendency label vector of the first semantic segment, and D2 is the sentiment tendency label vector of the second semantic segment. The dot product of the latest user sentiment state vector Q and each sentiment tendency label vector Di is calculated, and all dot product results are input into the Softmax function for normalization. The output of the Softmax function is a probability distribution containing H values, where the i-th value represents the attention weight of the current user sentiment state to the i-th semantic segment. This attention weight is defined as the sentiment association strength value of the semantic segment.

[0035] S33. Adjust the semantic segment selection strategy based on the currently extracted sentiment stability index; Specifically, a basic selection threshold is pre-set, and a dynamic adjustment coefficient is set, whereby the dynamic adjustment coefficient is a parameter that adjusts the strictness of semantic segment screening. The dynamic selection threshold is generated by multiplying the basic selection threshold by the dynamic adjustment coefficient by the current sentiment stability index and then adding the result to obtain the final dynamic selection threshold used for screening. When the sentiment stability index increases, the dynamic selection threshold increases accordingly; when the sentiment stability index decreases, the dynamic selection threshold decreases accordingly. When screening semantic segments based on this dynamic selection threshold, only semantic segments with sentiment association strength values ​​greater than the dynamic selection threshold are included in the candidate semantic segment set. This achieves the following: when sentiment stability is high, the screening criteria are strict, the candidate semantic segment set is small, and the association is high; when sentiment stability is low, the screening criteria are lenient, and the candidate semantic segment set is large.

[0036] S34. Based on the currently extracted emotional contagion index, decide whether to insert bridging content: If the emotional contagion index is higher than the preset bridging threshold, insert a preset bridging statement with emotional resonance content between the selected semantic segments; where the bridging threshold is a decision threshold value used to control whether to insert emotional bridging statements. Its function is to prevent the abuse of empathetic content when the user's willingness to interact emotionally is low, and to ensure that the bridging statement is only triggered when the emotional interaction window is open, that is, when the contagion is high and the regularity is strong, thereby improving the naturalness and efficiency of the interaction; Specifically, a bridging threshold is preset; the currently extracted emotional contagion index is compared and analyzed with this bridging threshold; if the emotional contagion index is higher than the bridging threshold, it is determined that emotional resonance content needs to be inserted to enhance the continuity of interaction; from its preset bridging statement library, a bridging statement with the highest matching degree with the current user's emotional state vector and the current marketing intent is selected; the selection process is completed by calculating the cosine similarity between the feature vector of the bridging statement and the vector representation of the current composite emotional context, and the bridging statement with the highest similarity is selected; if the emotional contagion index is not higher than the bridging threshold, it is determined that no bridging statement needs to be inserted.

[0037] S35. Based on the logical order defined by the position identifiers of each semantic segment, arrange and splice the selected semantic segments and the inserted bridging statements to generate grammatically coherent text, which serves as a dynamic marketing script with emotional adaptation.

[0038] In this embodiment of the invention, the detailed implementation steps of using the local NPU of an AI mobile phone to synthesize emotional TTS for dynamic marketing scripts, simultaneously superimposing emotionally enhanced audio features that match the current emotional state of the user, and outputting the synthesized speech stream for voice interaction include: S41. Based on the latest analysis window's emotional state vector in the user's emotional state vector sequence and the currently extracted emotional contagion index, a composite emotional description vector is constructed. In the pre-constructed sound effect feature space, a K-nearest neighbor search is performed, that is, the K sound effect feature prototypes with the smallest Euclidean distance to the composite emotional description vector are found. Among them, the sound effect feature prototype is a high-dimensional feature vector that uniquely associates with the original sound effect file after feature extraction and splicing normalization. The sound effect identifiers corresponding to the K sound effect feature prototypes are used as the basic background sound effect identifier set for this overlay. Specifically, the process of constructing a composite emotion description vector is as follows: the latest emotion state vector and the emotion contagion index are concatenated; specifically, if the emotion state vector is 6-dimensional, the emotion contagion index is extended as a scalar to a 6th plus one dimension, thereby forming a 7-dimensional composite emotion description vector. The process of constructing the sound effect feature space includes: collecting a background sound effect material library covering various emotional tones; performing audio feature extraction operations on each sound effect file in the material library, the extracted features including the mean vector of Mel frequency cepstral coefficients, the spectral roll-off point, and the long-term average zero-crossing rate of the sound effect; concatenating and normalizing the extracted features to form a high-dimensional feature vector; this feature vector corresponding to each sound effect file is called the sound effect feature prototype; storing the sound effect feature prototypes of all sound effect files in a data space, this database constitutes the sound effect feature space; where each sound effect feature prototype is uniquely associated with the identifier of its original sound effect file.

[0039] S42. Calculate the gain coefficient and spatial reverberation parameters of each sound effect in the basic background sound effect identifier set based on the currently extracted emotional stability index, specifically including: A set of baseline audio parameters, including baseline gain and baseline reverberation time, is pre-stored for each sound effect identifier; When calculating the current gain coefficient, a gain adjustment factor is determined, which is determined by the product of the first preset coefficient and (1 - emotional stability index); the baseline gain is multiplied by (1 plus the gain adjustment factor) to obtain the current gain coefficient. The specific calculation formula is: Current gain coefficient = baseline gain × (1 + first preset coefficient × (1 - emotional stability index)). When calculating the current reverberation parameters, the reverberation adjustment factor is determined, which is determined by the product of the second preset coefficient and (1 - emotional stability index); the reference reverberation time is multiplied by (1 plus the reverberation adjustment factor) to obtain the current reverberation parameters. The specific calculation formula is: current reverberation parameters = reference reverberation time × (1 + second preset coefficient × (1 - emotional stability index)). Understandably, the first preset coefficient is a fixed parameter used to adjust the magnitude of the emotional stability index's adjustment to the sound effect gain coefficient, and the second preset coefficient is a fixed parameter used to adjust the magnitude of the emotional stability index's adjustment to the sound effect reverberation parameter. The first and second preset coefficients are pre-set fixed parameters. Based on historical dialogue data from marketing scenarios and corresponding user feedback, the impact of changes in emotional stability on user auditory perception is statistically analyzed, and coefficient values ​​that produce a natural listening experience are set. Specifically, historical dialogue data covering multiple marketing scenarios and corresponding user feedback are collected, and emotional stability indices and user satisfaction scores are extracted from each dialogue. By analyzing the correlation between changes in emotional stability and user satisfaction, the optimal coefficient values ​​that maximize the improvement of user satisfaction experience for sound effect gain and reverberation parameters are sought. Different candidate coefficient values ​​are tested by iterating through historical data. The expected user satisfaction for the synthesized speech under each coefficient combination is calculated. The coefficient combination that maximizes average satisfaction or minimizes negative feedback is selected as the preset coefficient value. Once the values ​​of the first and second preset coefficients are determined, they can be used as fixed parameters without recalculation in each interaction. For example, the values ​​of the first and second preset coefficients range from 0 to 1. When the preset coefficient is 0, changes in the emotional stability index have no effect on gain and reverberation, i.e., the emotional adjustment function is turned off. When the preset coefficient is 1, changes in the emotional stability index have the greatest effect on gain and reverberation. In practical applications, values ​​between 0.3 and 0.7 can be selected as needed to achieve a balance between the naturalness of emotional expression and the stability of audio effects. Since the gain coefficient mainly affects the volume and the reverb parameter mainly affects the spatial sense of the sound, and the two belong to different dimensions of audio, independent preset coefficients are used for adjustment to achieve multi-dimensional adaptation to the user's emotional state. Through the above calculation, the lower the emotional stability index, the greater the increase of the calculated gain coefficient and reverb parameter values ​​relative to their baseline values, thereby enhancing the emotional expressiveness of the audio when the user's emotions fluctuate.

[0040] S43. In the local NPU, the backbone speech stream generated by the emotional TTS synthesis is mixed with the background sound effect stream adjusted according to the gain coefficient and reverberation parameters in real time through multiple channels to generate a composite audio stream.

[0041] S44. Based on the sentiment state vector of the latest analysis window in the user's sentiment state vector sequence, and the currently extracted sentiment contagiousness index, determine the target fundamental frequency curve, target syllable duration pattern, and target energy envelope parameters in the predefined prosodic parameter space, specifically including: Maintain a lookup table for target prosodic parameters. The input index of this lookup table is the joint encoding of the emotion category and the emotion contagion level. Determine the dominant emotion category based on the latest emotion state vector. Discretize the continuous emotion contagion index into several contagion levels through a preset interval division rule. For example, 0-0.3 is divided into low contagion level, 0.3-0.7 into medium contagion level, and 0.7-1 into high contagion level. The current real value of the emotional contagiousness index is matched with a preset numerical range to determine the numerical range in which the emotional contagiousness index value is located, and the discrete contagiousness level corresponding to the range is output. Using the combination of dominant emotional category and contagiousness level as an index, the corresponding target prosodic parameter template is retrieved from the lookup table. This template specifies the target fundamental frequency curve, target syllable duration pattern and target energy envelope parameters. The target fundamental frequency curve is defined by providing a set of control points uniformly or non-uniformly distributed on the time axis. Each control point contains a time coordinate and the target fundamental frequency value. The target fundamental frequency values ​​between adjacent control points are determined by linear interpolation or spline interpolation, thus forming a complete curve of the target fundamental frequency changing over time. The target syllable duration pattern is determined by a provided mapping table, which defines the relative duration scaling factor corresponding to different speech unit types. The target syllable duration is obtained by multiplying the standard duration by the relative duration scaling factor. The target energy envelope parameter is obtained by preprocessing the target speech signal, transforming it from the time domain to the frequency domain through Fourier transform, and obtaining the spectral distribution of the signal. The energy of different frequency components in the spectrum is weighted and adjusted according to the emotional state. The weighted and adjusted spectrum is then transformed back to the time domain through inverse Fourier transform, thus obtaining the envelope contour of the target energy changing over time.

[0042] S45. During the TTS synthesis process performed by the local NPU, the prosodic parameters inherent in the text-to-speech engine are adjusted frame-by-frame using a parameter interpolation algorithm to optimize the target fundamental frequency curve, target syllable duration pattern, and target energy envelope parameters. Specifically, this includes: When a TTS engine synthesizes speech, it internally generates a set of default prosodic parameter sequences for the text sequence to be synthesized, including a fundamental frequency sequence, a syllable duration sequence, and an energy sequence. The adjustment process is performed on a per-speech-frame basis: for each frame, based on its position on the synthesized speech timeline, the target prosodic parameter value at that moment is calculated from the target prosodic parameter template through linear interpolation; a first-order recursive filtering algorithm is used to smoothly approximate the engine-generated current frame prosodic parameter value to the calculated target prosodic parameter value; specifically, the adjusted target prosodic parameter value is equal to the adjusted target prosodic parameter value of the previous frame plus a step size coefficient multiplied by the difference between the target prosodic parameter value and the adjusted target prosodic parameter value of the previous frame, where the step size coefficient controls the approximation speed to ensure a natural and smooth transition of prosodic parameters.

[0043] In this embodiment of the invention, during voice interaction, the detailed implementation steps for dynamically deciding whether to switch speech strategies across scenarios based on the real-time topology changes of the emotion state transition network and the sequence of user intent recognition results include: S51. During the interaction process, periodically calculate the modularity index of the emotional state transition network, track and record the sequence of changes in the modularity index over time, and calculate the modularity change rate of the current period. Specifically, set a fixed period duration or the number of analysis windows for period processing; at the end of each period, based on the current state of the emotional state transition network, call the community detection algorithm to calculate the modularity value of the emotional state transition network, and append this modularity value to the modularity time series; obtain the modularity change rate of the current period by calculating the difference between the current modularity value and the modularity value of the Nth period in the time series, and then dividing it by the total time span of N periods, where N is an integer greater than or equal to one; the modularity change rate is used to quantify the instantaneous speed of network structure evolution.

[0044] S52. Simultaneously, continuously analyze the user intent recognition result sequence, identify the change pattern of the dominant intent label in the user intent recognition result sequence, and count the number of times the dominant intent label changes within a preset observation window. Specifically, define a sliding observation window containing the most recent M analysis windows; update the content of the sliding observation window each time a new analysis window generates an intent result; traverse the M dominant intent labels within the sliding observation window, and compare whether the dominant intent labels of two adjacent sliding observation windows are consistent; each time a inconsistent dominant intent label is found, increment an internal counter by one; after the observation window is updated, the value of this counter is the number of intent changes within the current window.

[0045] S53. If the rate of change of modularity exceeds the preset rate of change threshold within several consecutive calculation cycles, the network structure is determined to be in an unstable state. The rate of change threshold is calculated based on the historical statistical value of the rate of change of modularity. The historical statistical value is the historical average and the historical standard deviation. The rate of change threshold is defined by adding a certain multiple of the historical average and the historical standard deviation.

[0046] S54. If the following two preset conditions are met simultaneously within a consecutive preset number of analysis windows, the user intent is determined to be in a state of confusion. Among them, the preset condition 1: the total number of changes to the dominant intent label obtained in the consecutive analysis window sequence exceeds the preset change number threshold; the preset condition 2: the proportion of all intent change events marked as unreasonable changes in the consecutive analysis window sequence exceeds the preset change proportion threshold. Specifically, the system sets the number of consecutive analysis windows, the threshold for the number of intent changes, and the threshold for the change ratio. The intent change frequency threshold defines the maximum number of times an intent can be switched normally within a given time window. The change ratio threshold defines the maximum percentage of unreasonable changes allowed among all change events. The system checks the specified number of consecutive analysis windows tracing back from the current moment and counts the total number of times the dominant intent label in the windows has changed. If this total number is less than or equal to the intent change frequency threshold, it is determined that the user has not entered a chaotic state. If the total number is greater than the intent change frequency threshold, subsequent analysis continues. The subsequent analysis is as follows: for each intent change event occurring in the sequence of consecutive analysis windows, a predefined regular task flow logic rule base is queried. This rule base defines reasonable intent transition pairs. If the intent pair consisting of the preceding and following intents of this change event exists in the regular task flow logic rule base, it is marked as a reasonable change; otherwise, it is marked as an unreasonable change. The system calculates the percentage of unreasonable change events out of the total number of change events. If this percentage is greater than the preset change ratio threshold, the user's intent is ultimately determined to be in a chaotic state.

[0047] S55. If both the unstable network structure and the confused user intent are determined at the same time, a cross-scenario switching decision for the sales script strategy is triggered, the currently used marketing script template is stopped, and a new template is selected from the pre-set backup sales script strategy library for loading. The process of selecting and loading a new template from a pre-set library of backup script strategies includes: S551. Convert the sequence of user intent recognition results within the most recent predetermined time period before triggering the switching decision into an intent state transition relationship graph; wherein, the intent nodes in the intent state transition relationship graph are different intent labels that appear, the directed edges represent the before and after transition relationship between intents, and the edge weight is determined by the number of transitions. S552. For each backup script template in the backup script strategy library, obtain its preset typical intent transfer pattern diagram. S553. Calculate the graph structure similarity between the intent state transition relationship graph transformed from the actual user intent recognition result sequence and the typical intent transition pattern graph of each backup dialogue template, specifically including: Determine a unified set of nodes, which is the union of all intent tags that have appeared in the intent state transition relationship graph and all preset intent tags in the typical intent transition pattern graph of the current backup script template to be compared; We construct weighted adjacency matrices for the intent state transition graph and the typical intent transition pattern graph, respectively. For the intent state transition graph, we construct a square matrix where the rows and columns correspond to each intent label in the unified node set. If there is a directed edge in the intent state transition graph pointing from one intent label to another, the element value of the corresponding row and column in the matrix is ​​the normalized transition weight of that directed edge; otherwise, the corresponding element value is zero, thus accurately reflecting the situation where there is no corresponding intent transition in the intent state transition graph. Finally, we obtain the weighted adjacency matrix of the intent state transition graph. For the typical intent transition pattern graph, we use a square matrix of the same size, and the values ​​of the matrix elements are filled according to the typical intent transition strength predefined in the template, thus obtaining the weighted adjacency matrix of the typical intent transition pattern graph. Calculate the cosine similarity between two weighted adjacency matrices; expand the two weighted adjacency matrices into one-dimensional vectors by rows, calculate the dot product of the two one-dimensional vectors, and then divide each dot product by the magnitude of the two one-dimensional vectors. The quotient is the cosine similarity value. Calculate the Earth's distance of movement between the intention state transition graph and the intention node distribution of the typical intention transition pattern graph: Calculate the importance weight of each intention label node in each graph; for a node in the intention state transition graph, its importance weight is the average of the sum of the weights of all incoming edges and the sum of the weights of all outgoing edges; for a node in the typical intention transition pattern graph, its importance weight is preset by the template; normalize the importance weights of all nodes in each graph so that their sum is one, thus obtaining two discrete probability distributions; calculate the minimum cost required to transform the node distribution of the intention state transition graph into the node distribution of the typical intention transition pattern graph by solving a linear programming problem; this cost is the Earth's distance of movement. Converting Earth's movement distance into a node distribution similarity metric: Divide the calculated Earth's movement distance by the maximum theoretical distance value and normalize it to obtain a normalized distance value; Subtract this normalized distance value from one to obtain a value between 0 and 1, which is the node distribution similarity metric. The larger this value is, the more similar the node distributions are. The final graph structure similarity is obtained by weighted fusion. A fusion weight coefficient is pre-set to balance the similarity of adjacency structure and the similarity of node distribution. The graph structure similarity is calculated by the following formula: multiply the cosine similarity value by the fusion weight coefficient, add the node distribution similarity measure multiplied by one and subtract the fusion weight coefficient, and the sum is the final graph structure similarity. S554. Select the backup script template with the highest diagram structure similarity as the new current template and load it.

[0048] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. 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.

[0049] It should be noted that all formulas in this manual are calculated by removing dimensions and taking their numerical values. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters and thresholds in the formulas are set by those skilled in the art according to the actual situation.

[0050] Although embodiments of the invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the claims and their equivalents.

Claims

1. An AI-powered mobile voice interaction method for automated marketing, characterized in that: The method includes: S1. During the process of initiating an automated marketing call by an AI mobile phone, the user's voice stream is acquired, and the user's voice stream is processed in parallel using a pre-built emotion-intent joint recognition model, namely emotion state recognition and real-time user intent recognition, to generate a sequence of user emotion state vectors and a sequence of user intent recognition results. S2. Based on the user's emotional state vector sequence, dynamically construct and continuously update the emotional state transition network, and extract the emotional stability and emotional contagion indicators in the emotional state transition network in real time. S3. Combining the emotional stability and emotional contagiousness indicators, semantic fragments of the pre-set marketing script templates are recombined to generate emotionally appropriate dynamic marketing scripts. S4. Utilize the local NPU of the AI ​​phone to synthesize emotional TTS for dynamic marketing scripts, simultaneously overlay emotionally enhanced audio features that match the current emotional state of the user, and output the synthesized speech stream for voice interaction; among which, the emotionally enhanced audio features include background sound effects and speech rhythm. S5. During voice interaction, based on the real-time topology changes of the emotion state transition network and the sequence of user intent recognition results, dynamic decisions are made on whether to switch the dialogue strategy across scenarios.

2. The AI ​​mobile phone voice interaction method for automated marketing according to claim 1, characterized in that, S1 includes: The user's speech stream is segmented and windowed to extract prosodic contour features and spectral tilt features, forming a speech aggregation feature vector; Arrange the aggregated speech feature vectors of the current frame and its preceding and following adjacent frames in chronological order to form a temporal feature segment; Temporal feature fragments are input into the shared low-level feature encoder of a pre-trained sentiment-intention joint recognition model; The emotion-intention joint recognition model includes a shared low-level feature encoder, an emotion recognition branch network, and an intention recognition branch network. The shared low-level feature encoder extracts common features, which are then processed by the parallel emotion recognition branch network and the intention recognition branch network. The high-order features output by the shared low-level feature encoder are simultaneously input into the emotion recognition branch network and the intention recognition branch network. The emotion recognition branch network outputs the probability distribution of the current center frame belonging to each preset emotion category; the intent recognition branch network outputs the intent classification label of the current center frame.

3. The AI ​​mobile phone voice interaction method for automated marketing according to claim 2, characterized in that, S1 includes: A sliding window analysis is performed on the user's voice stream using an analysis window that contains a preset number of consecutive voice frames; Based on the probability distribution of sentiment categories output by the sentiment recognition branch network for all frames within the analysis window, the dominant sentiment category of the analysis window is determined by maximum a posteriori probability estimation. The one-hot encoded vector of the dominant sentiment category is used as the sentiment state vector of the analysis window; the sentiment state vectors of all analysis windows are arranged in chronological order to generate a sequence of user sentiment state vectors. For each analysis window, the dominant intent label for that analysis window is determined by aggregating the intent classification labels of all frames within the analysis window output by the intent recognition branch network. The dominant intent labels in the continuous analysis window are smoothed and corrected to form a sequence of user intent recognition results.

4. The AI ​​mobile phone voice interaction method for automated marketing according to claim 3, characterized in that, S2 includes: Define the set of nodes in the emotional state transition network as all preset emotional state categories; initialize the weights of all directed edges between nodes in the emotional state transition network to zero; Read the user's emotional state vector sequence sequentially. For two adjacent emotional state vectors in the user's emotional state vector sequence, use the emotional states they represent as the source node and target node in the emotional state transition network, respectively. In the emotional state transition network, a directed edge from the source node to the target node is searched. If it exists, the weight value of the directed edge is increased by a base weight. If it does not exist, the directed edge is created and its weight value is initialized to the base weight. At the same time, the weights of all existing directed edges in the emotional state transition network are multiplied by an exponential decay coefficient less than 1 according to their creation time.

5. The AI ​​mobile phone voice interaction method for automated marketing according to claim 4, characterized in that, S2 includes: Based on the updated emotional state transition network weights, a modularity-based community detection algorithm is used to divide the emotional state transition network nodes, and closely connected emotional state transition network nodes are grouped into the same community; the variance of the edge weights between all emotional state transition network nodes in the same community is calculated, and the reciprocal of the variance is normalized to the [0,1] interval, which is used as the emotional stability index at the current moment. We calculate the weights of all directed edges in the emotional state transition network and the conditional probability distribution of each source node transitioning to all its target nodes. Based on the conditional probability distribution, we quantify the state transition entropy of the emotional state transition network using the Shannon entropy formula and linearly map the negative of the state transition entropy value to the [0,1] interval as an indicator of the emotional contagion at the current moment.

6. The AI ​​mobile phone voice interaction method for automated marketing according to claim 1, characterized in that, S3 includes: The pre-set marketing script templates are parsed into multiple independent semantic fragments, and each semantic fragment is labeled with its preset emotional tendency tag and its position in the script logic; Construct an attention matching network: use the sentiment state vector of the latest analysis window in the user sentiment state vector sequence as the query vector, and the sentiment tendency label vectors of all semantic segments as the key vectors. Calculate the attention weight between the query vector and each key vector; use this attention weight as the sentiment association strength value of the corresponding semantic segment. Adjust the semantic segment selection strategy based on the currently extracted sentiment stability index; The decision to insert bridging content is based on the currently extracted sentiment contagion index: if the sentiment contagion index is higher than the preset bridging threshold, then the preset bridging statement is inserted between the selected semantic segments. Based on the logical order defined by the positional identifiers of each semantic fragment, the selected semantic fragments and the inserted bridging statements are arranged and spliced ​​to generate grammatically coherent text, which serves as a dynamic marketing script with emotional adaptation.

7. The AI ​​mobile phone voice interaction method for automated marketing according to claim 1, characterized in that, S4 includes: Based on the latest emotional state vector in the user's emotional state vector sequence and the currently extracted emotional contagion index, a composite emotional description vector is constructed. In the pre-constructed sound effect feature space, a K-nearest neighbor search is performed, that is, to find the K sound effect feature prototypes with the smallest Euclidean distance to the composite emotional description vector. Here, the sound effect feature prototype is a high-dimensional feature vector that uniquely associates with the original sound effect file after feature extraction, splicing and normalization. The sound effect identifiers corresponding to the K sound effect feature prototypes are used as the basic background sound effect identifier set for this overlay. Calculate the gain coefficient and spatial reverberation parameter of each sound effect in the basic background sound effect identifier set based on the currently extracted emotional stability index; In the local NPU, the backbone speech stream generated by emotional TTS synthesis is mixed with the background sound effect stream adjusted according to the gain coefficient and reverberation parameters in real time through multiple channels to generate a composite audio stream.

8. The AI ​​mobile phone voice interaction method for automated marketing according to claim 7, characterized in that, S4 includes: Based on the emotional state vector of the latest analysis window in the user's emotional state vector sequence, and the currently extracted emotional contagiousness index, the target fundamental frequency curve, target syllable duration pattern, and target energy envelope parameters are determined in the predefined prosodic parameter space. During the local NPU's TTS synthesis process, the prosodic parameters inherent in the text-to-speech engine are adjusted frame by frame using a parameter interpolation algorithm to optimize the target fundamental frequency curve, target syllable duration pattern, and target energy envelope parameters.

9. The AI ​​mobile phone voice interaction method for automated marketing according to claim 1, characterized in that, S5 includes: During the interaction, the modularity index of the emotional state transition network is periodically calculated, the sequence of changes in the modularity index over time is tracked and recorded, and the modularity change rate of the current period is calculated. Simultaneously, the user intent recognition result sequence is continuously analyzed to identify the change pattern of the dominant intent label in the user intent recognition result sequence, and the number of times the dominant intent label changes within the preset observation window is counted. Based on the rate of change of modularity and the number of times the dominant intent label changes, it is determined whether the network structure has entered an unstable state and whether the user intent has entered a chaotic state. If both network structure instability and user intent confusion are determined simultaneously, a cross-scenario switching decision for the sales script strategy is triggered, stopping the currently used marketing script template and selecting a new template from the pre-set backup sales script strategy library for loading.

10. The AI ​​mobile phone voice interaction method for automated marketing according to claim 9, characterized in that, The step of selecting and loading a new template from a pre-set library of backup script strategies includes: The sequence of user intent recognition results within the most recent predetermined time period before triggering the switching decision is converted into an intent state transition diagram; For each alternative dialogue template in the alternative dialogue strategy library, obtain its preset typical intent transfer pattern diagram; Calculate the graph structure similarity between the intent state transition relationship graph transformed from the actual user intent recognition result sequence and the typical intent transition pattern graph of each backup dialogue template; The backup script template with the highest graph structure similarity is selected and loaded as the new current template.

Images