A human-computer interaction response system and method based on speech recognition

Abstract

The application relates to the field of intelligent voice interaction, and particularly discloses a man-machine interaction response system and method based on voice recognition, which comprises a voice collection module, a target speaker recognition module, a keyword detection module, a dialogue management module and an interaction response output module; the method establishes target speaker representation through mixed voice in a psychological consultation scene, carries out target user identity recognition, target voice extraction, online segmentation labeling and segment gating, sends target user active voice segments into a keyword detection link, generates a psychological consultation response strategy in combination with a voice transcription result, a speaker recognition result and a conversation context state, and then outputs feedback through voice and visual modes. The application can improve the response object recognition accuracy, anti-interference capability and interaction continuity in a multi-person scene.

Description

Technical Field

[0001] This invention relates to the field of intelligent voice interaction technology, and more specifically, to a human-computer interaction response system and method based on speech recognition. Background Technology

[0002] With the development of smart terminals, service robots, and voice interaction systems, speech recognition-based human-computer interaction response technology has been widely applied in question-and-answer services, companionship services, health assistance, and consultation assistance scenarios. In psychological consultation assistance scenarios, users typically prefer to express their complaints, emotional experiences, and problem background through natural speech. Therefore, the stability of the speech recognition link, the accuracy of the response, and the maintenance of dialogue continuity directly affect the interaction effect.

[0003] Most existing voice interaction systems adopt a basic chain of "voice acquisition + speech recognition + intent determination + voice response". For example, Zhao Yangyang's "A Review of Research on Task-Oriented Dialogue Systems" published in the *Chinese Journal of Computers* in 2020... Figure 2 As shown, this paper deeply analyzes the pipeline approach for task-oriented dialogue systems from the perspectives of theoretical models, research progress, usability, and existing problems and challenges. The overall approach is "speech recognition—semantic understanding—dialogue management—response output." In 2003, Xie Guodong published "A Method for Parsing Spoken Chinese Oriented to Intermediate Semantic Representation Formats" in the journal *Chinese Information Processing*, proposing a method for parsing spoken Chinese that combines statistics and rules, such as... Figure 3 As shown, the structure of a human-computer dialogue system—"speech recognition - language parsing - dialogue management - speech synthesis"—is clearly presented. In a quiet, single-person environment, the above solution can achieve relatively stable response functionality. However, in a consultation room environment, there are often multiple people speaking in the same setting, others interrupting, and overlapping speech. If the collected mixed speech is directly fed into speech recognition, it is easy to recognize the speech of non-target speakers, leading to deviations in consultation intent, incorrect responses, and confusion of the context. Especially in psychological counseling support scenarios, such errors are more likely to disrupt the continuity and privacy of the consultation experience. Summary of the Invention

[0004] To overcome the aforementioned deficiencies of the prior art, this invention provides a human-computer interaction response system and method based on speech recognition. This system extracts embedding vectors from the collected mixed speech, establishes a target speaker representation, jointly models the time-frequency features of the mixed speech and the target speaker representation, outputs the target speech reconstruction result, performs segmented analysis along the time axis on the target speech reconstruction result and outputs annotation results, sends the speech segment identified as belonging to the target user into the keyword detection link, and combines the speaker recognition result, keyword recognition result, and conversation state to generate a personalized psychological counseling response for the target user, thereby solving the problems mentioned in the background art.

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

[0006] A human-computer interaction response system based on speech recognition, comprising:

[0007] The voice acquisition module collects conversational voices in psychological counseling scenarios and outputs mixed voices.

[0008] The target speaker recognition module identifies the target user and determines the time sequence of mixed speech, filters out the speech segments belonging to the target user, and suppresses speech from non-target speakers and overlapping speech interference.

[0009] The keyword detection module performs voice keyword detection on active voice segments of the target user and outputs keyword recognition results that represent consultation intent, consultation topics and interactive control.

[0010] The dialogue management module identifies the current respondent and generates psychological counseling response strategies.

[0011] The interactive response output module outputs voice response information and provides human-computer interaction feedback based on psychological counseling response strategies.

[0012] As a further aspect of this invention, a voice acquisition module collects conversational speech in a psychological counseling scenario and outputs mixed speech, including the following specific details: The voice acquisition module can employ a single microphone, dual microphones, or a microphone array structure, and can also be integrated into a robot, mobile terminal, consultation kiosk, self-service machine, or desktop device. The acquired mixed speech stream is preferentially sent to the target speaker recognition module, while simultaneously forming a parallel input relationship with the keyword detection module.

[0013] As a further aspect of the present invention, the target speaker recognition module performs target user identification and temporal segmentation on mixed speech, filters out speech segments belonging to the target user, and suppresses non-target speaker speech and overlapping speech interference. Specifically, the target speaker recognition module includes a registered speech modeling unit, a target speaker extraction unit, an online speaker separation and annotation unit, and a segment gating unit. The registered speech modeling unit collects the target user's registered speech before the consultation begins or when the user uses the service for the first time, extracts the speaker embedding vector, and establishes a target speaker representation.

[0014] The target speaker extraction unit transforms the mixed speech into the time-frequency domain to obtain the time-frequency features of the mixed speech; at the same time, it calls the target speaker representation already generated by the registered speech modeling unit, and inputs the target speaker representation as a conditional vector representing the voice identity features of the target user into the speech extraction model, and outputs the target speech reconstruction result.

[0015] The online speaker separation and annotation unit performs segmented analysis on the target speech reconstruction results along the time axis and outputs annotation results, which include target user active tags, non-target speaker tags, and overlapping speech tags.

[0016] Based on the annotation results, the segment gating unit only sends the speech segments that are labeled as active tags of the target user to the keyword detection module, while caching, downgrading, or triggering repeat prompts for non-target speaker segments and overlapping speech segments.

[0017] As a further aspect of the present invention, the keyword detection module performs voice keyword detection on active voice segments of the target user and outputs keyword recognition results representing consultation intent, consultation topic, and interaction control. This includes the following specific steps: After receiving active voice segments of the target user released by the segment gating unit, the keyword detection module performs continuous voice analysis on the active voice segments of the target user, identifying keyword information that can represent the current consultation content and interaction intent from the continuous spoken language flow. Simultaneously, the keyword detection module transcribes the same gated target user voice to obtain a voice transcription result corresponding to the keyword recognition result, and provides the voice transcription result and the keyword recognition result together to the dialogue management module.

[0018] As a further aspect of the present invention, the dialogue management module determines the current response target and generates a psychological counseling response strategy, including the following specific content: After receiving the keyword information recognition results, speech transcription results, and target user speech segment recognition results output by the keyword detection module and the target speaker recognition module, the dialogue management module does not process each result in isolation. Instead, it combines the dialogue context state corresponding to the current stage of the conversation to make a comprehensive judgment on the current input content, so as to determine whether the current interaction was initiated by the target user, what consultation stage the current expression belongs to, and what response method the system should adopt.

[0019] As a further aspect of the present invention, the interactive response output module outputs voice response information and provides human-computer interaction feedback according to the psychological counseling response strategy, including the following specific content: After receiving the psychological counseling response strategy generated by the dialogue management module, the interactive response output module converts the response strategy into feedback content for the target user and selects the appropriate output method to present it according to the current conversation state.

[0020] A human-computer interaction response method based on speech recognition includes the following steps:

[0021] Step 1: Collect conversational audio from a psychological counseling scenario and output mixed audio.

[0022] Step 2: Perform target user identification and time segmentation on the mixed speech, filter out the speech segments belonging to the target user, and suppress the speech of non-target speakers and overlapping speech interference.

[0023] Step 3: Detect voice keywords in the active voice segments of the target user and output the keyword recognition results that represent the consultation intent, consultation topic and interaction control.

[0024] Step 4: Identify the current respondent and generate a psychological counseling response strategy.

[0025] Step 5: Based on the psychological counseling response strategy, output voice response information and provide human-computer interaction feedback.

[0026] The technical effects and advantages of this invention, a human-computer interaction response system and method based on speech recognition, are as follows: First, by establishing a target speaker representation, this invention performs target user identification, target speech extraction, online segmentation and annotation, and segment gating on mixed speech collected in a psychological counseling scenario. Only speech segments confirmed to be active by the target user are sent to the subsequent keyword detection link. This effectively suppresses interference from non-target speaker speech, temporary interruptions, and overlapping speech, reducing the risk of consultation intent deviation, incorrect response targets, and contextual state confusion, and improving the system's recognition accuracy and response targeting in scenarios with multiple users. Simultaneously, this invention integrates keyword recognition results and speech transcription results... The system integrates speaker recognition results with the conversation context state to dynamically generate psychological counseling response strategies that match the current consultation content across different stages, including access, chief complaint collection, follow-up questioning and clarification, guidance and feedback, and conclusion. This enhances dialogue continuity, conversation management capabilities, and personalized interaction levels. Furthermore, the system can output feedback through voice, text summaries, candidate buttons, reassurance interface prompts, breathing training animations, and virtual avatars or robot actions. Therefore, this invention offers technical effects and advantages such as more accurate identification of target response objects, stronger anti-interference capabilities, better interaction continuity, more comprehensive privacy protection, and a more natural and complete human-computer interaction experience. Attached Figure Description

[0027] Figure 1 This is a schematic diagram of the structure of a human-computer interaction response system based on speech recognition according to the present invention.

[0028] Figure 2 This is a framework diagram of pipeline methods for existing task-oriented dialogue systems.

[0029] Figure 3 This is a block diagram of the existing spoken language translation system and human-computer dialogue system.

[0030] Figure 4 This is a schematic diagram of the human-computer interaction response task creation interface on the platform side of the present invention.

[0031] Figure 5 This is a schematic diagram of the voice test and connection verification interface on the platform side of the present invention.

[0032] Figure 6 This is a flowchart of a human-computer interaction response method based on speech recognition according to the present invention. Detailed Implementation

[0033] 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.

[0034] Example 1

[0035] like Figure 1 As shown, a human-computer interaction response system based on speech recognition includes:

[0036] The voice acquisition module collects conversational voices in psychological counseling scenarios and outputs mixed voices.

[0037] The target speaker recognition module identifies the target user and determines the time sequence of mixed speech, filters out the speech segments belonging to the target user, and suppresses speech from non-target speakers and overlapping speech interference.

[0038] The keyword detection module performs voice keyword detection on active voice segments of the target user and outputs keyword recognition results that represent consultation intent, consultation topics and interactive control.

[0039] The dialogue management module identifies the current respondent and generates psychological counseling response strategies.

[0040] The interactive response output module outputs voice response information and provides human-computer interaction feedback based on psychological counseling response strategies.

[0041] Furthermore, the voice acquisition module acquires conversational voices in psychological counseling scenarios and outputs mixed voices. In practical applications, the voice environment in psychological counseling scenarios is not a completely ideal single-person silent environment. Instead, there may be situations where the target user continues to express themselves, others interrupt temporarily, and background noise coexists. Therefore, the voice acquisition module often acquires mixed voices after the superposition of multiple voice components.

[0042] The voice acquisition module can adopt a single microphone, dual microphone, or microphone array structure to adapt to the voice input needs of different device forms; at the same time, the voice acquisition module can also be integrated into robots, mobile terminals, information kiosks, self-service machines, or desktop devices to meet the human-computer interaction requirements of different psychological counseling auxiliary application scenarios.

[0043] After the voice acquisition module completes voice acquisition, it prioritizes sending the mixed voice to the target speaker recognition module for initial recognition processing based on whether the current speaker is the target user. Simultaneously, the mixed voice forms a parallel input relationship with the keyword detection module in the system architecture. However, this parallel relationship does not mean that the keyword detection module directly performs formal recognition on the original mixed voice. Rather, it means that the keyword detection module maintains a collaborative connection with the target speaker recognition module in the system architecture. Specifically, the formal recognition input of the keyword detection module is not the original mixed voice itself, but rather the active voice segment of the target user released by the target speaker recognition module after completing its corresponding recognition processing and undergoing segment gating.

[0044] Furthermore, the target speaker recognition module performs target user identification and temporal segmentation on the mixed speech, filters out speech segments belonging to the target user, and suppresses non-target speaker speech and overlapping speech interference. This includes a registration speech modeling unit, a target speaker extraction unit, an online speaker separation and annotation unit, and a segment gating unit. The registration speech modeling unit is activated before the consultation begins or when the user uses the system for the first time. It is used to collect the target user's registration speech and establish the target speaker representation required for subsequent recognition. Specifically, before the consultation begins or when the user uses the system for the first time, the system enters a registration guidance state, prompting the target user to input speech in rounds within a preset time window after each prompt tone. After collecting the speech segments from each round, invalid segments with overlapping speech, temporary interruptions, or speech interruptions are removed to obtain the target user's registration speech. After acquiring the registered speech of the target user, the registered speech modeling unit first performs speech activity detection and silence segment removal to eliminate invalid pauses and obvious non-speech segments. Then, the retained valid speech is processed into frames according to a preset frame length and frame shift, and acoustic features used to characterize speaker features are extracted. Subsequently, the acoustic features are input into the speaker embedding model, which performs feature encoding on the speech at the frame level and aggregates the temporal information in the variable-length speech into a fixed-length representation through temporal statistical pooling or attention statistical pooling, thereby obtaining a speaker embedding vector that can characterize the speech features of the target user. The speaker embedding model is a neural network that maps variable-length speech to fixed-length vectors. It consists of four parts: the first part is the input feature layer, which transforms the registered speech into frame-level acoustic features such as filter bank features and MFCC (Mel frequency cepstral coefficients); the second part is the frame-level feature extraction layer, which expands the receptive field layer by layer in the time dimension and extracts local temporal features related to the speaker; the third part is the temporal convergence layer, which converges the frame-level output of the entire speech in the time axis into a fixed-length representation, most commonly statistical pooling, which calculates the mean and standard deviation of the hidden layer output of the entire speech; the fourth part is the segment-level mapping layer and classification layer, which first transforms the pooling result into a low-dimensional speaker representation through one or more fully connected layers, and then adds a softmax classification layer set according to the number of speakers in the training phase, using the speaker classification task to constrain the network to learn to distinguish the representations of different speakers. Based on the speaker embedding vector, the registration speech modeling unit then establishes a correspondence between it and the target user's identity identifier and saves it as a registration template. When the registered speech contains multiple valid speech segments, multiple speaker embedding vectors can be extracted separately and then converged to form a target speaker representation corresponding to the target user, so that subsequent input speech can be similar to the target speaker representation and the identity can be confirmed.

[0045] The target speaker extraction unit transforms the mixed speech to the time-frequency domain to obtain the time-frequency features of the mixed speech. Simultaneously, it calls the target speaker representation already generated by the registered speech modeling unit and inputs this representation as a conditional vector representing the target user's speech identity features into the speech extraction model. Subsequently, the speech extraction model uses the time-frequency features of the mixed speech as the speech content input and the target speaker representation as the speaker conditional input to jointly model the speech content input and speaker conditional input. This allows the speech extraction model to estimate a target speech mask at each time frame and frequency unit that is more biased towards preserving the target user's speech and suppressing non-target speaker speech and other interference components. The target speech mask is then multiplied point-by-point with the amplitude spectrum corresponding to the mixed speech to obtain the estimated amplitude spectrum of the target user's speech. This is combined with the phase information of the mixed speech for inverse transformation, thereby outputting the target speech reconstruction result. The speech extraction model is a model structure consisting of a hybrid speech feature encoding unit, a target speaker condition guidance unit, and a target speech estimation unit. The hybrid speech feature encoding unit is used to characterize the time-frequency content of the current input speech, the target speaker condition guidance unit is used to introduce target user identity information, and the target speech estimation unit is used to output an estimation result that is more inclined to preserve the target user's speech and suppress non-target speaker speech and interference components under the constraint of target user identity.

[0046] The online speaker separation and annotation unit performs segmented analysis on the target speech reconstruction results along the time axis and outputs annotation results. The annotation results include target user active tags, non-target speaker tags, and overlapping speech tags. Specifically, the target speech reconstruction results are first divided into continuous time blocks along the time axis. For each time block, frame-level speaker representations are extracted, and the speaker activity state of the time block is calculated in combination with the generated target speaker representations. Then, the annotation results are iteratively updated according to the time block sequence. If a valid speech activity is detected in the current time block and matches the target speaker representation, a target user active tag is output. If a valid speech activity is detected in the current time block but does not match the target speaker representation, a non-target speaker tag is output. If multiple speech activities are estimated simultaneously in the same time block, overlapping speech tags are output in an overlapping speech manner, thereby realizing online segmented analysis and annotation along the time axis.

[0047] Based on the annotation results, the segment gating unit only sends the speech segments that are labeled as active tags of the target user to the keyword detection module, while caching, downgrading, or triggering repeat prompts for non-target speaker segments and overlapping speech segments.

[0048] Furthermore, the keyword detection module performs voice keyword detection on the active voice segments of the target user and outputs keyword recognition results representing consultation intent, consultation topic, and interaction control. This includes: after receiving the active voice segments of the target user released by the segment gating unit, the keyword detection module performs continuous voice analysis on the active voice segments of the target user, and identifies keyword information that can represent the current consultation content and interaction intent from the continuous spoken language flow. Among them, consultation topic keywords are used to represent the topic category to which the target user's current question belongs, context description keywords are used to represent the target user's description of relevant experiences, state changes, and question background, confirmation and negation keywords are used to represent the target user's confirmation or negation feedback to the system's question, and conversation control keywords are used to represent the target user's control intent to continue, pause, repeat, or end the current human-computer interaction process.

[0049] Meanwhile, the keyword detection module transcribes the same gated target user's speech to obtain a speech transcription result corresponding to the keyword recognition result, and provides the speech transcription result and the keyword recognition result together to the dialogue management module.

[0050] In psychological counseling support scenarios, target users' speech is often characterized by frequent pauses, irregular sentence structures, repetitions, and colloquialisms. Therefore, relying solely on complete syntactic parsing can easily increase recognition errors due to the instability of the expression. Based on this, this embodiment retains the continuous speech keyword detection link to quickly capture counseling intent, topic categories, and interaction control information from the target user's continuous expressions. At the same time, the speech transcription unit is used to transcribe the gated target user's speech, enabling the dialogue management module to comprehensively use keywords and contextual text to more accurately understand and process the current conversation content.

[0051] Furthermore, the dialogue management module determines the current response target and generates a psychological counseling response strategy. This includes: after receiving the keyword information recognition results, speech transcription results, and target user speech segment recognition results output by the keyword detection module and the target speaker recognition module, the dialogue management module does not process each result in isolation. Instead, it combines the current conversation context state corresponding to the current conversation stage to make a comprehensive judgment on the current input content, in order to determine whether this round of interaction was initiated by the target user, what consultation stage the current expression belongs to, and what response method the system should adopt. Unlike ordinary question-and-answer systems that primarily use single-round question-and-answer matching, the dialogue management module in this embodiment sets up an access state, a complaint collection state, a follow-up clarification state, a guidance feedback state, and a summary and conclusion state for psychological counseling assistance scenarios. The access state confirms the current conversation user and the intention to begin the consultation, preventing the system from mistakenly taking the speech of a non-target speaker as formal consultation input at the initial stage of the conversation. The complaint collection state guides the target user to describe their current distress, and continuously records the target user's expression based on keyword recognition and speech-to-text results. The follow-up clarification state further inquires about the time, frequency, triggers, and scope of impact based on keywords and transcribed content, in order to form a more complete contextual understanding of the current distress. The guidance feedback state calls up reassurance templates, suggestion templates, or practice templates to output feedback, enabling the system to provide an interactive response that matches the current consultation content after recognizing the target user's expression. The summary and conclusion state summarizes the content of this round of interaction and provides options for the next step, forming a relatively complete conversation loop. When the dialogue management module performs state transitions between the above states, it continuously receives speaker recognition results and dynamically adjusts the current state processing strategy based on the speaker recognition results. Once it detects that a non-target speaker is interrupting or that the proportion of overlapping speech is increasing, it can dynamically raise the gate threshold, suspend the recording of the main complaint, or issue a privacy protection prompt to reduce the impact of non-target speech on the current conversation processing, thereby enabling the target speaker recognition module and the dialogue management module to form a continuous linkage relationship during the interaction process.

[0052] Furthermore, the interactive response output module, based on the psychological counseling response strategy, outputs voice response information and provides human-computer interaction feedback. This includes: after receiving the psychological counseling response strategy generated by the dialogue management module, the interactive response output module converts the response strategy into feedback content for the target user and selects the appropriate output method to present it according to the current conversation state. The output method can be either speech synthesis to broadcast the current response content to the target user in natural speech, or it can simultaneously output text summaries, candidate buttons, breathing training animations, or soothing interface prompts through the display interface, thereby allowing the target user to obtain visual interactive information while receiving voice feedback. In implementation scenarios with virtual avatars or robot carriers, the interactive response output module can further drive the virtual avatar to perform nodding actions, lighting changes, and other corresponding human-computer interaction feedback to enhance the continuity and perceptibility of the feedback. It should be noted that the interactive response output module does not simply read out text content. Instead, it works together with the voice acquisition module, the target speaker recognition module, the keyword detection module, and the dialogue management module to form a complete human-computer interaction closed loop. That is, after the user speaks, the system first confirms the speaker, then extracts keywords and text, and generates a response based on the conversation status. The system then provides feedback to the target user through voice and visual means, allowing the target user to speak again and enter the next round of interaction.

[0053] In this embodiment, the speech recognition-based human-computer interaction response system can be deployed in a call center or a voice interaction management platform. Before formally executing voice acquisition, target user identification, keyword detection, and response output, administrators can first complete task creation and basic parameter configuration through the task creation interface, ensuring that operational parameters such as the dialogue model, voice interaction link, and execution period correspond to the subsequent human-computer interaction response process. In an optional platform deployment scenario, the system can be integrated into the voice interaction management platform for task configuration and testing verification, such as... Figure 4 The diagram shows the interface for task creation on the platform side of the system of the present invention. After entering the corresponding function page, the operator clicks "Create Task" to create a running task to carry out the interactive response process of the present invention.

[0054] To verify the response effect, voice broadcast effect, and human-computer interaction consistency of the interactive response output module in a real call environment, the system can also provide a voice test function before formal operation. Operators can select a test mode with preset background noise conditions to simulate application scenarios with environmental noise; they can also choose to conduct connection tests with target users to construct a highly realistic voice interaction process that closely resembles a real business environment. Through this voice test process, the system's voice reception, response output, interaction rhythm, and call connection status can be pre-verified. Figure 5The diagram shown is a schematic of the interface for voice testing and connection verification on the platform side of the system of the present invention.

[0055] like Figure 6 As shown, a human-computer interaction response method based on speech recognition includes the following steps:

[0056] Step 1: Collect conversational audio from a psychological counseling scenario and output mixed audio.

[0057] Step 2: Perform target user identification and time segmentation on the mixed speech, filter out the speech segments belonging to the target user, and suppress the speech of non-target speakers and overlapping speech interference.

[0058] Step 3: Detect voice keywords in the active voice segments of the target user and output the keyword recognition results that represent the consultation intent, consultation topic and interaction control.

[0059] Step 4: Identify the current respondent and generate a psychological counseling response strategy.

[0060] Step 5: Based on the psychological counseling response strategy, output voice response information and provide human-computer interaction feedback.

[0061] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0062] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A human-computer interaction response system based on speech recognition, characterized in that, include: The voice acquisition module collects conversational voices in psychological counseling scenarios and outputs mixed voices. The target speaker recognition module performs target user identification and time-sequence segmentation on mixed speech, filters out speech segments belonging to the target user, and suppresses speech interference from non-target speakers and overlapping speech. The keyword detection module performs voice keyword detection on active voice segments of the target user and outputs keyword recognition results that represent consultation intent, consultation topics and interactive control. The dialogue management module identifies the current respondent and generates psychological counseling response strategies. The interactive response output module outputs voice response information and provides human-computer interaction feedback based on psychological counseling response strategies. The target speaker recognition module extracts acoustic features to characterize the speaker based on the registered voice of the target user, inputs the acoustic features into the speaker embedding model to obtain speaker embedding vectors, extracts multiple speaker embedding vectors, and converges them to form a target speaker representation corresponding to the target user. The mixed speech is then transformed to the time-frequency domain to obtain the time-frequency features of the mixed speech. The target speaker representation is used as a conditional vector representing the voice identity features of the target user. The time-frequency features and the conditional vector are input into the speech extraction model to output the target speech reconstruction result. The target speech reconstruction result is segmented and analyzed along the time axis and the annotation result is output. Based on the annotation result, the speech segments labeled as target users are sent to the keyword detection module.

2. The human-computer interaction response system based on speech recognition according to claim 1, characterized in that, After receiving a voice segment from a target user, the keyword detection module performs continuous voice analysis on the active voice segments of the target user, identifies keyword information that can represent the current consultation content and interaction intent from the continuous spoken language flow, and transcribes the target user's voice to obtain a speech transcription result corresponding to the recognition result of the keyword information.

3. The human-computer interaction response system based on speech recognition according to claim 2, characterized in that, The keyword information recognition results, speech transcription results, and target user speech segment recognition results output by the target speaker recognition module are input together into the dialogue management module. The dialogue management module does not process each result in isolation, but combines the current conversation context state corresponding to the current stage of the conversation to determine the current response object and generate a psychological counseling response strategy.

4. The human-computer interaction response system based on speech recognition according to claim 3, characterized in that, The interactive response output module outputs voice feedback according to the psychological counseling response strategy, and can drive the display interface to display the corresponding feedback content, or drive the virtual image or robot to perform corresponding interactive actions.

5. The human-computer interaction response system based on speech recognition according to claim 1, characterized in that, The target speaker recognition model transforms the mixed speech to the time-frequency domain to obtain the time-frequency features of the mixed speech. The target speaker representation is used as a conditional vector representing the target user's speech identity features. The time-frequency features and conditional vector are input into the speech extraction model, which outputs the target speech reconstruction result. This includes: the speech extraction model uses the time-frequency features of the mixed speech as the speech content input and the target speaker representation as the speaker conditional input, jointly modeling the speech content input and speaker conditional input. This allows the speech extraction model to estimate a target speech mask that is more biased towards preserving the target user's speech and suppressing non-target speaker speech and other interference components at each time frame and frequency unit. Then, the target speech mask is multiplied point-by-point with the amplitude spectrum corresponding to the mixed speech to obtain the estimated amplitude spectrum of the target user's speech. Finally, an inverse transform is performed using the phase information of the mixed speech to output the target speech reconstruction result.

6. The human-computer interaction response system based on speech recognition according to claim 5, characterized in that, The speech extraction model is a model structure consisting of a hybrid speech feature encoding unit, a target speaker condition guidance unit, and a target speech estimation unit. The hybrid speech feature encoding unit is used to characterize the time-frequency content of the current input speech, the target speaker condition guidance unit is used to introduce target user identity information, and the target speech estimation unit is used to output an estimation result that is more inclined to preserve the target user's speech and suppress non-target speaker speech and interference components under the constraint of target user identity.

7. The human-computer interaction response system based on speech recognition according to claim 1, characterized in that, The speaker embedding model is a neural network that maps variable-length speech to fixed-length vectors. It consists of four parts: the first part is the input feature layer, which transforms the registered speech into frame-level acoustic features such as filter bank features and Mel frequency cepstral coefficients; the second part is the frame-level feature extraction layer, which expands the receptive field layer by layer in the time dimension and extracts local temporal features related to the speaker; the third part is the temporal convergence layer, which converges the frame-level output of the entire speech in the time axis into a fixed-length representation; the fourth part is the segment-level mapping layer and classification layer, which first transforms the pooling result into a low-dimensional speaker representation through one or more fully connected layers, and then adds a softmax classification layer set according to the number of trainees during the training phase, using the speaker classification task to constrain the network to learn to distinguish the representations of different speakers.

8. The human-computer interaction response system based on speech recognition according to claim 1, characterized in that, The voice acquisition module adopts a single microphone, dual microphone, or microphone array structure and is integrated into robots, mobile terminals, information kiosks, self-service machines, or desktop devices.

9. The human-computer interaction response system based on speech recognition according to claim 3, characterized in that, The dialogue management module sets access status, chief complaint collection status, follow-up questioning and clarification status, guidance and feedback status, and summary and end status. After receiving the keyword recognition result, the speech transcription result, and the speaker recognition result, it makes a comprehensive judgment on the current input content in combination with the current state of the conversation to determine whether the current interaction was initiated by the target user, the consultation stage to which the current expression belongs, and the psychological counseling response strategy corresponding to the consultation stage.

10. A human-computer interaction response method based on speech recognition, characterized in that, The human-computer interaction response system based on speech recognition, as described in any one of claims 1-9, comprises the following steps: Step 1: Collect conversational audio from a psychological counseling scenario and output mixed audio. Step 2: Perform target user identification and time sequence segmentation on the mixed speech, filter out the speech segments belonging to the target user, and suppress the speech of non-target speakers and overlapping speech interference; Step 3: Detect voice keywords in the active voice segments of the target user and output the keyword recognition results that represent the consultation intent, consultation topic and interaction control; Step 4: Identify the current respondent and generate a psychological counseling response strategy; Step 5: Based on the psychological counseling response strategy, output voice response information and provide human-computer interaction feedback.

Images