Voiceprint feature extraction method for elderly patients in voice outpatient registration

Abstract

The present application relates to the technical field of speech signal processing, in particular to a voice clinic reporting elderly patient voiceprint feature extraction method. The method comprises: performing similarity analysis on adjacent speech frames of difference segments in the original speech segment of the target patient to obtain a plurality of similarity values; performing information entropy calculation based on the plurality of similarity values to obtain a decay index; compensating the initial voiceprint feature vector corresponding to the original speech segment based on the target patient's registered voiceprint feature vector, reference feature vector and decay index to obtain a target voiceprint feature vector; and performing voiceprint comparison based on the target voiceprint feature vector and the registered voiceprint feature vector to perform voice clinic reporting. The present application can avoid the false rejection or false pass problem caused by the fixed threshold in the related art being unable to adapt to the voiceprint feature fluctuation of elderly patients, and can improve the accuracy of voiceprint feature extraction for elderly patients.

Description

Technical Field

[0001] This invention relates to the field of speech signal processing technology, specifically to a method for capturing the voiceprint features of elderly patients reporting to a medical clinic. Background Technology

[0002] With the popularization of smart healthcare, voice-activated outpatient check-in systems are gradually being used among elderly patients.

[0003] In related technologies, voice-based outpatient registration systems typically employ a fixed-threshold voiceprint recognition system, requiring patients to fully recite the name of their department (e.g., "Cardiovascular Medicine") for registration. However, elderly patients experience significant fluctuations in voiceprint features due to physiological voice decay (unstable breathing, trembling, or elision), leading to insufficient stability in system feature extraction. This causes the fixed threshold to be unable to adapt to the inter-frame fluctuations in the voiceprint feature vector, resulting in decreased voiceprint discrimination under the fixed threshold. Consequently, the risk of false rejection or false approval during voice-based outpatient registration increases.

[0004] This shows that the accuracy of voiceprint feature capture for elderly patients is relatively low in related technologies. Summary of the Invention

[0005] The purpose of this invention is to provide a method for capturing the voiceprint features of elderly patients during voice-based outpatient registration, in order to solve the technical problem of low accuracy in capturing the voiceprint features of elderly patients in related technologies.

[0006] In a first aspect, embodiments of the present invention provide a method for capturing the voiceprint features of elderly patients reporting for appointment at a medical clinic, the method comprising: Similarity analysis is performed on adjacent speech frames of the difference segments in the original speech segment of the target patient to obtain multiple similarity values. The similarity value is used to characterize the degree of similarity of acoustic features between corresponding adjacent speech frames. The difference segments are used to characterize speech segments in the original speech segment whose acoustic features have decayed. The difference segments are obtained based on the analysis of the original speech segment and the standard speech template corresponding to the department to be reported to. The department to be reported to is used to indicate the department that the target patient should report to. The department to be reported to is determined by matching the original speech segment and multiple standard speech templates. Information entropy is calculated based on multiple similarity values ​​to obtain the decay index, which is used to characterize the severity of pronunciation degradation of the original speech segment. Based on the registered voiceprint feature vector, reference feature vector, and decay index of the target patient, the initial voiceprint feature vector corresponding to the original speech segment is compensated to obtain the target voiceprint feature vector. The reference feature vector is obtained by feature extraction based on the reference speech segment recorded when the target patient pronounces the speech in a non-standard reading manner. The registered voiceprint feature vector is used to indicate the voiceprint feature vector of the baseline speech segment registered by the target patient during the first visit and registration. Voiceprint comparison is performed between the target voiceprint feature vector and the registered voiceprint feature vector to enable voice-based clinic check-in.

[0007] In one embodiment, the step of obtaining the difference fragment includes: The DTW algorithm is used to calculate the distance frame by frame between the original speech segment of the target patient and the standard speech template corresponding to the department to be reported to, and obtain multiple local distance values. The number of local distance values ​​is positively correlated with the number of frames of the original speech segment of the target patient. The multiple local distance values ​​are arranged in temporal order according to the corresponding frames in the original speech segment to obtain a difference sequence; The percentile method is used to extract the threshold of the difference sequence to obtain the difference threshold; By analyzing the relative magnitudes of multiple local distance values ​​and the difference threshold in the difference sequence, multiple candidate distance values ​​are obtained, among which the candidate distance values ​​are greater than the difference threshold. The original speech segments corresponding to temporally adjacent frames among the multiple candidate distance values ​​are identified as the difference segments.

[0008] In one embodiment, the similarity analysis of adjacent speech frames of the differing segments in the original speech segment of the target patient yields multiple similarity values, including: Feature extraction is performed on each frame of the difference segments in the original speech of the target patient to obtain the feature vector corresponding to each frame; For each pair of adjacent frames in the difference segment, the cosine similarity is calculated based on the feature vectors corresponding to the adjacent frames to obtain the similarity value of each pair of adjacent frames. Among them, multiple similarity values ​​are the similarity values ​​of multiple pairs of adjacent frames.

[0009] In one embodiment, the calculation of information entropy based on multiple similarity values ​​to obtain the decay index includes: Based on dividing the value range of multiple similar values ​​into multiple preset intervals, the number of the multiple similar values ​​falling into each of the preset intervals is counted; The information entropy is calculated based on the number of items falling into each of the preset intervals, and the decay index is obtained.

[0010] In one embodiment, the compensation of the initial voiceprint feature vector corresponding to the original speech segment based on the registered voiceprint feature vector, reference feature vector, and decay index of the target patient to obtain the target voiceprint feature vector includes: When the deviation distance is greater than the distance threshold, the difference between the registered voiceprint feature vector and the reference feature vector of the target patient is analyzed to obtain an additional deviation vector. The additional deviation vector is used to characterize the degree of additional deviation of the voiceprint feature of the target patient relative to the inherent decay pattern of the target patient. The deviation distance is used to characterize the degree of deviation between the voiceprint feature of the original speech segment and the voiceprint feature of the target patient at the time of registration. The deviation distance is obtained based on the difference analysis between the original speech segment and the reference speech segment. Based on the initial voiceprint feature vector, the additional deviation vector, the decay index, and the preset compensation matrix, the initial voiceprint feature vector corresponding to the original speech segment is compensated to obtain the target voiceprint feature vector.

[0011] In one embodiment, the step of obtaining the deviation distance includes: Feature extraction is performed on the original speech segments of the target patient to obtain the initial voiceprint feature vector; The cosine distance between the initial voiceprint feature vector and the registered voiceprint feature vector is calculated to obtain the deviation distance. The registered voiceprint feature vector is used to indicate the feature vector obtained by feature extraction of the reference speech segment when the target patient registers in the department to be reported.

[0012] In one embodiment, the step of obtaining the distance threshold includes: The historical average deviation distance is calculated by taking the mean of the historical deviation distances at each time of the target patient’s historical successful check-in records. The standard deviation of historical deviation is calculated based on the historical deviation distance at each time of the target patient’s historical successful check-in record. Data analysis is performed based on historical average deviation distance, historical standard deviation, and decay index to obtain a distance threshold, wherein the distance threshold is positively correlated with the historical average deviation distance, the distance threshold is positively correlated with the historical standard deviation, and the distance threshold is positively correlated with the decay index.

[0013] In one embodiment, the step of obtaining the department to which one is to report includes: The original speech segment of the target patient is matched with the standard speech templates corresponding to each of the multiple booked departments to obtain the matching distance between the original speech segment and each booked department. Each booked department has multiple standard speech templates, which are used to indicate the standard speech segment of the official full name and the standard speech segment of the medical abbreviation of the corresponding department. Based on the matching distance between the original voice segment and multiple pre-booked departments, the department to which the target patient is to report is selected from the multiple pre-booked departments. The matching distance between the original voice segment and the department to which the patient is to report is less than the matching distance between the original voice segment and any other pre-booked department other than the department to which the patient is to report.

[0014] In one embodiment, matching the original speech segment of the target patient with the standard speech templates corresponding to each of the multiple booked departments to obtain the matching distance between the original speech segment and each booked department includes: Based on the weighted dynamic time warping matching of multiple standard speech templates corresponding to the candidate departments with the original speech segments of the target patients, the weighted matching distance between the original speech segments and each standard speech template corresponding to the candidate departments is obtained. The candidate department is any one of multiple booked departments. The minimum weighted matching distance between the original speech segment and each standard speech template corresponding to the candidate department is determined as the matching distance between the original speech segment and the candidate department.

[0015] In one embodiment, the step of selecting the department to which the target patient should report from multiple booked departments based on the matching distance between the original voice segment and the corresponding departments includes: The relative magnitudes of the matching distances between the original speech segment and multiple booked departments are analyzed to obtain the first matching distance and the second matching distance. The first matching distance is used to indicate the minimum value among the matching distances between the original speech segment and multiple booked departments, and the second matching distance is used to indicate the second minimum value among the matching distances between the original speech segment and multiple booked departments. Calculate the ratio of the second matching distance to the first matching distance to obtain the confidence interval ratio; If the confidence interval ratio is greater than or equal to the ratio threshold, the department with the appointment corresponding to the first matching distance is determined as the department to be visited by the target patient.

[0016] The present invention has the following beneficial effects: This invention performs similarity analysis on adjacent speech frames of differing segments in the original speech segment of a target patient to obtain multiple similarity values. Information entropy is then calculated based on these similarity values ​​to obtain a decay index. Based on the target patient's registered voiceprint feature vector, reference feature vector, and decay index, the initial voiceprint feature vector corresponding to the original speech segment is compensated to obtain a target voiceprint feature vector. Voiceprint comparison is then performed between the target voiceprint feature vector and the registered voiceprint feature vector for voice-based outpatient registration. This invention considers the impact of physiological voice decay in elderly patients on voiceprint features. By compensating the initial voiceprint feature vector to obtain the target voiceprint feature vector, voiceprint comparison is performed based on the target voiceprint feature vector. This avoids the problem in related technologies where fixed thresholds cannot adapt to the fluctuations in the voiceprint features of elderly patients, leading to increased risks of false rejection or false approval. This invention improves the accuracy of voiceprint feature capture for elderly patients. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating a method for capturing the voiceprint features of elderly patients during voice-based outpatient registration, as provided in an embodiment of the present invention. Detailed Implementation

[0018] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0019] This invention is applicable to scenarios involving elderly patients checking into outpatient clinics via voice. For example, elderly patients may need to visit multiple departments on the same day, and due to age-related physiological voice decay such as unstable breath control, trembling, elision, or uneven speech rate, they may be unable to consistently recite the full names of the departments. The following, in conjunction with the accompanying drawings, details a specific scheme for capturing the voiceprint features of elderly patients during outpatient clinic check-ins, as provided by this invention.

[0020] This invention proposes a method for capturing the voiceprint features of elderly patients during voice-based outpatient registration. The method provided by this invention can be executed by a voiceprint feature capture system, specifically by the controller within that system. Please refer to [link to relevant documentation]. Figure 1 The diagram illustrates a flowchart of a method for capturing the voiceprint features of elderly patients during voice-based outpatient registration, according to an embodiment of the present invention. The method includes the following steps: Step S1: Perform similarity analysis on adjacent speech frames of the different segments in the original speech segment of the target patient to obtain multiple similarity values.

[0021] The similarity value characterizes the degree of acoustic feature similarity between corresponding adjacent speech frames, while the difference segment characterizes speech segments in the original speech segment whose acoustic features have deteriorated. The difference segment is obtained based on the analysis of the original speech segment and the standard speech template corresponding to the department to be reported to. The department to be reported to indicates the department that the target patient should currently report to, and the department to be reported to is determined by matching the original speech segment with multiple standard speech templates.

[0022] The mobile device responds to the target patient's voice check-in trigger operation, activates the microphone, and begins collecting the target patient's voice stream. Upon receiving the target patient's end-of-collection command (such as a stop operation triggered by the target patient), it stops collecting the voice stream, obtaining the recorded voice segment. This recorded voice segment is then processed for noise reduction and uploaded to the server via the network. In other words, starting the voice stream collection is contingent upon the "target patient's voice check-in trigger operation," and ending the voice stream collection is contingent upon the "target patient's end-of-collection command." This means that the collection of the voice stream (also known as voice information) is under the active control of the target patient from beginning to end. It is understood that this invention provides a corresponding notification obligation, allowing the target patient to decide whether to participate in the program. Furthermore, in this embodiment of the invention, after voiceprint feature extraction is completed, i.e., after the voice information is used up, the original voice data (i.e., the voice stream) is automatically destroyed and no longer stored or used for other purposes.

[0023] The controller and server can communicate via wired or wireless network to obtain the original speech segment of the target patient from the server. This original speech segment is a clean speech signal with environmental noise removed by the server; for example, the server removes stationary environmental noise from the speech segment using spectral subtraction.

[0024] Specifically, the similarity analysis of adjacent speech frames of the differing segments in the original speech segment of the target patient yields multiple similarity values, including: Feature extraction is performed on each frame of the difference segments in the original speech of the target patient to obtain the feature vector corresponding to each frame; For each pair of adjacent frames in the difference segment, the cosine similarity is calculated based on the feature vectors corresponding to the adjacent frames to obtain the similarity value of each pair of adjacent frames. Among them, multiple similarity values ​​are the similarity values ​​of multiple pairs of adjacent frames.

[0025] For each frame of the differential segment in the original speech segment of the target patient, the controller can convert it into Mel frequency and then perform cepstral analysis to obtain the Mel-frequency cepstral coefficients (MFCC) feature vector corresponding to each frame.

[0026] Subsequently, the controller can calculate the similarity value of adjacent frames based on the MFCC feature vector corresponding to each frame. For example, the similarity value of the i-th pair of adjacent frames... It can be represented as: In the formula, Let i be the MFCC feature vector corresponding to the i-th frame. Let be the magnitude of the MFCC feature vector corresponding to the i-th frame. This is the MFCC feature vector corresponding to the (i+1)th frame. Let be the magnitude of the MFCC feature vector corresponding to the (i+1)th frame. The similarity value between the i-th pair of adjacent frames. The closer to 1, the more similar the acoustic features of the i-th frame and the (i+1)-th frame.

[0027] In one example, the steps to obtain the difference fragments include: The DTW algorithm is used to calculate the distance frame by frame between the original speech segment of the target patient and the standard speech template corresponding to the department to be reported to, and obtain multiple local distance values. The number of local distance values ​​is positively correlated with the number of frames of the original speech segment of the target patient. The multiple local distance values ​​are arranged in temporal order according to the corresponding frames in the original speech segment to obtain a difference sequence; The percentile method is used to extract the threshold of the difference sequence to obtain the difference threshold; By analyzing the relative magnitudes of multiple local distance values ​​and the difference threshold in the difference sequence, multiple candidate distance values ​​are obtained, among which the candidate distance values ​​are greater than the difference threshold. The original speech segments corresponding to temporally adjacent frames among the multiple candidate distance values ​​are identified as the difference segments.

[0028] The server also stores standard voice templates for the departments to be reported to. Similarly, the controller can obtain the standard voice templates for the departments to be reported to via the network.

[0029] The controller pre-stores a dynamic time warping (DTW) algorithm. This DTW algorithm is used to perform equal-weighted fine alignment of the target patient's original speech segment with the standard speech template corresponding to the department to which the patient is reporting. The controller calculates the inter-frame distance between the original speech segment and the standard speech template frame by frame along the optimal alignment path, obtaining multiple local distance values. Subsequently, the controller arranges these local distance values ​​according to the temporal order of their corresponding frames within the original speech segment, obtaining a difference sequence. A percentile method is then used to extract a threshold from this difference sequence; for example, the 85th percentile is used as the difference threshold. The controller then compares the relative magnitude of each local distance value in the difference sequence with the difference threshold. Local distance values ​​greater than the difference threshold are selected as candidate distance values. The portions of the original speech segments corresponding to temporally adjacent frames within the multiple candidate distance values ​​are then identified as difference segments. It is understood that local distance values ​​less than or equal to the difference threshold are not considered candidate distance values.

[0030] Optionally, before using the percentile method for threshold extraction, the controller can smooth the difference sequence, for example, by using a sliding window averaging method to smooth the difference sequence, in order to eliminate the interference of random fluctuations in individual frames on threshold extraction.

[0031] Analysis revealed that vocal degeneration in elderly patients typically manifests as unstable acoustic features over consecutive time periods. Therefore, by locating the continuous speech region with the most concentrated decay in the original speech segment, i.e., the difference segment, and then analyzing the difference segment, the dilution of the decay signal caused by the overall averaging of the original speech segment can be avoided. This allows for more accurate capture of abrupt changes in acoustic features between frames, enabling a refined quantification of the degree of vocal degeneration in elderly patients and thus improving the accuracy of obtaining the decay index.

[0032] In one example, the steps to obtain the department to which one is to report include: The original speech segment of the target patient is matched with the standard speech templates corresponding to each of the multiple booked departments to obtain the matching distance between the original speech segment and each booked department. Each booked department has multiple standard speech templates, which are used to indicate the standard speech segment of the official full name and the standard speech segment of the medical abbreviation of the corresponding department. Based on the matching distance between the original voice segment and multiple pre-booked departments, the department to which the target patient is to report is selected from the multiple pre-booked departments. The matching distance between the original voice segment and the department to which the patient is to report is less than the matching distance between the original voice segment and any other pre-booked department other than the department to which the patient is to report.

[0033] Elderly patients often have difficulty accurately repeating the official full name (such as Department of Cardiovascular Medicine) when checking in at the outpatient clinic, which leads to difficult interaction with the check-in method that requires a complete repetition of the official full name of the department. In view of this, the embodiments of the present invention configure a plurality of standard speech templates for each pre-registered department, which respectively correspond to various expression forms such as the official full name and medical abbreviation of the same department, and match the original speech segment of the target patient with the multiple templates of each pre-registered department respectively, and select the department with the minimum matching distance from them as the department to be checked in. Since the multiple templates of each department all point to the same department, even if the target patient expresses with a medical abbreviation, the template of the corresponding department can still be matched, so that the department to be checked in can still be identified when the patient cannot accurately repeat the standard name, thereby improving the success rate and interaction friendliness of voice outpatient check-in.

[0034] Optionally, configuring a plurality of standard speech templates for each pre-registered department may further include: a standard speech segment of a common folk name and a standard speech segment of a doctor's surname designation, wherein the doctor's surname designation is determined according to the surname of the doctor on duty of the pre-registered department on the current day. For example, the common folk name of "Department of Cardiovascular Medicine" is "Cardiology Department"; the doctor's surname designation of "Department of Cardiovascular Medicine" is "Doctor Wang". In this way, when the target patient expresses with a common folk name or a doctor's surname designation, the template of the corresponding department can still be matched.

[0035] Wherein, the process of matching the original speech segment of the target patient with the standard speech templates corresponding to each of the plurality of pre-registered departments respectively to obtain the matching distance between the original speech segment and each pre-registered department comprises: Performing weighted dynamic time warping matching on the plurality of standard speech templates corresponding to the candidate department and the original speech segment of the target patient respectively, to obtain the weighted matching distance between the original speech segment and each standard speech template corresponding to the candidate department, wherein the candidate department is any one of the plurality of pre-registered departments; Determining the minimum value among the weighted matching distances between the original speech segment and each standard speech template corresponding to the candidate department as the matching distance corresponding to the original speech segment and the candidate department.

[0036] The controller pre-configures a frame-level weight sequence for each standard speech template, wherein the frame-level weight is determined according to the importance of different semantic components in the template. Taking the Department of Cardiovascular Medicine as an example, it corresponds to 4 standard speech templates: Department of Cardiovascular Medicine (standard full name), Department of Cardiology (medical abbreviation), Heart Department (common folk name), and Doctor Wang (doctor's surname designation). Among them, the weight of the frame corresponding to the core entity words such as "Department" and "Wang" in "Doctor Wang" is 1.5, the weight of the frame corresponding to the modifying and limiting words such as "Cardiovascular", "Gastroenterology" and "Morning" is 0.5, and the weight of the frame corresponding to the conjunctions and function words such as "le" and "de" is 0.2.

[0037] The controller can access the hospital's data interface to query the information system and obtain records of all departments where the target patient has made appointments but has not yet reported for the day. This allows the controller to obtain the official full name of the department, appointment time, and doctor's name for the target patient's appointments. For any department with an appointment (i.e., a candidate department), the controller can perform weighted dynamic time warping (DTW) matching between the target patient's original speech segment and multiple standard speech templates corresponding to the candidate department, calculating the weighted matching distance between the original speech segment and each standard speech template corresponding to the candidate department.

[0038] Specifically, for each standard speech template of the candidate department, the controller can first obtain the weighted local distance between each speech frame in the original speech segment and each speech frame of the standard speech template. Then, the DTW algorithm is used to perform time warping on the frame sequence of the original speech segment and the standard speech template to obtain the optimal alignment path between the start frame to the end frame of the original speech segment and the start frame to the end frame of the standard speech template. The weighted local distance is dynamically accumulated along the optimal alignment path to obtain the cumulative weighted distance. Then, the cumulative weighted distance is divided by the total number of frames of the standard speech template to obtain the frame average weighted distance, thereby obtaining the weighted matching distance between the original speech segment and each standard speech template corresponding to the candidate department.

[0039] For example, the weighted local distance between the i-th frame of the original speech segment and the j-th frame of the standard speech template currently being matched for the candidate department. It can be represented as: In the formula, The weight of the j-th frame of the standard speech template currently being matched for the candidate department. is the Euclidean distance between the MFCC feature vector of the i-th frame of the original speech segment and the MFCC feature vector of the j-th frame of the standard speech template currently being matched for the candidate department. Let be the MFCC feature vector of the i-th speech frame in the original speech segment. This is the MFCC feature vector of the j-th frame of the standard speech template currently being matched, corresponding to the candidate department. The implementation method of the controller obtaining the MFCC feature vector of each speech frame in each reserved standard speech template is similar to the implementation method of the controller obtaining the MFCC feature vector of each frame in the difference segment of the original speech segment, and will not be described in detail here.

[0040] For example, the cumulative weighted distance can be expressed as: In the formula, It is the cumulative weighted distance between the first i frames of the original speech segment and the first j frames of the standard speech template currently being matched for the candidate department. That is, during the dynamic time warping process, the cumulative weighted distance between the first to the i frames of the original speech segment and the first to the j frames of the standard speech template currently being matched for the candidate department under the optimal alignment path. It is the weighted local distance between the i-th frame of the original speech segment and the j-th frame of the standard speech template currently being matched for the candidate department. The cumulative weighted distance between the first i-1 frames of the original speech segment and the first j frames of the standard speech template currently being matched for the candidate department means that the standard speech template currently being matched for the candidate department has one more frame than the original speech segment. The cumulative weighted distance between the first i frames of the original speech segment and the first j-1 frames of the standard speech template currently being matched for the candidate department means that the standard speech template currently being matched for the candidate department is one frame shorter than the original speech segment. The cumulative weighted distance between the first i-1 frames of the original speech segment and the first j-1 frames of the standard speech template currently being matched for the candidate department means the same number of frames as the standard speech template currently being matched for the candidate department.

[0041] Since the lengths (total number of frames) of different standard speech templates may vary, the cumulative weighted distance of longer standard speech templates is naturally larger, causing the matching results to favor shorter templates. Based on this, the controller can divide the cumulative weighted distance between the original speech segment and the currently matched standard speech template corresponding to the candidate department at the termination frame by the total number of frames of that standard speech template, to obtain the weighted matching distance between the original speech segment and the currently matched standard speech template corresponding to the candidate department.

[0042] For example, assuming the candidate department is the kth department, the matching distance between the original speech segment and the candidate department is... It can be represented as: In the formula, This provides all standard voice templates for the k-th department. Original speech segment With the kth department A standard voice template The weighted matching distance between them. Original speech segment With the kth department A standard voice template The cumulative weighted distance. For the kth department A standard voice template Total number of frames.

[0043] The process of selecting the target patient's department to report to from multiple booked departments based on the matching distance between the original voice segment and the corresponding departments includes: The relative magnitudes of the matching distances between the original speech segment and multiple booked departments are analyzed to obtain the first matching distance and the second matching distance. The first matching distance is used to indicate the minimum value among the matching distances between the original speech segment and multiple booked departments, and the second matching distance is used to indicate the second minimum value among the matching distances between the original speech segment and multiple booked departments. Calculate the ratio of the second matching distance to the first matching distance to obtain the confidence interval ratio; If the confidence interval ratio is greater than or equal to the ratio threshold, the department with the appointment corresponding to the first matching distance is determined as the department to be visited by the target patient.

[0044] Elderly patients may exhibit speech errors under stress, such as referring to the gastroenterology department as the respiratory department, or they may confuse their descriptions, such as mentioning both the cardiology and respiratory departments of Dr. Wang. In such cases, directly identifying the department corresponding to the smallest matching distance as the department to be registered for might lead to incorrect department identification. Therefore, this invention further analyzes the distribution characteristics of matching distances and calculates the ratio of the smallest matching distance to the second smallest matching distance (i.e., the confidence interval ratio) to assess the reliability of the current matching result. When the confidence interval ratio is greater than or equal to a threshold, indicating high reliability, the department with the first matching distance (i.e., the smallest matching distance) that has already been booked is identified as the department to be registered for. This avoids misidentification due to the target patient's speech errors or confusion, thus improving the accuracy of department identification.

[0045] For example, confidence interval ratio It can be represented as: In the formula, The second matching distance, This is the first matching distance.

[0046] The ratio threshold is pre-stored by the controller, for example, 1.6. The controller can compare the confidence interval ratio with the ratio threshold. If the confidence interval ratio is greater than or equal to the ratio threshold, the controller determines the booked department corresponding to the first matching distance as the department to be visited by the target patient.

[0047] Optionally, in this embodiment of the invention, if the confidence interval ratio is less than the ratio threshold and greater than or equal to a preset threshold (e.g., 1.2), the controller can trigger a confirmation interaction process. This confirmation interaction process includes: the controller acquiring the booked departments corresponding to the first matching distance and the second matching distance respectively, and generating a secondary confirmation command to send to the target patient's mobile phone, instructing the mobile phone to synthesize a voice prompt, such as: "Do you want to report to [the common name of candidate department A] or [the common name of candidate department B]? Please say 'first' or 'second'," and receiving the target patient's confirmation voice, performing limited-word set voice command recognition on the confirmation voice, i.e., only distinguishing between "first" and "second," and determining the department to report to from the two candidate departments based on the recognition result.

[0048] Optionally, if the confidence interval ratio is less than a preset threshold, the controller can trigger the speech segmentation process. This speech segmentation process includes: the controller dividing the original speech segment into sliding segments with a fixed window length (e.g., 1 second), calculating the weighted matching distance between each sliding segment and each standard speech template of each booked department, and obtaining the candidate department with the smallest weighted matching distance for each sliding segment. If the candidate department with the smallest weighted matching distance for the top-ranked sliding segments (e.g., the first 50% of sliding segments) is Department A, and the candidate department with the smallest weighted matching distance for the bottom-ranked sliding segments (e.g., the last 50% of sliding segments) is Department B, and Department A and Department B are different, then it is determined that the target patient mentioned two different departments in the current speech segment. At this time, the controller uses the boundary position where Department A switches to Department B as the segmentation boundary to divide the original speech segment, obtaining the first speech sub-segment and the second speech sub-segment. Subsequently, the controller can calculate the weighted matching distance between the first speech segment and the standard speech templates corresponding to each booked department, and confirm the corresponding department based on the confidence interval ratio. It also calculates the weighted matching distance between the second speech segment and the standard speech templates corresponding to each booked department, and confirms the corresponding department based on the confidence interval ratio, thus enabling step-by-step check-in to each department sequentially according to the speech segment order. In this way, when a target patient's speech contains two different department names, the two department names can be separated into their respective speech segments for separate recognition, improving the accuracy of department identification.

[0049] Step S2: Calculate the information entropy based on multiple similarity values ​​to obtain the decay index.

[0050] The decay index is used to characterize the severity of pronunciation degradation of the original speech segment.

[0051] In this embodiment of the invention, when the total number of frames in the difference segment is less than 3, it means that the difference segment contains fewer than two pairs of adjacent frames, which cannot provide a statistically significant inter-frame similarity distribution and is insufficient to support information entropy calculation. In this case, the controller can directly determine that the decay index is 0, and no further operation is required. When the total number of frames in the difference segment is greater than or equal to 3, the controller can calculate the decay index in the following manner.

[0052] Specifically, the decay index is obtained by calculating information entropy based on multiple similarity values, including: Based on dividing the value range of multiple similar values ​​into multiple preset intervals, the number of the multiple similar values ​​falling into each of the preset intervals is counted; The information entropy is calculated based on the number of items falling into each of the preset intervals, and the decay index is obtained.

[0053] The range of multiple similar values ​​is [0,1]. The controller pre-stores five preset intervals, such as [0,0.6), [0.6,0.7), [0.7,0.8), [0.8,0.9), and [0.9,1]. Subsequently, the controller can calculate the information entropy and obtain the decay index by the number of multiple similar values ​​falling into each preset interval.

[0054] For example, decay index It can be represented as: In the formula, To fall into the first The ratio of the number of preset intervals to the total number of similar values.

[0055] Step S3: Based on the registered voiceprint feature vector, reference feature vector, and decay index of the target patient, compensate the initial voiceprint feature vector corresponding to the original speech segment to obtain the target voiceprint feature vector.

[0056] Among them, the registered voiceprint feature vector is used to indicate the voiceprint feature vector of the baseline speech segment registered by the target patient during the initial medical record-keeping process.

[0057] The reference feature vector is obtained by extracting features from a reference speech segment recorded when the target patient pronounces the text in a non-standard reading manner (such as reading in a natural and relaxed state). For example, when a target patient registers for the first time, the controller sends a guidance command to the mobile device. The mobile device responds to the command by guiding the target patient to read a preset text in a natural manner (allowing for prolongation and slight trembling) and records this speech segment as a reference speech segment. This reference speech segment is stored on the server and bound to the target patient's identifier. After determining the target patient's identity identifier (such as patient ID), the controller retrieves the target patient's reference speech segment from the server based on the target patient's identifier and performs feature extraction on the reference speech segment to obtain the target patient's reference feature vector. The controller pre-stores an x-vector voiceprint feature extraction network, which can be used to extract features from the reference speech segment to obtain the target patient's reference feature vector.

[0058] During the process of determining the department to be reported to through department matching, the controller determines the target patient's identity based on the identity identifier carried in the original voice segment. Alternatively, when the target patient triggers voice reporting, the controller determines the target patient's identity based on the account already logged in on the mobile device.

[0059] Specifically, the compensation of the initial voiceprint feature vector corresponding to the original speech segment based on the registered voiceprint feature vector, reference feature vector, and decay index of the target patient to obtain the target voiceprint feature vector includes: When the deviation distance is greater than the distance threshold, the difference between the registered voiceprint feature vector and the reference feature vector of the target patient is analyzed to obtain an additional deviation vector. The additional deviation vector is used to characterize the degree of additional deviation of the voiceprint feature of the target patient relative to the inherent decay pattern of the target patient. The deviation distance is used to characterize the degree of deviation between the voiceprint feature of the original speech segment and the voiceprint feature of the target patient at the time of registration. The deviation distance is obtained based on the difference analysis between the original speech segment and the reference speech segment. Based on the initial voiceprint feature vector, the additional deviation vector, the decay index, and the preset compensation matrix, the initial voiceprint feature vector corresponding to the original speech segment is compensated to obtain the target voiceprint feature vector.

[0060] A deviation distance greater than the distance threshold means that the target patient's current pronunciation voiceprint features deviate from the registered voiceprint beyond the target patient's historical normal fluctuation range. In other words, the voiceprint decay is significant enough to affect the reliability of voiceprint recognition. Therefore, the initial voiceprint feature vector needs to be compensated to eliminate the impact of this additional deviation on the recognition result, resulting in a corrected target voiceprint feature vector. Conversely, if the deviation distance is less than or equal to the distance threshold, it means that the deviation of the current pronunciation voiceprint features is still within the target patient's historical normal fluctuation range. In this case, there is no need to compensate the initial voiceprint feature vector; it can be directly used as the target voiceprint feature vector for subsequent voiceprint comparison.

[0061] The controller can use the x-vector voiceprint feature extraction network to extract features from the original speech segment of the target patient to obtain an initial voiceprint feature vector. It can be understood that the reference feature vector of the target patient has the same dimension as the initial voiceprint feature vector corresponding to the original speech segment.

[0062] For example, the target voiceprint feature vector It can be represented as: In the formula, This is the initial voiceprint feature vector. This is the additional deviation vector. This is the reference feature vector for the target patient. Register the voiceprint feature vector. The decay index. The preset compensation matrix is ​​calculated offline by the controller based on the registered voiceprint feature vector and reference feature vector of the target patient. The registered voiceprint feature vector is a speech segment recorded by the target patient in a quiet environment during their initial medical record-keeping, which is then extracted by the x-vector voiceprint feature extraction network and pre-stored on the server. The preset compensation matrix is ​​a diagonal matrix, with all off-diagonal elements set to 0. For example, each element on the diagonal of the preset compensation matrix... It can be represented as: For example, This is the value of the d-th dimension in the registered voiceprint feature vector. This refers to the value of the d-th dimension in the reference feature vector. The d-th dimension standard deviation is pre-stored for the controller. This d-th dimension standard deviation is calculated based on the difference between the initial voiceprint feature vector and the reference feature vector in the d-th dimension of each successful check-in record of the target patient. It is used to characterize the fluctuation range of the target patient's voiceprint features in the d-th feature dimension. To prevent extremely small positive numbers with a denominator of 0, for example, the The value is 10 －8 .

[0063] In one example, the steps to obtain the deviation distance include: Feature extraction is performed on the original speech segments of the target patient to obtain the initial voiceprint feature vector; The cosine distance between the initial voiceprint feature vector and the registered voiceprint feature vector is calculated to obtain the deviation distance. The registered voiceprint feature vector is used to indicate the feature vector obtained by feature extraction of the reference speech segment when the target patient registers in the department to be reported.

[0064] It is understandable that the registered voiceprint feature vector has the same dimension as the initial voiceprint feature vector corresponding to the original speech segment.

[0065] For example, deviation distance It can be represented as: In the formula, This is the initial voiceprint feature vector. This is the magnitude of the initial voiceprint feature vector. To register voiceprint feature vectors, The magnitude of the registered voiceprint feature vector.

[0066] In one example, the steps to obtain the distance threshold include: The historical average deviation distance is calculated by taking the mean of the historical deviation distances at each time of the target patient’s historical successful check-in records. The standard deviation of historical deviation is calculated based on the historical deviation distance at each time of the target patient’s historical successful check-in record. Data analysis is performed based on historical average deviation distance, historical standard deviation, and decay index to obtain a distance threshold, wherein the distance threshold is positively correlated with the historical average deviation distance, the distance threshold is positively correlated with the historical standard deviation, and the distance threshold is positively correlated with the decay index.

[0067] The server stores the target patient's historical successful check-in records. The controller can access the server to obtain the target patient's historical successful check-in records. These historical successful check-in records are used to indicate all the historical successful check-in records of the target patient in all departments. They are not limited to a specific department, nor are they limited to the department to be checked in at the moment.

[0068] The target patient's historical successful check-in records show that after each check-in, the controller calculates and uploads the historical deviation distance to the server based on the initial voiceprint feature vector of that check-in and the target patient's registered voiceprint feature vector.

[0069] For example, distance threshold It can be represented as: In the formula, This represents the historical average deviation. The decay index. This represents the historical deviation from the standard deviation. A dynamic adjustment coefficient is pre-stored in the controller to control the influence of the decay index on the distance threshold. This dynamic adjustment coefficient is, for example, 0.35. Alternatively, the controller can obtain the historical deviation distance and decay index for each successful reporting record, and perform linear fitting based on the historical deviation distance and decay index values ​​for each reporting record to obtain a linear fitting expression. The slope of this linear fitting expression is the dynamic adjustment coefficient.

[0070] By calculating the distance threshold based on the target patient's historical data and introducing a decay index to dynamically adjust the threshold according to the degree of degradation of the current pronunciation, it is possible to accurately determine whether the voiceprint deviation of the target patient's current pronunciation exceeds the normal voice fluctuation range of the target patient. This improves the accuracy of compensation triggering decisions, avoids triggering unnecessary compensation operations when no abnormal decay occurs, and enables timely triggering of compensation when significant decay occurs, thereby improving the effectiveness of voiceprint feature correction.

[0071] In this embodiment of the invention, if the number of successful check-in records of the target patient is less than a threshold (e.g., 10 times), the controller can obtain the target patient's age from the server and retrieve the historical successful check-in records of people in the same age group as the target patient. Based on these historical successful check-in records, the controller obtains the mean and standard deviation of the historical deviation distance. Then, using the aforementioned distance threshold calculation formula, the controller calculates the distance threshold based on these mean and standard deviation. This leverages the similarity in vocal physiological characteristics among people in the same age group. When historical data for the target patient is insufficient, historical data from people in the same age group can be used as a substitute, ensuring reliable data support for calculating the distance threshold and preventing threshold failure due to data sparsity.

[0072] Step S4: Compare the target voiceprint feature vector with the registered voiceprint feature vector to enable voice clinic check-in.

[0073] The controller calculates the cosine similarity between the target voiceprint feature vector and the registered voiceprint feature vector. If the cosine similarity is greater than or equal to the preset confirmation threshold (e.g., 0.65), the voiceprint comparison is confirmed to be successful, i.e., the registration is successful. If the cosine similarity is less than the preset confirmation threshold, the voiceprint comparison is confirmed to be unsuccessful, i.e., the registration fails.

[0074] In summary, this invention performs similarity analysis on adjacent speech frames of differing segments in the original speech segment of the target patient to obtain multiple similarity values. Information entropy is then calculated based on these similarity values ​​to obtain a decay index. Based on the target patient's registered voiceprint feature vector, reference feature vector, and decay index, the initial voiceprint feature vector corresponding to the original speech segment is compensated to obtain a target voiceprint feature vector. Subsequently, voiceprint comparison is performed between the target voiceprint feature vector and the registered voiceprint feature vector for voice-based outpatient registration. This invention considers the impact of physiological voice decay in elderly patients on voiceprint features. By compensating the initial voiceprint feature vector to obtain the target voiceprint feature vector, voiceprint comparison is performed based on the target voiceprint feature vector. This avoids the problem in related technologies where fixed thresholds cannot adapt to the fluctuations in the voiceprint features of elderly patients, leading to increased risks of false rejection or false approval. This invention improves the accuracy of voiceprint feature capture for elderly patients.

[0075] It should be noted that the processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired result. In some implementations, multitasking and parallel processing are possible or may be advantageous.

Claims

1. A method for capturing the voiceprint features of elderly patients reporting for outpatient services via voice, characterized in that, The method includes: Similarity analysis is performed on adjacent speech frames of the difference segments in the original speech segment of the target patient to obtain multiple similarity values. The similarity value is used to characterize the degree of similarity of acoustic features between corresponding adjacent speech frames. The difference segments are used to characterize speech segments in the original speech segment whose acoustic features have decayed. The difference segments are obtained based on the analysis of the original speech segment and the standard speech template corresponding to the department to be reported to. The department to be reported to is used to indicate the department that the target patient should report to. The department to be reported to is determined by matching the original speech segment and multiple standard speech templates. Information entropy is calculated based on multiple similarity values ​​to obtain the decay index, which is used to characterize the severity of pronunciation degradation of the original speech segment. Based on the registered voiceprint feature vector, reference feature vector, and decay index of the target patient, the initial voiceprint feature vector corresponding to the original speech segment is compensated to obtain the target voiceprint feature vector. The reference feature vector is obtained by feature extraction based on the reference speech segment recorded when the target patient pronounces the speech in a non-standard reading manner. The registered voiceprint feature vector is used to indicate the voiceprint feature vector of the baseline speech segment registered by the target patient during the first visit and registration. Voiceprint comparison is performed between the target voiceprint feature vector and the registered voiceprint feature vector to enable voice-based clinic check-in.

2. The method for capturing the voiceprint features of elderly patients reporting for outpatient services according to claim 1, characterized in that, The steps to obtain the difference fragments include: The DTW algorithm is used to calculate the distance frame by frame between the original speech segment of the target patient and the standard speech template corresponding to the department to be reported to, and obtain multiple local distance values. The number of local distance values ​​is positively correlated with the number of frames of the original speech segment of the target patient. The multiple local distance values ​​are arranged in temporal order according to the corresponding frames in the original speech segment to obtain a difference sequence; The percentile method is used to extract the threshold of the difference sequence to obtain the difference threshold; By analyzing the relative magnitudes of multiple local distance values ​​and the difference threshold in the difference sequence, multiple candidate distance values ​​are obtained, among which the candidate distance values ​​are greater than the difference threshold. The original speech segments corresponding to temporally adjacent frames among the multiple candidate distance values ​​are identified as the difference segments.

3. The method for capturing the voiceprint features of elderly patients reporting for outpatient services according to claim 2, characterized in that, The similarity analysis of adjacent speech frames of the differing segments in the original speech segment of the target patient yields multiple similarity values, including: Feature extraction is performed on each frame of the difference segments in the original speech of the target patient to obtain the feature vector corresponding to each frame; For each pair of adjacent frames in the difference segment, the cosine similarity is calculated based on the feature vectors corresponding to the adjacent frames to obtain the similarity value of each pair of adjacent frames. Among them, multiple similarity values ​​are the similarity values ​​of multiple pairs of adjacent frames.

4. The method for capturing the voiceprint features of elderly patients reporting for outpatient services according to claim 3, characterized in that, The decay index is obtained by calculating information entropy based on multiple similarity values, including: Based on dividing the value range of multiple similar values ​​into multiple preset intervals, the number of the multiple similar values ​​falling into each of the preset intervals is counted; The information entropy is calculated based on the number of items falling into each of the preset intervals, and the decay index is obtained.

5. The method for capturing the voiceprint features of elderly patients reporting for outpatient services according to claim 1, characterized in that, The initial voiceprint feature vector corresponding to the original speech segment is compensated based on the registered voiceprint feature vector, reference feature vector, and decay index of the target patient to obtain the target voiceprint feature vector, including: When the deviation distance is greater than the distance threshold, the difference between the registered voiceprint feature vector and the reference feature vector of the target patient is analyzed to obtain an additional deviation vector. The additional deviation vector is used to characterize the degree of additional deviation of the voiceprint feature of the target patient relative to the inherent decay pattern of the target patient. The deviation distance is used to characterize the degree of deviation between the voiceprint feature of the original speech segment and the voiceprint feature of the target patient at the time of registration. The deviation distance is obtained based on the difference analysis between the original speech segment and the reference speech segment. Based on the initial voiceprint feature vector, the additional deviation vector, the decay index, and the preset compensation matrix, the initial voiceprint feature vector corresponding to the original speech segment is compensated to obtain the target voiceprint feature vector.

6. The method for capturing the voiceprint features of elderly patients reporting for outpatient services according to claim 5, characterized in that, The steps to obtain the deviation distance include: Feature extraction is performed on the original speech segments of the target patient to obtain the initial voiceprint feature vector; The cosine distance between the initial voiceprint feature vector and the registered voiceprint feature vector is calculated to obtain the deviation distance. The registered voiceprint feature vector is used to indicate the feature vector obtained by feature extraction of the reference speech segment when the target patient registers in the department to be reported.

7. The method for capturing the voiceprint features of elderly patients reporting for outpatient services according to claim 5, characterized in that, The steps to obtain the distance threshold include: The historical average deviation distance is calculated by taking the mean of the historical deviation distances at each time of the target patient’s historical successful check-in records. The standard deviation of historical deviation is calculated based on the historical deviation distance at each time of the target patient’s historical successful check-in record. Data analysis is performed based on historical average deviation distance, historical standard deviation, and decay index to obtain a distance threshold, wherein the distance threshold is positively correlated with the historical average deviation distance, the distance threshold is positively correlated with the historical standard deviation, and the distance threshold is positively correlated with the decay index.

8. The method for capturing the voiceprint features of elderly patients reporting to a voice clinic according to claim 1, characterized in that, The steps to obtain the department you are to report to include: The original speech segment of the target patient is matched with the standard speech templates corresponding to each of the multiple booked departments to obtain the matching distance between the original speech segment and each booked department. Each booked department has multiple standard speech templates, which are used to indicate the standard speech segment of the official full name and the standard speech segment of the medical abbreviation of the corresponding department. Based on the matching distance between the original voice segment and multiple pre-booked departments, the department to which the target patient is to report is selected from the multiple pre-booked departments. The matching distance between the original voice segment and the department to which the patient is to report is less than the matching distance between the original voice segment and any other pre-booked department other than the department to which the patient is to report.

9. The method for capturing the voiceprint features of elderly patients reporting for outpatient services according to claim 8, characterized in that, The process of matching the original speech segment of the target patient with the standard speech templates corresponding to each of the multiple booked departments to obtain the matching distance between the original speech segment and each booked department includes: Based on the weighted dynamic time warping matching of multiple standard speech templates corresponding to the candidate departments with the original speech segments of the target patients, the weighted matching distance between the original speech segments and each standard speech template corresponding to the candidate departments is obtained. The candidate department is any one of multiple booked departments. The minimum weighted matching distance between the original speech segment and each standard speech template corresponding to the candidate department is determined as the matching distance between the original speech segment and the candidate department.

10. The method for capturing the voiceprint features of elderly patients reporting to a voice clinic according to claim 8, characterized in that, The step of selecting the target patient's assigned department from multiple booked departments based on the matching distance between the original voice segment and the corresponding departments includes: The relative magnitudes of the matching distances between the original speech segment and multiple booked departments are analyzed to obtain the first matching distance and the second matching distance. The first matching distance is used to indicate the minimum value among the matching distances between the original speech segment and multiple booked departments, and the second matching distance is used to indicate the second minimum value among the matching distances between the original speech segment and multiple booked departments. Calculate the ratio of the second matching distance to the first matching distance to obtain the confidence interval ratio; If the confidence interval ratio is greater than or equal to the ratio threshold, the department with the appointment corresponding to the first matching distance is determined as the department to be visited by the target patient.

Images