AI automatic call processing method and system

Abstract

The present invention discloses an AI automatic call processing method, including collecting audio data during the AI ​​automatic call, obtaining each call voice frame, obtaining each voice modal component, obtaining a sentence disconnection peak point and a pronunciation time series, calculating a call fluctuation discontinuity index of each voice modal component based on the number of sentence disconnection peak points in each voice modal component and the volatility of each pronunciation time series, obtaining a user voice component based on the call fluctuation discontinuity index of all voice modal components, obtaining an effective prediction factor based on the similarity between each pronunciation time series in the user voice component and the sentence disconnection peak point, determining the number of prediction reference items, obtaining call prediction audio data, and determining an AI real-time reply text sequence. An AI automatic call processing method provided by an embodiment of the present invention improves the real-time performance of AI automatic call processing while ensuring the accuracy of AI automatic call processing.

Description

Technical Field

[0001] The present invention relates to the field of call technology, and in particular to an AI automatic call processing method and system. Background Art

[0002] Transmitting information through voice is the most important, effective, commonly used and convenient form of information exchange for humans. Voice calls make communication more vivid and powerful through factors such as tone, audio and speaking speed. Whether it is a business meeting, telephone interview or other forms of communication, voice calls do not require a lot of time and energy, and can express meaning and emotions more accurately and promptly, eliminate misunderstandings, and help users better feel the other person's emotions and tone, further enhancing mutual trust and closeness.

[0003] AI automatic dialing is not restricted by time and can provide services around the clock, handling a large number of incoming and outbound calls, ensuring that customers can get a response at any time, significantly reducing the demand for manual seats and lowering labor costs. It is widely used in corporate customer service, market research, sales follow-up and other fields.

[0004] Existing speech recognition technology converts the input speech into sound signals, performs a series of signal processing, frame segmentation, acoustic feature extraction, and then compares and matches with the stored acoustic model and language model to obtain the recognition result. However, due to the high time complexity of matching, it is difficult to take into account the real-time and accuracy of customer call content recognition during AI automatic call making.

[0005] It can be seen that how to solve the problem of ensuring the accuracy of AI automatic call processing while improving the real-time processing has become a technical problem that technical personnel in this field need to solve urgently. Summary of the invention

[0006] The present invention provides an AI automatic call processing method and system to solve the technical problem of ensuring the accuracy of AI automatic call processing while improving the real-time processing, so as to achieve accurate and real-time recognition of customer call content during the AI ​​automatic call processing.

[0007] In order to solve the above technical problems, an embodiment of the present invention provides an AI automatic call processing method and system.

[0008] Collect audio data during AI automatic call making, obtain each call voice frame, and use the modal decomposition algorithm to decompose each call voice frame into multiple voice modal components;

[0009] Acquire sentence disconnection peak points and pronunciation time series according to each speech modal component, and calculate the call fluctuation discontinuity index of each speech modal component based on the number of sentence disconnection peak points in each speech modal component and the volatility of each pronunciation time series;

[0010] The user voice component is obtained based on the call fluctuation intermittency index of all voice modal components, the effective prediction factor is obtained based on the similarity between the pronunciation time series in the user voice component and the peak point of sentence disconnection, and the number of prediction reference items is obtained according to the effective prediction factor and the call fluctuation intermittency index of the user voice component;

[0011] Based on each call voice frame and the number of prediction reference items, call prediction audio data is obtained to determine the AI ​​real-time reply text sequence.

[0012] As one preferred solution, the method of obtaining the sentence disconnection peak point and pronunciation time series includes:

[0013] Extract all peak points in each speech modal component using a peak search algorithm, and obtain the sentence disconnection peak point of each speech modal component based on the difference in time intervals between adjacent peak points in each speech modal component;

[0014] The midpoint between each sentence disconnection peak point and the next adjacent peak point in each speech modal component is recorded as the sentence segmentation peak point. All sentence disconnection peak points in each speech modal component are used as division points to divide each speech modal component into multiple pronunciation time series.

[0015] As one preferred solution, the step of obtaining the sentence disconnection peak point of each speech modal component includes:

[0016] Acquire peak clusters based on the time intervals between adjacent peak points in each speech modal component;

[0017] The average value of the sentence neighborhood disconnection coefficients of all peak points in each peak cluster is taken as the peak disconnection index of each peak cluster, and all peak points of the peak cluster with the largest peak disconnection index of each speech modal component are recorded as the sentence disconnection peak points of each speech modal component.

[0018] As one preferred solution, obtaining peak value clusters includes:

[0019] For each peak point in each speech modal component, the time interval between each peak point and the next adjacent peak point is taken as the sentence neighborhood disconnection coefficient of each peak point. The sentence neighborhood disconnection coefficient of all peak points in each speech modal component is taken as the input of the clustering algorithm. The clustering algorithm is used to divide all peak points in each speech modal component into two peak clustering clusters.

[0020] As one of the preferred solutions, the method for calculating the call fluctuation interruption index of each speech modal component is:

[0021] Calculate the irregular fluctuation discrimination index of each speech modal component based on the fluctuation of each pronunciation time series of each speech modal component;

[0022] The call fluctuation discontinuity index of each speech modal component is positively correlated with the irregular fluctuation discrimination index and the number of corresponding sentence disconnection peak points.

[0023] As one of the preferred solutions, the calculation method of the irregular fluctuation discrimination index of each speech modal component is:

[0024] The range of each speech modal component and the time lag mutual correlation coefficient at each lag order are calculated, and the maximum value of the time lag mutual correlation coefficient of each speech modal component at all lag orders is recorded as the audio fluctuation autocorrelation coefficient of each speech modal component. The irregular fluctuation discrimination index of each speech modal component is positively correlated with the corresponding range and negatively correlated with the corresponding audio fluctuation autocorrelation coefficient.

[0025] As one of the preferred solutions, the specific method of obtaining the user voice component includes:

[0026] The call fluctuation interruption indexes of all voice modal components are arranged in descending order, and the voice modal component corresponding to the call fluctuation interruption index with the largest value is taken as the user voice component.

[0027] As one preferred solution, obtaining effective prediction factors includes:

[0028] The sum of the time intervals between all the sentence disconnection peak points and the next adjacent peak points in the user voice component is divided by the length of the user voice component as the user invalid audio coverage rate;

[0029] The pronunciation time series of the user's voice components are combined in pairs, and the DTW distance between the pronunciation time series of each combination is used as the pronunciation variation index of each combination. The effective prediction factor is negatively correlated with the pronunciation variation index of all combinations and the invalid audio coverage rate of the user.

[0030] As one preferred solution, obtaining the number of prediction reference items includes:

[0031] The call prediction variation factor is positively correlated with the call fluctuation intermittency index of the user voice component and negatively correlated with the effective prediction factor;

[0032] The reference item adjustment coefficient is the normalized value of the call prediction variation factor, and the number of predicted reference items is the integer value of the product of the reference item adjustment coefficient and the preset maximum number of reference items.

[0033] As one of the preferred solutions, obtaining call prediction audio data and determining the AI ​​real-time reply text sequence includes:

[0034] The number of prediction reference items is used as the number of call speech frames participating in the prediction in the linear prediction coding algorithm, and the linear prediction coding algorithm is used to predict the next call speech frame to obtain call prediction audio data;

[0035] The language model is used to convert the call prediction audio data into a predicted text sequence, and the predicted text sequence is matched with the human text sequence in the AI ​​dialogue database. The reply text sequence corresponding to the human text sequence with the highest matching degree is used as the AI ​​pre-reply text sequence. When the sequence similarity between the next call voice frame and the call prediction audio data is greater than a preset threshold, the AI ​​pre-reply text sequence is used as the real-time reply text sequence. Otherwise, the language model is used to convert the next call voice frame into a real-time text sequence, and the real-time text sequence is matched with the human text sequence in the AI ​​dialogue database. The reply text sequence corresponding to the human text sequence with the highest matching degree is used as the AI ​​real-time reply text sequence.

[0036] Compared with the prior art, the embodiments of the present invention have the following advantages:

[0037] (1) Collect audio data during AI automatic call making, obtain each call voice frame, obtain each voice modal component based on each call voice frame, comprehensively consider the discontinuity and volatility of each voice modal component of each call voice frame, calculate the call fluctuation discontinuity index of each voice modal component, and use it to measure the possibility that each voice modal component is human user audio, obtain the user voice component, extract the audio data of the human user, and subsequently analyze only the user voice component, thereby improving the real-time performance of AI automatic call making processing;

[0038] (2) The dissimilarity between the audio segments of human users and the proportion of invalid audio are analyzed. The effective prediction factor is obtained based on the similarity between the pronunciation time series in the user's voice component and the peak point of sentence disconnection. The number of prediction reference items is obtained by combining the call fluctuation intermittent index of the user's voice component. The difficulty and accuracy of predicting the audio during the AI ​​automatic dialing process based on historical audio data are comprehensively considered, thereby improving the reliability of determining the number of prediction reference items.

[0039] (3) Based on the call voice frames and the number of prediction reference items, the call prediction audio data is obtained to determine the AI ​​real-time reply text sequence, which ensures the accuracy of the AI ​​automatic call processing while further improving the real-time performance of the AI ​​automatic call processing. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 This is a flowchart of AI automatic call processing in one embodiment of the present invention;

[0041] Figure 2 is a schematic diagram of a peak point of sentence disconnection in one embodiment of the present invention;

[0042] Figure 3 It is a schematic diagram of obtaining the predicted number of reference items in one embodiment of the present invention. DETAILED DESCRIPTION

[0043] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. The purpose of providing these embodiments is to make the disclosure of the present invention more thorough and comprehensive. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0044] In the description of this application, the terms "first", "second", "third", etc. are used for descriptive purposes only and are not to be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Thus, a feature defined as "first", "second", "third", etc. may explicitly or implicitly include one or more of the feature. In the description of this application, unless otherwise specified, "plurality" means two or more.

[0045] In the description of the present application, it should be noted that, unless otherwise clearly specified and limited, the terms "installed", "connected", and "connected" should be understood in a broad sense. For example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection, or it can be indirectly connected through an intermediate medium, or it can be a connection between the two elements. The terms "vertical", "horizontal", "left", "right", "upper", "lower" and similar expressions used herein are only for illustrative purposes, and do not indicate or imply that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as a limitation on the present invention. The term "and / or" used herein includes any and all combinations of one or more related listed items. For those of ordinary skill in the art, the specific meanings of the above terms in this application can be understood according to specific circumstances.

[0046] In the description of this application, it should be noted that, unless otherwise defined, all technical and scientific terms used in the present invention have the same meaning as those commonly understood by those skilled in the art. The terms used in the specification of the present invention are only for the purpose of describing specific embodiments and are not intended to limit the present invention. For those of ordinary skill in the art, the specific meanings of the above terms in this application can be understood by specific circumstances.

[0047] An embodiment of the present invention provides an AI automatic call processing method. For details, see Figure 1 , Figure 1 The figure shows a flow chart of AI automatic call dialing process in one embodiment of the present invention.

[0048] Step S1: Collect audio data during the AI ​​automatic call and obtain each call voice frame.

[0049] In order to identify the audio data of human users during the process of AI automatically making calls, the audio data of the call is recorded during each AI automatically making call as call voice data.

[0050] Since the environment in which human users live is relatively complex, the quality of voice calls will inevitably be affected by the external environment, noise introduced by the transmission medium, and internal noise of the communication equipment. Conversation sounds, wind sounds, echoes, current sounds, etc. may appear in the call voice data, which greatly interferes with the recognition of the content of human users' calls. Therefore, before recognizing the content of human users' calls, the call voice data is first denoised, pre-emphasized, framed, and windowed to obtain each call voice frame.

[0051] It should be noted that the length of each call voice frame is a preset value. In this embodiment, the length of each call voice frame is 20ms, and the window function of the window is a Hamming window. The length of each call voice frame and the selection of the window function are other implementation methods that the implementer can choose at his own discretion, and this application does not impose any special restrictions on this.

[0052] Step S2: Acquire each speech modal component according to each call speech frame, acquire the sentence disconnection peak point and the pronunciation time series according to each speech modal component, and calculate the call fluctuation discontinuity index of each speech modal component based on the number of sentence disconnection peak points in each speech modal component and the volatility of each pronunciation time series.

[0053] Since the two parties in an AI automatic call are an artificial intelligence system and a human user, the audio data at different time stages during the call may belong to the human user's audio data, or it may belong to artificial intelligence audio data, silence, background noise and other interference information. In order to extract the human user's audio data first and improve the real-time performance of subsequent AI automatic call processing, a modal decomposition algorithm is used to decompose each call voice frame into multiple voice modal components.

[0054] It should be noted that the modal decomposition algorithm is a well-known technology and will not be elaborated in this application. This embodiment selects the EMD empirical mode decomposition algorithm to decompose each call voice frame. The selection of the modal decomposition algorithm is other implementation methods that the implementer can choose at his own discretion, and this application does not impose any special restrictions on this.

[0055] It should be further explained that during the AI ​​automatic call process, the artificial intelligence and the human user complete the call through a dialogue, and the audio data of the artificial intelligence is often the sound generated by a computer program or algorithm. Therefore, the audio data of the human user is more intermittent than the background noise, and has stronger emotional fluctuations than the audio data of the artificial intelligence.

[0056] Since there may be multiple adjacent peaks in the audio data of the same sentence, the discontinuity of each speech modal component of each call voice frame is first analyzed, and all peak points in each speech modal component are extracted using a peak search algorithm. For each peak point in each speech modal component, the time interval between each peak point and the next adjacent peak point is used as the sentence neighborhood disconnection coefficient of each peak point. The sentence neighborhood disconnection coefficient of all peak points in each speech modal component is used as the input of the clustering algorithm, and the clustering algorithm is used to divide all peak points in each speech modal component into two peak clustering clusters.

[0057] Furthermore, the average value of the sentence neighborhood disconnection coefficients of all peak points in each peak cluster is taken as the peak disconnection index of each peak cluster, and all peak points of the peak cluster with the largest peak disconnection index of each speech modal component are recorded as the sentence disconnection peak points of each speech modal component.

[0058] It should be noted that, when the sentence neighborhood disconnection coefficient of the peak point is larger, it means that the peak point is more likely to be the last peak point of a segment of audio data, the larger the peak disconnection index value of the peak clustering cluster, the more suitable the peak points in the peak clustering cluster are for dividing the speech modal components into multiple audio segments.

[0059] The midpoint between each sentence disconnection peak point and the next adjacent peak point in each speech modal component is recorded as the sentence segmentation peak point. All sentence disconnection peak points in each speech modal component are used as division points to divide each speech modal component into multiple pronunciation time series.

[0060] It should be noted that the peak finding algorithm, clustering algorithm and the calculation of the average value are well-known technologies. In this embodiment, the AMPD automatic multi-scale peak finding algorithm is selected to extract the peak points in the speech modal component, and the k-means clustering algorithm is selected to cluster the peak points in the speech modal component. In this embodiment, the necessary parameter clustering number in the k-means clustering algorithm is set to 2. For the selection of the peak finding algorithm and the clustering algorithm, as other implementation methods, the implementer can choose them at his own discretion, and this application does not impose any special restrictions on this.

[0061] A schematic diagram of a sentence disconnection peak point in one embodiment of the present invention is as follows Figure 2 As shown, the dots represent peak points and the squares represent sentence disconnection peak points.

[0062] Then the volatility of each pronunciation time series is analyzed, and the range of each speech modal component and the time lag mutual correlation coefficient at each lag order are calculated. The maximum value of the time lag mutual correlation coefficient of each speech modal component at all lag orders is recorded as the audio fluctuation autocorrelation coefficient of each speech modal component. The irregular fluctuation discrimination index of each speech modal component is positively correlated with the corresponding range and negatively correlated with the corresponding audio fluctuation autocorrelation coefficient.

[0063] The call fluctuation discontinuity index of each speech modal component is positively correlated with the irregular fluctuation discrimination index and the number of corresponding sentence disconnection peak points.

[0064] It should be noted that a negative correlation refers to the opposite directions of change of the two variables, which can be division, subtraction, etc. A positive correlation refers to the same directions of change of the two variables, which can be addition, multiplication, etc.

[0065] As an embodiment of the present application, the audio fluctuation autocorrelation coefficient of each speech modal component plus a preset tuning factor is used as the audio fluctuation adjusted autocorrelation coefficient of each speech modal component, and the ratio of the range of each speech modal component to the audio fluctuation adjusted autocorrelation coefficient is used as the irregular fluctuation discrimination index of each speech modal component.

[0066] It should be noted that the preset tuning factor is a manually preset value. In this embodiment, the preset tuning factor is 0.0001. As for the value of the preset tuning factor, as other implementation methods, the implementer can choose it at will, and this application does not impose any special restrictions on this.

[0067] When the range of the speech modal component is larger and the autocorrelation coefficient of the audio fluctuation is smaller, it means that the fluctuation amplitude of the speech modal component is larger, it is more difficult to predict future fluctuations based on historical data, and the irregular fluctuation discrimination index value is larger; when the irregular fluctuation discrimination index of the speech modal component is larger and the number of sentence disconnection peak points is larger, it means that the speech modal component is more likely to be intermittent and irregular audio fluctuations, and it is more likely to be the audio data of human users during AI automatic dialing.

[0068] Step S3: obtaining the user voice component based on the call fluctuation intermittency index of all voice modal components, obtaining the effective prediction factor based on the similarity between the pronunciation time series in the user voice component and the sentence disconnection peak point, and obtaining the number of prediction reference items according to the effective prediction factor and the call fluctuation intermittency index of the user voice component.

[0069] Taking into account the discontinuity and volatility of each voice modal component, the call fluctuation discontinuity indexes of all voice modal components are arranged in descending order, and the voice modal component corresponding to the call fluctuation discontinuity index with the largest value is taken as the user voice component.

[0070] The sum of the time intervals between all sentence disconnection peak points and the next adjacent peak points in the user voice component is divided by the length of the user voice component as the user invalid audio coverage rate. The pronunciation time series of the user voice components are combined in pairs, and the DTW distance between the pronunciation time series of each combination is used as the pronunciation variation index of each combination. The effective prediction factor is negatively correlated with the pronunciation variation index of all combinations and the user invalid audio coverage rate.

[0071] As an embodiment of the present application, the average value of the pronunciation variation index of all combinations composed of the pronunciation time series of the user's voice components is taken as the user's pronunciation variation factor, the product of the user's invalid audio coverage and the user's pronunciation variation factor plus the preset parameter adjustment factor is taken as the prediction blocking adjustment coefficient, and the inverse of the prediction blocking adjustment coefficient is taken as the effective prediction factor.

[0072] It should be noted that the preset tuning factor is a manually preset value. In this embodiment, the preset tuning factor is 0.0001. As for the value of the preset tuning factor, as other implementation methods, the implementer can choose it at will, and this application does not impose any special restrictions on this.

[0073] It should be further explained that there is invalid audio between the sentence disconnection peak point and the next adjacent peak point. When the time interval between each sentence disconnection peak point and the next adjacent peak point in the user voice component accounts for a larger proportion, it means that the proportion of invalid audio in the user voice component is larger, and the user invalid audio coverage value is larger, and it is more difficult to predict the audio during the AI ​​automatic dialing process based on historical audio data, and the effective prediction factor value is smaller; when the DTW distance between each pronunciation time series of the user voice component is larger, it means that the difference between each audio segment of the human user is higher, and the pronunciation variation index value is larger, and it is more difficult to predict the audio during the AI ​​automatic dialing process based on historical audio data, and the effective prediction factor value is smaller.

[0074] Furthermore, the call prediction variation factor is positively correlated with the call fluctuation interruption index of the user voice component and negatively correlated with the effective prediction factor.

[0075] As an embodiment of the present application, the effective prediction factor plus the preset parameter adjustment factor is used as the effective adjustment prediction factor, and the ratio of the call fluctuation interruption index of the user voice component to the effective adjustment prediction factor is used as the call prediction variation factor.

[0076] It should be noted that when the call fluctuation intermittency index of the user voice component is larger and the effective prediction factor is smaller, it means that the volatility and intermittency of the user audio during the AI ​​automatic dialing process are higher and the regularity is lower. It is more difficult to predict the audio during the AI ​​automatic dialing process based on historical audio data, and the call prediction variation factor value is larger.

[0077] The reference item adjustment coefficient is the normalized value of the call prediction variation factor, and the number of predicted reference items is the integer value of the product of the reference item adjustment coefficient and the preset maximum number of reference items.

[0078] It should be noted that the preset parameter adjustment factor and the preset maximum number of reference items are both artificially preset values. In this embodiment, the preset parameter adjustment factor is 0.0001, and the preset maximum number of reference items is the number of call voice frames. As for the values ​​of the preset parameter adjustment factor and the preset maximum number of reference items, as other implementation methods, the implementer can select them at will, and this application does not impose any special restrictions on this. In one embodiment of the present invention, the schematic diagram of obtaining the predicted reference number is as shown in Figure 3 shown.

[0079] Step S4: Acquire call prediction audio data based on each call voice frame and the number of prediction reference items, and determine the AI ​​real-time reply text sequence.

[0080] The acoustic model in the speech recognition system is responsible for identifying sound features and mapping them to phonemes, while the language model is used to determine the most likely word sequence. Phonemes are the smallest speech units divided according to the natural properties of speech. They are analyzed based on the pronunciation actions in the syllable, and one action constitutes a phoneme. Therefore, by predicting the phonemes during the AI ​​automatic dialing process, the mapping efficiency of the acoustic model can be improved, ensuring the accuracy of the AI ​​automatic dialing process while improving the real-time processing.

[0081] Specifically, the number of prediction reference items is used as the number of call speech frames involved in the prediction in the linear prediction coding algorithm, and the linear prediction coding algorithm is used to predict the next call speech frame to obtain the call prediction audio data.

[0082] The language model is used to convert the call prediction audio data into a predicted text sequence, and the predicted text sequence is matched with the human text sequence in the AI ​​dialogue database. The reply text sequence corresponding to the human text sequence with the highest matching degree is used as the AI ​​pre-reply text sequence. When the sequence similarity between the next call voice frame and the call prediction audio data is greater than a preset threshold, the AI ​​pre-reply text sequence is used as the real-time reply text sequence. Otherwise, the language model is used to convert the next call voice frame into a real-time text sequence, and the real-time text sequence is matched with the human text sequence in the AI ​​dialogue database. The reply text sequence corresponding to the human text sequence with the highest matching degree is used as the AI ​​real-time reply text sequence.

[0083] It should be noted that the language model and the AI ​​dialogue database are constructed by professionals in the field, and each human text sequence in the AI ​​dialogue database corresponds to a reply text sequence. The calculation of sequence similarity is a well-known technology, and the preset threshold is a manually preset value. In this embodiment, the preset threshold value is 0.5, and the Pearson correlation coefficient is selected for the calculation of sequence similarity. As for the value of the preset threshold and the calculation of sequence similarity, as other implementation methods, the implementer can select them at his own discretion, and this application does not impose any special restrictions on this.

[0084] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiments can be implemented by instructing the relevant hardware through a computer program, and the program can be stored in a computer-readable storage medium, and when the program is executed, it can include the processes of the embodiments of the above-mentioned methods. The storage medium can be a disk, an optical disk, a read-only memory (ROM) or a random access memory (RAM), etc.

[0085] Accordingly, an embodiment of the present invention provides a computer-readable storage medium, wherein the computer-readable storage medium includes a stored computer program, wherein when the computer program is executed, the device where the computer-readable storage medium is located is controlled to perform the steps in the AI ​​automatic call processing method of the above embodiment, for example Figure 1 Steps S1 to S4 described in .

[0086] The above-mentioned embodiments only express several implementation methods of the present invention, and the description thereof is relatively specific and detailed, but it cannot be understood as limiting the scope of the patent of the present invention. It should be pointed out that, for ordinary technicians in this field, several variations and improvements can be made without departing from the concept of the present invention, which all belong to the protection scope of the present invention. Therefore, the protection scope of the patent of the present invention shall be subject to the attached claims.

Claims

1. An AI automatic call processing method, characterized in that: include: Collect audio data during AI automatic call making, obtain each call voice frame, and use the modal decomposition algorithm to decompose each call voice frame into multiple voice modal components; Acquire sentence disconnection peak points and pronunciation time series according to each speech modal component, and calculate the call fluctuation discontinuity index of each speech modal component based on the number of sentence disconnection peak points in each speech modal component and the volatility of each pronunciation time series; The user voice component is obtained based on the call fluctuation intermittency index of all voice modal components, the effective prediction factor is obtained based on the similarity between the pronunciation time series in the user voice component and the peak point of sentence disconnection, and the number of prediction reference items is obtained according to the effective prediction factor and the call fluctuation intermittency index of the user voice component; Acquire call prediction audio data based on each call voice frame and the number of prediction reference items, and determine the AI ​​real-time reply text sequence; The method of obtaining sentence disconnection peak points and pronunciation time series includes: using a peak search algorithm to extract all peak points in each speech modal component, and obtaining the sentence disconnection peak points of each speech modal component based on the difference in time intervals between adjacent peak points in each speech modal component; recording the midpoint between each sentence disconnection peak point and the next adjacent peak point in each speech modal component as a sentence segmentation peak point, and dividing each speech modal component into multiple pronunciation time series using all sentence disconnection peak points in each speech modal component as division points.

2. The AI ​​automatic call processing method according to claim 1, characterized in that: The obtaining of the sentence disconnection peak point of each speech modal component includes: Acquire peak clusters based on the time intervals between adjacent peak points in each speech modal component; The average value of the sentence neighborhood disconnection coefficients of all peak points in each peak cluster is taken as the peak disconnection index of each peak cluster, and all peak points of the peak cluster with the largest peak disconnection index of each speech modal component are recorded as the sentence disconnection peak points of each speech modal component.

3. The AI ​​automatic call processing method according to claim 2, characterized in that: The obtaining of peak clusters comprises: For each peak point in each speech modal component, the time interval between each peak point and the next adjacent peak point is taken as the sentence neighborhood disconnection coefficient of each peak point. The sentence neighborhood disconnection coefficient of all peak points in each speech modal component is taken as the input of the clustering algorithm. The clustering algorithm is used to divide all peak points in each speech modal component into two peak clustering clusters.

4. The AI ​​automatic call processing method according to claim 1, characterized in that: The calculation method of the call fluctuation interruption index of each voice modal component is: Calculate the irregular fluctuation discrimination index of each speech modal component based on the fluctuation of each pronunciation time series of each speech modal component; The call fluctuation discontinuity index of each speech modal component is positively correlated with the irregular fluctuation discrimination index and the number of corresponding sentence disconnection peak points.

5. The AI ​​automatic call processing method according to claim 4, characterized in that: The calculation method of the irregular fluctuation discrimination index of each speech modal component is: The range of each speech modal component and the time lag mutual correlation coefficient at each lag order are calculated, and the maximum value of the time lag mutual correlation coefficient of each speech modal component at all lag orders is recorded as the audio fluctuation autocorrelation coefficient of each speech modal component. The irregular fluctuation discrimination index of each speech modal component is positively correlated with the corresponding range and negatively correlated with the corresponding audio fluctuation autocorrelation coefficient.

6. The AI ​​automatic call processing method according to claim 1, characterized in that: The specific method of obtaining the user voice component includes: The call fluctuation interruption indexes of all voice modal components are arranged in descending order, and the voice modal component corresponding to the call fluctuation interruption index with the largest value is taken as the user voice component.

7. The AI ​​automatic call processing method according to claim 1, characterized in that: The obtaining of effective predictive factors comprises: The sum of the time intervals between all the sentence disconnection peak points and the next adjacent peak points in the user voice component is divided by the length of the user voice component as the user invalid audio coverage rate; The pronunciation time series of the user's voice components are combined in pairs, and the DTW distance between the pronunciation time series of each combination is used as the pronunciation variation index of each combination. The effective prediction factor is negatively correlated with the pronunciation variation index of all combinations and the invalid audio coverage rate of the user.

8. The AI ​​automatic call processing method according to claim 1, characterized in that: The obtaining of the number of prediction reference items includes: The call prediction variation factor is positively correlated with the call fluctuation intermittency index of the user voice component and negatively correlated with the effective prediction factor; The reference item adjustment coefficient is the normalized value of the call prediction variation factor, and the number of predicted reference items is the integer value of the product of the reference item adjustment coefficient and the preset maximum number of reference items.

9. The AI ​​automatic call processing method according to claim 1, characterized in that: The obtaining of call prediction audio data and determining the AI ​​real-time reply text sequence comprises: The number of prediction reference items is used as the number of call speech frames participating in the prediction in the linear prediction coding algorithm, and the linear prediction coding algorithm is used to predict the next call speech frame to obtain call prediction audio data; The language model is used to convert the call prediction audio data into a predicted text sequence, and the predicted text sequence is matched with the human text sequence in the AI ​​dialogue database. The reply text sequence corresponding to the human text sequence with the highest matching degree is used as the AI ​​pre-reply text sequence. When the sequence similarity between the next call voice frame and the call prediction audio data is greater than a preset threshold, the AI ​​pre-reply text sequence is used as the real-time reply text sequence. Otherwise, the language model is used to convert the next call voice frame into a real-time text sequence, and the real-time text sequence is matched with the human text sequence in the AI ​​dialogue database. The reply text sequence corresponding to the human text sequence with the highest matching degree is used as the AI ​​real-time reply text sequence.

Images