Telephone answering system based on AI

Abstract

The invention discloses an AI-based telephone answering system, and relates to the technical field of telephone answering. Comprising a real-time voice acquisition and preprocessing module, a voice speed detection and prediction module, a self-adaptive ASR dynamic regulation and control module, a refined voice slicing and decoding module, a keyword detection and voice recognition module, a semantic understanding and intention analysis module and an emergency response and scheduling module. And the real-time voice acquisition and preprocessing module is used for acquiring user voice data of the emergency call in real time, establishing a real-time audio stream transmission channel and quickly preprocessing the acquired audio signal. According to the method, the voice features of the user at the high speed are quickly and accurately captured, and the voice segmentation length of the ASR engine is dynamically adjusted, so that each voice segment is clearer and more accurate, cross-segmentation vocabulary segmentation and recognition errors caused by too high speed are avoided, and the user experience is improved. Therefore, omission and misjudgment risks of the emergency information of the user are effectively avoided in an emergency help-seeking scene.

Description

Technical Field

[0001] The present invention relates to the technical field of telephone answering, and in particular to an AI-based telephone answering system. Background Art

[0002] The AI-based telephone answering system is an intelligent interactive system that uses artificial intelligence technology to automatically handle incoming calls. It is widely used in customer service, voice assistants, corporate hotlines and other scenarios. The system usually combines speech recognition (ASR), natural language processing (NLP) and speech synthesis (TTS) technologies to achieve recognition, intent analysis and intelligent response of user voice. It can understand the needs of users and generate accurate answers based on knowledge bases, databases or deep learning models, and even conduct multiple rounds of conversations. In addition, the system can realize personalized services such as business processing, information query, appointment scheduling, etc. through sentiment analysis, context understanding and integration with CRM (customer relationship management) systems. Compared with the traditional IVR (interactive voice response) system, the AI ​​telephone answering system is more intelligent, reduces user keystrokes, improves service efficiency and user experience, and reduces the cost of manual customer service.

[0003] The prior art has the following deficiencies: When users make calls for emergency help, they often speak faster due to anxiety and nervousness. At this time, the AI ​​phone answering system may find it difficult to accurately recognize and understand the user's voice content, and then misjudge or ignore the user's emergency help intention. Once the AI ​​system fails to detect the emergency help signal sent by the user in time, it will not be able to quickly transfer the call to the manual customer service or alarm department, which will cause medical staff, firefighters or other emergency rescue forces to be unable to rush to the scene in the first time, missing the precious golden rescue time, causing irreparable personal injury or serious property loss, or even leading to more serious social security incidents.

[0004] The above information disclosed in this Background section is only for enhancement of understanding of the background of the present disclosure and therefore it may contain information that does not constitute the prior art that is already known to one of ordinary skill in the art. Summary of the invention

[0005] The purpose of the present invention is to provide an AI-based telephone answering system, which can quickly and accurately capture the voice features of users at high speaking speeds and dynamically adjust the voice segment length of the ASR engine to make each voice segment clearer and more accurate, thereby avoiding cross-segment vocabulary fragmentation and recognition errors caused by speaking too fast, thereby effectively avoiding the risk of omission and misjudgment of users' emergency information in emergency help scenarios, so as to solve the problems in the above-mentioned background technology.

[0006] In order to achieve the above-mentioned object, the present invention provides the following technical solutions: an AI-based telephone answering system, comprising a real-time voice acquisition and preprocessing module, a speech rate detection and prediction module, an adaptive ASR dynamic control module, a refined voice slicing and decoding module, a keyword detection and voice recognition module, a semantic understanding and intention analysis module, and an emergency response and dispatch module; Real-time voice collection and preprocessing module, which collects user voice data of emergency help calls in real time, establishes a real-time audio stream transmission channel, and quickly preprocesses the acquired audio signals; The speech rate detection and prediction module quickly extracts audio data features from audio data of a fixed length, conducts in-depth analysis of the extracted speech rate features, and uses a pre-trained machine learning model to make real-time predictions of the user's current speech rate. The adaptive ASR dynamic control module activates the real-time adaptive strategy of the backend ASR engine, automatically adjusts the original speech segmentation duration, and performs refined segmentation cutting on the audio data; The refined voice slicing and decoding module receives control instructions, performs refined segmentation and slicing of the user's voice signal, and then independently decodes each segmented voice segment; The keyword detection and speech recognition module sends each fine speech slice to the real-time ASR decoding unit independently to convert speech recognition into text content. At the same time, based on the pre-defined emergency keyword vocabulary, it detects and marks keywords in real time for each speech slice content. The semantic understanding and intention analysis module integrates the text recognition results of multiple consecutive short voice segments in real time, and performs comprehensive semantic understanding and analysis through the natural language understanding unit to determine the user's overall intention and specific emergency situation; The emergency response and dispatch module, when confirming that the user's request for help is indeed an emergency, quickly triggers the pre-defined emergency response mechanism, automatically transfers the help call to manual customer service or emergency rescue hotline, or automatically pushes the alarm information, location, and help content to the corresponding department.

[0007] Preferably, the audio signal is acquired at a set sampling frequency and bit depth through a high-quality microphone device and a communication link.

[0008] Preferably, the audio data of fixed duration refers to audio segments of the same length intercepted from a continuous audio stream in speech processing.

[0009] Preferably, fast audio data features are extracted from audio data of fixed duration, wherein the extracted features include the compression ratio of the duration of each syllable per unit time relative to the standard duration and the rapid change intensity of the audio signal from positive to negative and from negative to positive per unit time. After in-depth analysis of the extracted speech rate features, a syllable duration shortening reference value and a short-time zero-crossing rate mutation reference value are generated respectively. The syllable duration shortening reference value and the short-time zero-crossing rate mutation reference value are input as feature vectors into a machine learning model that has been trained and put into use. An audio rate evaluation coefficient is generated by the model, and the user's current speech rate is predicted in real time based on the audio rate evaluation coefficient.

[0010] Preferably, the audio rate evaluation coefficient generated by real-time prediction of the user's current speech rate using a pre-trained machine learning model is compared and analyzed with a pre-set audio rate evaluation coefficient reference threshold to identify that the current user's speech rate exceeds the normal range. The specific steps are as follows: If the audio rate assessment coefficient is greater than the audio rate assessment coefficient reference threshold, the current user's speaking speed is identified as exceeding the normal range; if the audio rate assessment coefficient is less than or equal to the audio rate assessment coefficient reference threshold, the current user's speaking speed is identified as within the normal range.

[0011] Preferably, for audio data of fixed duration, the specific steps of generating a syllable duration shortening reference value after in-depth analysis of the compression ratio of each syllable duration relative to the standard duration per unit time are as follows: In audio data of fixed duration, the speech signal is segmented at the syllable level and the actual duration of each syllable is measured. The start and end times of the syllable level are obtained using the phoneme boundary detection technology, and a reference syllable duration is set. For the actual duration of a syllable, its duration compression ratio is defined as follows: ,in: For the The duration compression ratio of each syllable indicates the degree of compression of the pronunciation speed of the current syllable relative to the standard syllable duration. is the reference syllable duration, i.e. the typical syllable duration in standard pronunciation. is the actual pronunciation duration of the current syllable, is a smoothing term to avoid division by zero errors; In the entire fixed-length audio segment, the duration compression ratios of multiple syllables are summarized to construct a syllable duration shortening reference value that represents the overall speaking speed state. The construction expression of the syllable duration shortening reference value is: ,in: is the reference value for shortening the syllable duration, representing the overall speaking speed of the fixed-duration audio segment. is the total number of syllables in the audio segment, is a nonlinear adjustment factor, and its value range is , used to control the sensitivity of exponential growth at high speech rates.

[0012] Preferably, for audio data of fixed duration, the specific steps of generating a short-time zero-crossing rate mutation reference value after in-depth analysis of the rapid change intensity of the audio signal from positive to negative and from negative to positive within a unit time are as follows: Convert a fixed-length audio signal into a discrete sequence Represented by, and divided into frames according to the preset window size, we get adjacent but non-overlapping frames; The sampling points in the frame, indexed from the start of the frame To frame end index The number of switching between positive and negative polarities, that is, the number of zero-crossing points, is detected point by point, and the short-time zero-crossing rate of the frame is defined by the following formula: ,in: Indicates The short-term zero-crossing rate of the frame, is a symbolic function; After obtaining the short-time zero-crossing rate at the frame level, the short-time zero-crossing rate mutation reference value is calculated by the short-time zero-crossing rate. The calculation expression is: ,in: is the reference value of short-time zero-crossing rate mutation, is an amplification factor greater than zero, which is used to adjust the sensitivity to the change in the zero-crossing rate. Indicates The short-term zero-crossing rate of the frame.

[0013] Preferably, a predefined adaptive ASR dynamic control mechanism is automatically triggered to activate the real-time adaptive strategy of the backend ASR engine and automatically adjust the original speech segment duration. The specific steps are as follows: When it is detected that the audio rate evaluation coefficient is greater than the audio rate evaluation coefficient reference threshold, the segmentation length of the ASR engine is automatically shortened within the next fixed time. The specific expression for shortening the segmentation length of the ASR engine is: ,in: is the current segment duration, To adjust the rate coefficient, it is used to control the degree of shortening the segment; Indicates the minimum acceptable value for segment length, It is the segmentation slice length of the shortened ASR engine.

[0014] In the above technical solution, the technical effects and advantages provided by the present invention are: The present invention quickly and accurately captures the user's voice features at high speaking speeds, and dynamically adjusts the length of the ASR engine's voice segments, making each voice segment clearer and more accurate, avoiding cross-segment vocabulary fragmentation and recognition errors caused by too fast a speaking speed, thereby effectively avoiding the risk of missing and misjudgment of user emergency information in emergency help scenarios. The successful application of this method ensures that information in emergency situations can be quickly and reliably identified and responded to, greatly shortening the time it takes for medical staff, firefighters or other rescue forces to arrive at the scene of the accident, ensuring that emergency rescue agencies can accurately obtain key information and carry out rescue work in a timely manner within the precious golden rescue time, effectively avoiding or minimizing casualties, property losses and negative social impacts caused by delayed response, and significantly improving the safety and reliability of the AI ​​telephone answering system in emergency situations. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present invention. For ordinary technicians in this field, other drawings can also be obtained based on these drawings.

[0016] Figure 1 It is a module schematic diagram of the AI-based telephone answering system of the present invention. DETAILED DESCRIPTION

[0017] Example embodiments will now be described more fully with reference to the accompanying drawings. However, example embodiments can be implemented in a variety of forms and should not be construed as limited to the examples set forth herein; rather, these example embodiments are provided so that the description of the present disclosure will be more comprehensive and complete, and the concept of the example embodiments will be fully conveyed to those skilled in the art.

[0018] The present invention provides Figure 1 The AI-based telephone answering system shown includes a real-time voice acquisition and preprocessing module, a speech rate detection and prediction module, an adaptive ASR dynamic control module, a refined voice slicing and decoding module, a keyword detection and voice recognition module, a semantic understanding and intent analysis module, and an emergency response and dispatch module; Real-time voice collection and preprocessing module: When a user calls the AI ​​phone system to express emergency help, the user's voice data is collected in real time, a real-time audio stream transmission channel is established, and the acquired audio signal is quickly preprocessed; Through high-quality microphone equipment and communication links, audio signals are acquired at appropriate sampling frequencies (such as 16kHz or higher) and bit depths. At the same time, the acquired audio signals are quickly pre-processed, such as noise reduction and filtering, to suppress background noise interference, improve audio data quality, and ensure the reliability of subsequent feature extraction and recognition.

[0019] Speech speed detection and prediction module: selects audio data of a fixed length, extracts the fast audio data features, conducts in-depth analysis of the extracted speech speed features, and uses a pre-trained machine learning model to make real-time predictions of the user's current speech speed; Fixed-length audio data refers to audio segments of the same length intercepted from a continuous audio stream in speech processing. For example, in speech rate prediction, speech recognition, or sentiment analysis tasks, audio windows of fixed time length (such as 0.5 seconds, 1.0 seconds, or 1.5 seconds) are usually used as basic processing units to ensure the uniformity of data input and the stability of calculations. This fixed-length setting can avoid the increased complexity of model training and reasoning caused by inconsistent audio data length. At the same time, fixed-length audio data helps to make equivalent comparisons of different speech segments during feature extraction and optimize the accuracy and robustness of ASR (automatic speech recognition) or speech analysis systems.

[0020] Extract fast audio data features from audio data of fixed duration, wherein the extracted features include the compression ratio of the duration of each syllable relative to the standard duration per unit time and the rapid change intensity of the audio signal from positive to negative and from negative to positive per unit time. After in-depth analysis of the extracted speech speed features, a syllable duration shortening reference value and a short-time zero-crossing rate mutation reference value are generated respectively, and the syllable duration shortening reference value and the short-time zero-crossing rate mutation reference value are used to characterize the user's current speech speed; The syllable duration shortening reference value and the short-time zero-crossing rate mutation reference value are input as feature vectors into the trained and used machine learning model. The audio rate evaluation coefficient is generated by the model, and the user's current speaking speed is predicted in real time based on the audio rate evaluation coefficient.

[0021] In audio data of fixed duration, if the average duration of each syllable per unit time is detected to be significantly compressed compared to the standard pronunciation duration, it means that the speaker pronounces more syllables in the same amount of time, or pronounces a single syllable shorter. This phenomenon of "compression" of syllable duration precisely reflects the increase in speaking speed, that is, the user is pronouncing at a faster pace, so that the entire call content contains more voice information in a limited time.

[0022] For audio data of fixed duration, the specific steps for generating a reference value for shortening the duration of a syllable after in-depth analysis of the compression ratio of the duration of each syllable relative to the standard duration per unit time are as follows: In audio data of fixed duration, the speech signal is first segmented at the syllable level and the actual duration of each syllable is measured. Then, the start and end times of the syllable level are obtained using phoneme boundary detection technology or endpoint detection algorithms (such as Viterbi decoding based on time-frequency features). A reference syllable duration is set, which can be based on the standard syllable duration defined by linguistics (for example, the typical syllable duration in Mandarin Chinese is 200~300ms) or the syllable duration annotation from a high-confidence speech synthesis system. For the actual duration of a syllable, its duration compression ratio is defined as follows: ,in: For the The duration compression ratio of each syllable indicates the degree of compression of the pronunciation speed of the current syllable relative to the standard syllable duration. is the reference syllable duration, i.e. the typical syllable duration in standard pronunciation. is the actual pronunciation duration of the current syllable, is a very small smoothing term to avoid division by zero errors, usually ; When the duration of a single syllable is compressed to a large extent, it means that the syllable is significantly compressed compared to the standard duration, which means that the speaker completes the pronunciation of the syllable in a shorter time, reflecting the faster speaking speed; The main purpose of this step is to establish the duration compression ratio of a single syllable, to provide basic data for the subsequent calculation of the reference value for syllable duration shortening, and to more sensitively capture the abnormal acceleration of individual syllables in order to cope with high-speed speech in different languages ​​and contexts.

[0023] In the entire fixed-length audio segment, the duration compression ratios of multiple syllables are summarized to construct a syllable duration shortening reference value that represents the overall speaking speed state. Based on the idea of ​​information entropy, a weighted sum is constructed using the exponential growth trend of the syllable duration compression ratio to make it more sensitive to high-speed speech. The generation expression of the syllable duration shortening reference value is: ,in: is the reference value for shortening the syllable duration, representing the overall speaking speed of the fixed-duration audio segment. is the total number of syllables in the audio segment, is a nonlinear adjustment factor, and its value range is , used to control the sensitivity of exponential growth at high speech rates, usually between 0.5 and 0.8; The syllable duration shortening reference value is directly used as the final indicator to characterize the speaking speed, and can effectively distinguish between speech segments with fast, normal or slow speaking speeds. The syllable duration shortening reference value has a good dynamic range and nonlinear mapping capability, so that the faster the speaking speed, the more obvious the increase of the syllable duration shortening reference value.

[0024] It can be seen from the syllable duration shortening reference value that, for audio data of fixed duration, the greater the performance value of the syllable duration shortening reference value generated after in-depth analysis of the compression ratio of the duration of each syllable in a unit time relative to the standard duration, the faster the user's current speaking speed is, and vice versa. This is because the syllable duration shortening reference value is calculated based on the compression ratio of the duration of each syllable in a unit time relative to the standard duration. When the user's speaking speed increases, the pronunciation duration of each syllable will be shortened, and the compression ratio relative to the standard syllable duration will increase, resulting in a cumulative increase in the compression ratio of multiple syllables, which increases the value of the syllable duration shortening reference value. On the contrary, when the user's speaking speed is slow, the pronunciation duration of the syllable approaches or exceeds the standard duration, the compression ratio decreases, and the performance value of the syllable duration shortening reference value decreases. Therefore, the syllable duration shortening reference value can be used as a reliable indicator to measure the speed of speech, directly reflecting the real-time speech speed changes of the user.

[0025] The intensity of the rapid change of an audio signal from positive to negative or from negative to positive in a unit time is usually called the short-time zero-crossing rate. Specifically, when the user's speaking speed is very fast, the voice segment in the audio signal tends to have more phonemes or syllables in a short period of time, which significantly increases the number of oscillations of the voice signal in a unit time, which is manifested as the audio waveform passing through the zero point from positive polarity to negative polarity, or from negative polarity through the zero point to positive polarity. Changes are more frequent and intensive. Therefore, in audio data of a fixed length, if this short-time zero-crossing rate is detected to be significantly increased and rapidly changing, it can be regarded as the user's pronunciation speed is fast, which is one of the important acoustic indicators that effectively measures and reflects the speed of the user's speech.

[0026] For audio data of fixed duration, the specific steps of generating a short-time zero-crossing rate mutation reference value after in-depth analysis of the rapid changes of the audio signal from positive to negative and from negative to positive in unit time are as follows: First, the fixed-length audio signal is converted into a discrete sequence Indicates that the window size is set according to the preset window size (such as W sampling points) to obtain adjacent but non-overlapping frames. The sampling points in the frame, indexed from the start of the frame To frame end index The number of switching between positive and negative polarities, that is, the number of zero-crossing points, is detected point by point, and the short-time zero-crossing rate of the frame is defined by the following formula: ,in: Indicates The short-term zero-crossing rate of the frame, It is a sign function. When the input is greater than 0, it outputs 1, when it is less than 0, it outputs -1, and when it is equal to 0, it can be regarded as 0. The larger the value is, the more frequent the switching between positive and negative polarities of the speech signal in the frame is. Through this step, the short-time zero-crossing rate sequence of each frame is obtained , which lays the foundation for the subsequent judgment of speech rate mutation. The role of this step is to quantify the zero-crossing phenomenon of the audio signal into a numerical description at the frame level, thereby providing core data support for further constructing a short-term zero-crossing rate mutation reference value.

[0027] After obtaining the short-term zero-crossing rate at the frame level, we measure its drastic changes between adjacent frames, and then capture the mutation characteristics caused by high speech speed in numerical form. To this end, we calculate the short-term zero-crossing rate mutation reference value through the short-term zero-crossing rate. The calculation expression is: ,in: is the reference value of short-time zero-crossing rate mutation, is an amplification factor greater than zero, used to adjust the sensitivity to the change in the zero-crossing rate; It is used to enhance nonlinear mapping so that the difference in zero-crossing rate with medium amplitude can be appropriately amplified. Indicates The short-time zero-crossing rate of the frame. When the difference in the number of zero crossings between adjacent frames becomes significantly larger, The value of will also increase accordingly, thereby cumulatively improving the short-time zero-crossing rate mutation reference value. If the short-time zero-crossing rate mutation reference value increases significantly over a period of time, it means that the user is rapidly switching between positive and negative polarity at an extremely fast speaking speed, thereby helping the system detect the "high speaking speed" state in real-time or offline analysis.

[0028] The core function of this step is to integrate the change of zero-crossing rate between frames into a mutation measurement index, which can more sensitively reflect the frequent waveform jumps caused by the sudden increase in speech speed.

[0029] It can be seen from the short-time zero-crossing rate mutation reference value that, for audio data of fixed duration, the short-time zero-crossing rate mutation reference value generated after in-depth analysis of the rapid change intensity of the audio signal from positive to negative and from negative to positive in unit time is larger, usually indicating that the user's current speaking speed is faster, and vice versa. This is because the short-time zero-crossing rate mutation reference value measures the rapid switching degree of the positive and negative polarity of the audio signal within a fixed duration, and the faster the speaking speed, the higher the phoneme conversion frequency in the pronunciation, resulting in more intense fluctuations in the zero-crossing rate of the waveform in a short time, thereby increasing the cumulative value of the short-time zero-crossing rate mutation reference value. On the contrary, when the speaking speed is slow, the user prolongs the syllables and pauses more during the pronunciation process, the zero-crossing change amplitude of the waveform is smaller, and the zero-crossing rate difference between adjacent frames is smaller, which makes the growth of the short-time zero-crossing rate mutation reference value slower or even stable. Therefore, the high or low short-time zero-crossing rate mutation reference value can be directly used as a characterization indicator of the speaking speed, providing a quantitative basis for real-time speech analysis.

[0030] A machine learning model that has been trained and put into use refers to a machine learning model that has gone through a series of standardized model construction, data training, parameter tuning, and performance verification, and the performance and stability of the model have met the design requirements, and has been deployed in the actual application environment. It can process input data in real time and stably, and output reliable prediction results. Specifically, this process first involves collecting a large amount of historical audio data with clear labels, especially covering various speech speed scenarios (including normal speaking speed, slow speaking speed, medium speed, fast speaking speed, and even extremely fast speaking speed), and clearly marking the accurate speech speed level or value corresponding to each audio data segment. Subsequently, based on the collected audio data, using specialized speech speed-related features such as syllable duration shortening reference value and short-time zero-crossing rate mutation reference value as input features, a model suitable for audio speech speed prediction tasks is constructed, such as a neural network model (such as CNN, LSTM, Transformer model) or a traditional machine learning model (such as random forest, support vector machine). In the training stage, it is necessary to gradually correct the model weights and parameters through repeated iterations and optimizations, reduce prediction errors, and improve the generalization ability and prediction accuracy of the model until the performance of the model on the validation data set reaches or exceeds the expected target.

[0031] When the above training process is completely completed, it means that the model has the ability to process data in actual application scenarios and can stably output accurate prediction results. At this time, the model is a "trained and put into use machine learning model". In the actual deployment stage, this model will be encapsulated into an independent, callable module and deployed on the actual application system or server. When the user makes a call, after collecting the user's audio signal in real time, the system will quickly complete the data preprocessing and feature extraction process, such as extracting the syllable duration shortening reference value and the short-time zero-crossing rate mutation reference value from the audio. These two feature vectors specifically reflect the trend of speech rate changes, and input them into the trained machine learning model in real time. At this time, the model will quickly calculate and infer the current input feature data based on the data distribution, rules and feature expressions learned in the previous training process, and output an audio rate evaluation coefficient that can quantify the user's current speech rate status in real time, and quickly identify, judge and predict the user's real-time speech rate status based on this. Therefore, "a machine learning model that has been trained and put into use" not only means that the process of building and training the model itself has been successfully completed, but also means that the application capabilities and actual value of the model have been truly realized, becoming an indispensable part of the actual operation of the system, and can effectively achieve the goal of real-time detection of user speech speed and intelligent response.

[0032] The machine learning model is not specifically limited here, and can achieve the goal of shortening the syllable duration to a reference value. and short-time zero-crossing rate mutation reference value Perform comprehensive analysis to generate audio rate estimation coefficients In order to realize the technical solution of the present invention, the present invention provides a specific implementation method; audio rate evaluation coefficient The generated expression is: , where , are the reference values ​​for shortening syllable duration and short-time zero-crossing rate mutation reference value The preset scaling factor of , are all greater than 0. The preset proportional coefficient (i.e. , ) refers to the generation of the audio rate estimation coefficient ( ), the syllable duration shortening reference values ​​are given respectively ( ) and short-time zero-crossing rate mutation reference value ( ) is a coefficient that represents the weight or contribution of different features. The purpose of these proportional coefficients is to adjust and control the degree of influence of different features on the final generated result, reflect the difference in importance between features, and enable the final generated audio rate evaluation coefficient to more accurately reflect the actual speaking rate. Usually, these coefficients are obtained through experimental verification, data training and optimization, and represent the importance of specific features in the combined model. For example, if it is found through experiments that the syllable duration shortening reference value contributes more to the speaking rate judgment, a larger proportional coefficient should be set to increase its influence; on the contrary, if the short-term zero-crossing rate mutation reference value contributes less, the proportional coefficient can be relatively reduced. In short, the preset proportional coefficient is to enable the final generated audio rate evaluation coefficient to more accurately reflect the actual user speaking rate, improve the speaking rate judgment accuracy and the reliability of the model prediction effect.

[0033] It can be seen from the audio rate evaluation coefficient that for audio data of fixed length, the greater the performance value of the syllable duration shortening reference value generated after in-depth analysis of the compression ratio of the duration of each syllable relative to the standard duration per unit time, the greater the performance value of the short-time zero-crossing rate mutation reference value generated after in-depth analysis of the rapid change of the audio signal from positive to negative and from negative to positive per unit time, that is, the greater the performance value of the audio rate evaluation coefficient generated by real-time prediction of the user's current speaking speed using a pre-trained machine learning model, the faster the user's current speaking speed, and vice versa.

[0034] The audio rate evaluation coefficient generated by real-time prediction of the user's current speech rate using a pre-trained machine learning model is compared and analyzed with the pre-set audio rate evaluation coefficient reference threshold to identify whether the current user's speech rate exceeds the normal range. The specific steps are as follows: If the audio rate evaluation coefficient is greater than the audio rate evaluation coefficient reference threshold, the current user's speaking speed is identified as exceeding the normal range, that is, the current user's speaking speed is faster at this time; if the audio rate evaluation coefficient is less than or equal to the audio rate evaluation coefficient reference threshold, the current user's speaking speed is identified as within the normal range, that is, the current user's speaking speed is normal at this time.

[0035] Adaptive ASR dynamic control module: If the machine learning model identifies that the current user's speaking speed exceeds the normal range, it will automatically trigger the pre-defined adaptive ASR dynamic control mechanism, activate the real-time adaptive strategy of the backend ASR engine, automatically adjust the original speech segmentation time (such as from 1 second to 0.5 seconds or even shorter), and perform refined segmentation cutting on the audio data; Automatically trigger the pre-defined adaptive ASR dynamic control mechanism, activate the real-time adaptive strategy of the backend ASR engine, and automatically adjust the original speech segment duration. The specific steps are as follows: When it is detected that the audio rate evaluation coefficient is greater than the audio rate evaluation coefficient reference threshold, the system will automatically shorten the segmentation length of the ASR engine within the next fixed time. The specific expression for shortening the segmentation length of the ASR engine is: ,in: is the current segment duration (e.g. 1 second), To adjust the rate coefficient, it is used to control the degree of shortening the segment; Indicates the minimum acceptable value of the segment length (to avoid excessive shortening that affects real-time performance). The segment slice length of the shortened ASR engine; If the difference between the detected audio rate assessment coefficient and the audio rate assessment coefficient reference threshold is larger, that is, the faster the speech speed is, the more obvious the shortening of the segment duration will be; otherwise, the shortening will be relatively weaker. Through this step, the system can divide the audio data into smaller segments in high-speed speech scenarios, thereby improving the accuracy of the ASR engine during recognition and reducing confusion caused by cross-segmentation. The main function of this mechanism is to implement real-time and dynamic processing strategies to ensure that when the user's speech speed suddenly increases, the system can adapt quickly to ensure the timely capture and recognition of key information (such as emergency help keywords), and prevent missed judgments and misjudgments caused by incorrect splicing or ignoring of large segments of speech at high speeds.

[0036] Refined voice slicing and decoding module: The backend ASR engine refines and slices the user's voice signal in a shorter time window according to the control instructions in the previous step, and then decodes each segmented voice fragment independently. In this way, the content of each voice slice will be clearer, reducing cross-segment confusion caused by too long fragments and too much content.

[0037] Keyword detection and speech recognition module: After the dynamic segmentation of speech slices, each fine speech slice is independently sent to the real-time ASR decoding unit for speech recognition and conversion into text content. At the same time, based on the pre-defined emergency keyword vocabulary (such as "help", "fire", "first aid", etc.), real-time keyword detection and tagging are performed on the content of each speech slice. Once a key emergency word is found, the weight and attention mechanism priority of the segment are immediately increased to trigger the emergency help process.

[0038] Semantic understanding and intention analysis module: Although the accuracy is improved after the voice segment is shortened, the integrity of the contextual semantics still needs to be considered. The system then fuses the text recognition results of multiple consecutive short voice segments in real time, and performs comprehensive semantic understanding and analysis through the natural language understanding (NLU) unit to determine the user's overall intention and specific emergency situation. For example, the user's specific help needs, location information and other important emergency information can be obtained through context fusion.

[0039] Emergency response and dispatch module: When it is confirmed that the user's request for help is indeed an emergency, the pre-defined emergency response mechanism is quickly triggered, and the call for help is automatically transferred to manual customer service or emergency rescue hotline (such as 120, 119 or 110) or the alarm information, location, and help content are automatically pushed to the corresponding department. In addition, to avoid any risks caused by AI misjudgment, the system should establish a bottom-line protection mechanism: when the confidence of the recognition result is low or there is uncertainty, the user will be automatically and quickly transferred to manual customer service to prevent delays in rescue.

[0040] The implementation of the above solution can significantly improve the accuracy and timeliness of emergency voice recognition by the AI ​​telephone answering system. By real-time monitoring of the user's speech speed changes, the segmentation time of the voice slices can be adaptively shortened to make the processing granularity of the audio data more refined, greatly reducing the cross-segment information confusion, missed recognition and misrecognition problems caused by excessive speaking speed under the traditional fixed segment length. Specifically, by quickly and accurately capturing the user's voice features at high speaking speed and dynamically adjusting the speech segment length of the ASR engine, each voice segment is made clearer and more accurate, avoiding cross-segment vocabulary fragmentation and recognition errors caused by excessive speaking speed, thereby effectively avoiding the risk of missing and misjudgment of user emergency information in emergency help scenarios. Ultimately, the successful application of this method ensures that information in emergency situations can be quickly and reliably identified and responded to, significantly shortens the time it takes for medical staff, firefighters or other rescue forces to arrive at the scene of the accident, and ensures that emergency rescue agencies can accurately obtain key information and carry out rescue work in a timely manner within the precious golden rescue time, effectively avoiding or minimizing casualties, property losses and negative social impacts caused by delayed response, and significantly improving the safety and reliability of the AI ​​telephone answering system in emergency situations.

[0041] The above formulas are all dimensionless and numerical calculations. The formula is a formula for the most recent real situation obtained by collecting a large amount of data and performing software simulation. The preset parameters in the formula are set by technicians in this field according to actual conditions.

[0042] The above description is only by way of illustration of certain exemplary embodiments of the present invention. It is undoubted that those skilled in the art can modify the described embodiments in various ways without departing from the spirit and scope of the present invention. Therefore, the above drawings and descriptions are illustrative in nature and should not be construed as limiting the scope of protection of the claims of the present invention.

[0043] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art who is familiar with the present technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. AI-based telephone answering system, characterized by: It includes real-time speech acquisition and preprocessing module, speech rate detection and prediction module, adaptive ASR dynamic control module, refined speech slicing and decoding module, keyword detection and speech recognition module, semantic understanding and intention analysis module, and emergency response and dispatch module; Real-time voice collection and preprocessing module, which collects user voice data of emergency help calls in real time, establishes a real-time audio stream transmission channel, and quickly preprocesses the acquired audio signals; The speech rate detection and prediction module quickly extracts audio data features from audio data of a fixed length, conducts in-depth analysis of the extracted speech rate features, and uses a pre-trained machine learning model to make real-time predictions of the user's current speech rate. The adaptive ASR dynamic control module activates the real-time adaptive strategy of the backend ASR engine, automatically adjusts the original speech segmentation duration, and performs refined segmentation cutting on the audio data; The refined voice slicing and decoding module receives control instructions, performs refined segmentation and slicing of the user's voice signal, and then independently decodes each segmented voice segment; The keyword detection and speech recognition module sends each fine speech slice to the real-time ASR decoding unit independently to convert speech recognition into text content. At the same time, based on the pre-defined emergency keyword vocabulary, it detects and marks keywords in real time for each speech slice content. The semantic understanding and intention analysis module integrates the text recognition results of multiple consecutive short voice segments in real time, and performs comprehensive semantic understanding and analysis through the natural language understanding unit to determine the user's overall intention and specific emergency situation; The emergency response and dispatch module, when confirming that the user's request for help is indeed an emergency, quickly triggers the pre-defined emergency response mechanism, automatically transfers the help call to manual customer service or emergency rescue hotline, or automatically pushes the alarm information, location, and help content to the corresponding department.

2. The AI-based telephone answering system according to claim 1, characterized in that: Acquire audio signals at a set sampling rate and bit depth through high-quality microphone devices and communication links.

3. The AI-based telephone answering system according to claim 1, characterized in that: Fixed-length audio data refers to audio segments of the same length extracted from a continuous audio stream in speech processing.

4. The AI-based telephone answering system according to claim 1, characterized in that: Fast audio data features are extracted from audio data of fixed duration, wherein the extracted features include the compression ratio of the duration of each syllable relative to the standard duration per unit time and the rapid change intensity of the audio signal from positive to negative and from negative to positive per unit time. After in-depth analysis of the extracted speech speed features, a syllable duration shortening reference value and a short-time zero-crossing rate mutation reference value are generated respectively. The syllable duration shortening reference value and the short-time zero-crossing rate mutation reference value are input as feature vectors into a machine learning model that has been trained and put into use. The audio rate evaluation coefficient is generated by the model, and the user's current speech rate is predicted in real time based on the audio rate evaluation coefficient.

5. The AI-based telephone answering system according to claim 4, characterized in that: The audio rate evaluation coefficient generated by real-time prediction of the user's current speaking speed using a pre-trained machine learning model is compared and analyzed with the pre-set audio rate evaluation coefficient reference threshold to identify whether the current user's speaking speed exceeds the normal range. The specific steps are as follows: If the audio rate assessment coefficient is greater than the audio rate assessment coefficient reference threshold, the current user's speaking speed is identified as exceeding the normal range; if the audio rate assessment coefficient is less than or equal to the audio rate assessment coefficient reference threshold, the current user's speaking speed is identified as within the normal range.

6. The AI-based telephone answering system according to claim 4, characterized in that: For audio data of fixed duration, the specific steps for generating a reference value for shortening the duration of a syllable after in-depth analysis of the compression ratio of the duration of each syllable relative to the standard duration per unit time are as follows: In audio data of fixed duration, the speech signal is segmented at the syllable level and the actual duration of each syllable is measured. The start and end times of the syllable level are obtained using the phoneme boundary detection technology, and a reference syllable duration is set. For the actual duration of a syllable, its duration compression ratio is defined as follows: ,in: For the The duration compression ratio of each syllable indicates the degree of compression of the pronunciation speed of the current syllable relative to the standard syllable duration. is the reference syllable duration, i.e. the typical syllable duration in standard pronunciation. is the actual pronunciation duration of the current syllable, is a smoothing term to avoid division by zero errors; In the entire fixed-length audio segment, the duration compression ratios of multiple syllables are summarized to construct a syllable duration shortening reference value that represents the overall speaking speed state. The construction expression of the syllable duration shortening reference value is: ,in: is the reference value for shortening the syllable duration, representing the overall speaking speed of the fixed-duration audio segment. is the total number of syllables in the audio segment, is a nonlinear adjustment factor, and its value range is , used to control the sensitivity of exponential growth at high speech rates.

7. The AI-based telephone answering system according to claim 4, characterized in that: For audio data of fixed duration, the specific steps of generating a short-time zero-crossing rate mutation reference value after in-depth analysis of the rapid changes of the audio signal from positive to negative and from negative to positive in unit time are as follows: Convert a fixed-length audio signal into a discrete sequence Represented by, and divided into frames according to the preset window size, we get adjacent but non-overlapping frames; The sampling points in the frame, indexed from the start of the frame To frame end index The number of switching points of positive and negative polarity is detected point by point, that is, the number of zero crossing points, and the following formula is used to define the The short-time zero-crossing rate of the frame is defined as: ,in: Indicates The short-term zero-crossing rate of the frame, is a symbolic function; After obtaining the short-time zero-crossing rate at the frame level, the short-time zero-crossing rate mutation reference value is calculated by the short-time zero-crossing rate. The calculation expression is: ,in: is the reference value of short-time zero-crossing rate mutation, is an amplification factor greater than zero, used to adjust the sensitivity to the change in the zero-crossing rate. Indicates The short-term zero-crossing rate of the frame.

8. The AI-based telephone answering system according to claim 5, characterized in that: Automatically trigger the pre-defined adaptive ASR dynamic control mechanism, activate the real-time adaptive strategy of the backend ASR engine, and automatically adjust the original speech segment duration. The specific steps are as follows: When it is detected that the audio rate evaluation coefficient is greater than the audio rate evaluation coefficient reference threshold, the segmentation length of the ASR engine is automatically shortened within the next fixed time. The specific expression for shortening the segmentation length of the ASR engine is: ,in: is the current segment duration, To adjust the rate coefficient, it is used to control the degree of shortening the segment; Indicates the minimum acceptable value for segment length, It is the segmentation slice length of the shortened ASR engine.

Images