AI-based telephone answering system

Abstract

The application discloses an AI-based telephone answering system, relates to the technical field of telephone answering, and comprises a real-time voice acquisition and preprocessing module, a speech speed detection and prediction module, a self-adaptive ASR dynamic regulation 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 dispatching module; the real-time voice acquisition and preprocessing module acquires user voice data of an emergency help telephone in real time, establishes a real-time audio stream transmission channel, and rapidly preprocesses the acquired audio signal. The application rapidly and accurately captures the voice features of a user under high speech speed, dynamically adjusts the voice segmentation length of an ASR engine, makes each voice segment more clear and accurate, avoids the splitting and recognition errors of cross-segment words caused by too fast speech speed, and effectively avoids the omission of user emergency information and the risk of misjudgment in an emergency help scene.

Description

Technical Field

[0001] This invention relates to the field of telephone answering technology, and more specifically to an AI-based telephone answering system. Background Technology

[0002] AI-based telephone answering systems are intelligent interactive systems that automatically process incoming calls using artificial intelligence technology. They are widely used in customer service, voice assistants, and corporate hotlines. These systems typically combine Automatic Speech Recognition (ASR), Natural Language Processing (NLP), and Text-to-Speech (TTS) technologies to recognize user speech, interpret intent, and provide intelligent responses. They can understand user needs and generate accurate answers based on knowledge bases, databases, or deep learning models, even engaging in multi-turn conversations. Furthermore, the system can provide personalized services such as business processing, information inquiries, and appointment scheduling through sentiment analysis, contextual understanding, and integration with CRM (Customer Relationship Management) systems. Compared to traditional IVR (Interactive Voice Response) systems, AI telephone answering systems are more intelligent, reduce user keystrokes, improve service efficiency and user experience, and lower the cost of human customer service representatives.

[0003] The existing technology has the following shortcomings:

[0004] When users call to express emergency help, their speech often quickens due to anxiety and tension. In such situations, AI call response systems may struggle to accurately recognize and understand the user's voice, leading to misjudgment or ignoring the user's urgent need for assistance. If the AI ​​system fails to detect the user's emergency signal in time, it cannot quickly transfer the call to a human operator or emergency services. This results in paramedics, firefighters, or other emergency rescue teams being unable to reach the scene immediately, missing crucial rescue time, causing irreparable personal injury or serious property damage, and potentially even escalating into a more serious public safety incident.

[0005] The information disclosed in the background section is only intended to enhance the understanding of the background of this disclosure, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0006] The purpose of this invention is to provide an AI-based telephone response system that can quickly and accurately capture the voice features of users speaking at high speeds and dynamically adjust the length of voice segments in the ASR engine, making each voice segment clearer and more accurate. This avoids word fragmentation and recognition errors caused by excessively fast speaking speed, thereby effectively preventing the omission and misjudgment of users' emergency information in emergency help scenarios, thus solving the problems mentioned in the background technology.

[0007] To achieve the above objectives, the present invention provides the following technical solution: an AI-based telephone response system, comprising a real-time voice acquisition and preprocessing module, a speech rate detection and prediction module, an adaptive ASR dynamic adjustment module, a refined speech slicing and decoding module, a keyword detection and speech recognition module, a semantic understanding and intent analysis module, and an emergency response and dispatch module;

[0008] The real-time voice acquisition and preprocessing module acquires user voice data from emergency calls in real time, establishes a real-time audio stream transmission channel, and performs rapid preprocessing on the acquired audio signals.

[0009] The speech rate detection and prediction module quickly extracts audio data features from audio data of a fixed duration, performs in-depth analysis of the extracted speech rate features, and then uses a pre-trained machine learning model to predict the user's current speech rate in real time.

[0010] The adaptive ASR dynamic adjustment module activates the real-time adaptive strategy of the backend ASR engine, automatically adjusts the original speech segment duration, and performs fine segmentation and cutting of audio data.

[0011] The fine-grained speech slicing and decoding module receives control commands, performs fine-grained segmentation and slicing of the user's speech signal, and then decodes each segmented speech independently.

[0012] The keyword detection and speech recognition module independently sends each fine speech slice into the real-time ASR decoding unit for speech recognition and conversion into text content. At the same time, based on a predefined emergency keyword lexicon, it performs real-time keyword detection and labeling on each speech slice content.

[0013] The semantic understanding and intent analysis module fuses the text recognition results of multiple consecutive short speech segments in real time, and performs comprehensive semantic understanding analysis through the natural language understanding unit to determine the user's overall intent and specific emergency situation;

[0014] The emergency response and dispatch module, when confirming that the user's request for help is indeed an emergency, quickly triggers a predefined emergency response mechanism, automatically transferring the call to human customer service or emergency rescue hotline, or automatically pushing alarm information, location, and request details to the appropriate department.

[0015] Preferably, audio signals are acquired using high-quality microphone equipment and communication links at a set sampling frequency and bit depth.

[0016] Preferably, fixed-duration audio data refers to audio segments of the same length extracted from a continuous audio stream in speech processing.

[0017] Preferably, rapid audio data features are extracted from audio data of fixed duration. The extracted features include the compression ratio of the duration of each syllable relative to the standard duration per unit time and the degree of rapid change 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, syllable duration shortening reference values ​​and short-term zero-crossing rate mutation reference values ​​are generated. The syllable duration shortening reference values ​​and short-term zero-crossing rate mutation reference values ​​are used as feature vectors and input into the trained and put into use machine learning model. The model generates audio rate evaluation coefficients, and the user's current speech rate is predicted in real time based on the audio rate evaluation coefficients.

[0018] 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 users whose speech rate exceeds the normal range. The specific steps are as follows:

[0019] If the audio rate evaluation coefficient is greater than the audio rate evaluation coefficient reference threshold, the current user's speech rate is identified as exceeding the normal range; if the audio rate evaluation coefficient is less than or equal to the audio rate evaluation coefficient reference threshold, the current user's speech rate is identified as within the normal range.

[0020] Preferably, the specific steps for generating a syllable duration shortening reference value after in-depth analysis of the compression ratio of each syllable's duration relative to the standard duration within a unit of time for audio data of a fixed duration are as follows:

[0021] In audio data of fixed duration, the speech signal is segmented at the syllable level, and the actual duration of each syllable is measured. Phoneme boundary detection technology is used to obtain the start and end times at the syllable level. A reference syllable duration is set, and the duration compression ratio for a given syllable is defined as follows:

[0022] ,in: For the first The duration compression ratio of a syllable indicates the degree of compression of the pronunciation speed of the current syllable relative to the standard syllable duration. The reference syllable duration is the typical syllable duration under standard pronunciation conditions. The actual duration of pronunciation of the current syllable. To create a smooth term and avoid division by zero errors;

[0023] Within a fixed-duration audio segment, the duration compression ratios of multiple syllables are aggregated to construct a syllable duration reduction reference value that characterizes the overall speech rate. The expression for constructing the syllable duration reduction reference value is as follows:

[0024] ,in: This serves as a reference value for syllable duration shortening, representing the overall speech rate of this fixed-duration audio segment. This represents the total number of syllables in the audio segment. This is a non-linear adjustment factor, with a value range of [value range missing]. This is used to control the sensitivity to exponential growth under high speech rates.

[0025] Preferably, the specific steps for generating short-time zero-crossing rate abrupt change reference values ​​after conducting in-depth analysis of the rapid changes in the audio signal from positive to negative and from negative to positive within a unit of time for audio data of a fixed duration are as follows:

[0026] A fixed-duration audio signal is represented by a discrete sequence. This is represented and divided into frames according to a pre-set window size, resulting in... The adjacent but non-overlapping frames; for the first The sampling points in the frame, starting from the frame start index End of frame index The number of positive and negative polarity switching points is detected point by point, i.e., the number of zero crossings, and the short-time zero crossing rate of the frame is defined by the following formula:

[0027] ,in: Indicates the first The short-time zero-crossing rate of the frame, It is a symbolic function;

[0028] After obtaining the frame-level short-time zero-crossing rate, the short-time zero-crossing rate abrupt reference value is calculated using the short-time zero-crossing rate. The calculation expression is as follows:

[0029] ,in: This serves as a short-term reference value for zero-crossing rate mutations. This is a positive amplification factor used to adjust the sensitivity to changes in the zero-crossing rate. Indicates the first The short-time zero-crossing rate of the frame.

[0030] Preferably, a predefined adaptive ASR dynamic adjustment 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:

[0031] When the audio rate evaluation coefficient is detected to be greater than the audio rate evaluation coefficient reference threshold, the segment slice length of the ASR engine is automatically shortened within the next fixed duration. The specific expression for shortening the ASR engine segment slice is as follows:

[0032] ,in: This represents the current segment duration. The rate coefficient is used to control the degree of segment shortening; This represents the minimum acceptable segment length. This refers to the segment slice length of the shortened ASR engine.

[0033] The technical effects and advantages provided by the present invention in the above technical solution are as follows:

[0034] This invention rapidly and accurately captures the speech features of users speaking at high speeds and dynamically adjusts the speech segment length of the ASR engine, making each speech segment clearer and more precise. This avoids word fragmentation and recognition errors caused by excessively fast speech, effectively preventing the omission and misjudgment of users' urgent information in emergency situations. The successful application of this method ensures rapid and reliable identification and response to information in emergency situations, significantly shortening the time for medical personnel, firefighters, or other rescue forces to reach the accident scene. It guarantees that emergency rescue organizations can accurately obtain key information and promptly carry out rescue work within the precious golden rescue time, effectively avoiding or minimizing casualties, property losses, and negative social impacts caused by response delays, and significantly improving the safety and reliability of AI telephone response systems in emergency situations. Attached Figure Description

[0035] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0036] Figure 1 This is a schematic diagram of the modules of the AI-based telephone answering system of the present invention. Detailed Implementation

[0037] Exemplary embodiments will now be described more fully with reference to the accompanying drawings. However, these exemplary embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, they are provided so that the description of this disclosure will be more complete and fully convey the concept of the exemplary embodiments to those skilled in the art.

[0038] This invention provides, for example Figure 1 The AI-based telephone response system shown includes a real-time voice acquisition and preprocessing module, a speech rate detection and prediction module, an adaptive ASR dynamic adjustment module, a fine-grained speech slicing and decoding module, a keyword detection and speech recognition module, a semantic understanding and intent analysis module, and an emergency response and dispatch module.

[0039] Real-time voice acquisition and preprocessing module: When a user calls the AI ​​phone system to express an emergency request for help, the system acquires the user's voice data in real time, establishes a real-time audio stream transmission channel, and performs rapid preprocessing on the acquired audio signal.

[0040] Audio signals are acquired using high-quality microphone equipment and communication links at appropriate sampling frequencies (e.g., 16kHz or higher) and bit depths. Simultaneously, the acquired audio signals undergo rapid preprocessing, 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.

[0041] Speech rate detection and prediction module: Select audio data of fixed duration, extract fast audio data features, conduct in-depth analysis of the extracted speech rate features, and use a pre-trained machine learning model to predict the user's current speech rate in real time.

[0042] Fixed-duration audio data refers to audio segments of uniform length extracted from a continuous audio stream in speech processing. For example, in tasks such as speech rate prediction, speech recognition, or sentiment analysis, fixed-duration audio windows (e.g., 0.5 seconds, 1.0 second, or 1.5 seconds) are typically used as the basic processing unit to ensure data input consistency and computational stability. This fixed-duration setting avoids increased model training and inference complexity caused by inconsistent audio data lengths. Furthermore, fixed-duration audio data facilitates equivalent comparisons of different speech segments during feature extraction, optimizing the accuracy and robustness of ASR (Automatic Speech Recognition) or speech analysis systems.

[0043] Rapid audio data features are extracted from fixed-duration audio data. The extracted features include the compression ratio of each syllable duration relative to the standard duration per unit time and the degree of rapid change 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, syllable duration shortening reference value and short-term zero-crossing rate mutation reference value are generated respectively. The syllable duration shortening reference value and short-term zero-crossing rate mutation reference value are used to characterize the user's current speech rate.

[0044] The reference values ​​for syllable duration shortening and short-term zero-crossing rate mutation are used as feature vectors and input into the trained and deployed machine learning model. The model generates audio rate evaluation coefficients, and the user's current speech rate is predicted in real time based on the audio rate evaluation coefficients.

[0045] In audio data of a fixed duration, if the average duration of each syllable per unit time is significantly compressed compared to the standard pronunciation duration, it indicates that the speaker is uttering more syllables in the same amount of time, or that the pronunciation of individual syllables is shorter. This phenomenon of syllable duration "compression" precisely reflects an increased speech rate, meaning that the user is speaking at a faster pace, thus allowing the entire conversation to contain more speech information within a limited time.

[0046] The specific steps for generating a reference value for syllable duration shortening after conducting in-depth analysis of the compression ratio of each syllable's duration relative to the standard duration within a unit of time for audio data of fixed duration are as follows:

[0047] In fixed-duration audio data, the speech signal is first segmented at the syllable level, and the actual duration of each syllable is measured. Then, phoneme boundary detection techniques or endpoint detection algorithms (such as Viterbi decoding based on time-frequency features) are used to obtain the start and end times at the syllable level. A reference syllable duration is defined, which can be based on a linguistically defined standard syllable duration (e.g., a typical syllable duration of 200-300ms in Mandarin Chinese) or syllable duration annotations from a high-confidence speech synthesis system. For the actual duration of a given syllable, its duration compression ratio is defined as follows:

[0048] ,in: For the first The duration compression ratio of a syllable indicates the degree of compression of the pronunciation speed of the current syllable relative to the standard syllable duration. The reference syllable duration is the typical syllable duration under standard pronunciation conditions. The actual duration of pronunciation of the current syllable. To avoid division by zero errors, a very small smoothing term is usually taken as... ;

[0049] When the duration of a single syllable is compressed significantly, it means that the syllable has been significantly compressed compared to the standard duration, which means that the speaker has completed the pronunciation of the syllable in a shorter time, reflecting an increased speech rate;

[0050] The main purpose of this step is to establish the duration compression ratio of individual syllables, provide basic data for subsequent calculation of syllable duration reduction reference values, and be able to more sensitively capture abnormal acceleration of individual syllables in order to cope with high speech rates in different languages ​​and contexts.

[0051] Within a fixed-duration audio segment, the compression ratios of multiple syllables are aggregated to construct a syllable duration reduction reference value that characterizes the overall speech rate. Based on the idea of ​​information entropy, a weighted sum is constructed using the exponential growth trend of the syllable duration compression ratio, making it more sensitive to high-speed speech. The expression for generating the syllable duration reduction reference value is as follows:

[0052] ,in: This serves as a reference value for syllable duration shortening, representing the overall speech rate of this fixed-duration audio segment. This represents the total number of syllables in the audio segment. This is a non-linear adjustment factor, with a value range of [value range missing]. This is used to control the sensitivity to exponential growth under high speech rates, and is usually set between 0.5 and 0.8.

[0053] The syllable duration shortening reference value serves as the final indicator of speech rate, effectively distinguishing between fast, normal, and slow speech segments. This syllable duration shortening reference value possesses a good dynamic range and non-linear mapping capability, meaning that the faster the speech rate, the more significant the increase in the syllable duration shortening reference value.

[0054] As can be seen from the syllable duration reduction reference value, for audio data of a fixed duration, a higher syllable duration reduction reference value, generated after in-depth analysis of the compression ratio of each syllable's duration relative to the standard duration within a unit of time, indicates a faster current speaking speed, and vice versa. This is because the syllable duration reduction reference value is calculated based on the compression ratio of each syllable's duration relative to the standard duration within a unit of time. When a user speaks faster, the pronunciation duration of each syllable shortens, resulting in a larger compression ratio relative to the standard syllable duration. This leads to a cumulative increase in the compression ratio of multiple syllables, causing the syllable duration reduction reference value to rise. Conversely, when a user speaks slower, the pronunciation duration of syllables tends to approach or exceeds the standard duration, resulting in a lower compression ratio and thus a lower syllable duration reduction reference value. Therefore, the syllable duration reduction reference value can serve as a reliable indicator for measuring speaking speed, directly reflecting the user's real-time speaking speed changes.

[0055] The degree to which an audio signal rapidly changes from positive to negative or vice versa within a unit of time is commonly referred to as the short-time zero-crossing rate. Specifically, when a user speaks very quickly, the speech segments in the audio signal often contain more phonemes or syllables within a short period. This significantly increases the number of oscillations in the speech signal per unit of time, manifesting as more frequent and denser changes in the audio waveform from positive to negative polarity, or vice versa. Therefore, in audio data of a fixed duration, if such a significantly increased and rapidly changing short-time zero-crossing rate is detected, it can be considered that the user's speaking speed is relatively fast, and is one of the important acoustic indicators that effectively measures and reflects the user's speaking speed.

[0056] The specific steps for generating short-time zero-crossing rate abrupt change reference values ​​after conducting in-depth analysis of the rapid changes in the audio signal from positive to negative and from negative to positive within a unit of time for audio data of fixed duration are as follows:

[0057] First, the audio signal of fixed duration is processed using a discrete sequence. This indicates that the window size should be set according to the preset window size (e.g., ...). W (each sampling point) is used to divide the frame, resulting in... The first, adjacent but non-overlapping frames. For the first... The sampling points in the frame, starting from the frame start index End of frame index The number of positive and negative polarity switching points is detected point by point, i.e., the number of zero crossings, and the short-time zero crossing rate of the frame is defined by the following formula:

[0058] ,in: Indicates the first The short-time zero-crossing rate of the frame, This is a sign function; it outputs 1 when the input is greater than 0, -1 when the input is less than 0, and can be considered as 0 when the input is equal to 0. The larger the value, the more frequently the speech signal within the frame switches between positive and negative polarities.

[0059] This step obtains the short-time zero-crossing rate sequence for each frame. This lays the foundation for subsequent judgment of speech rate abrupt changes. The purpose of this step is to quantify the zero-crossing phenomenon of the audio signal into a frame-level numerical description, thereby providing core data support for further constructing short-time zero-crossing rate abrupt change reference values.

[0060] After obtaining the frame-level short-time zero-crossing rate, its dramatic changes between adjacent frames are measured to capture the abrupt changes caused by high speech rate in numerical form. To this end, a reference value for the short-time zero-crossing rate abrupt change is calculated using the short-time zero-crossing rate. The expression for the calculation is as follows:

[0061] ,in: This serves as a short-term reference value for zero-crossing rate mutations. It is a magnification factor greater than zero, used to adjust the sensitivity to changes in the zero-crossing rate. This is used to enhance nonlinear mapping, so that even moderate differences in zero-crossing rates can be appropriately amplified. Indicates the first The short-time zero-crossing rate of a frame. When the difference in the number of zero-crossings between adjacent frames increases significantly, The value will also increase accordingly, thereby accumulating and raising the short-term zero-crossing rate mutation reference value. If the short-term zero-crossing rate mutation reference value increases significantly within a certain period of time, it indicates that the user is rapidly switching between positive and negative polarities at an extremely fast speaking speed, thus helping the system detect the "high speaking speed" state in real-time or offline analysis.

[0062] The core function of this step is to integrate the changes in the inter-frame zero-crossing rate into a mutation metric, which can more sensitively reflect the frequent waveform jumps caused by a sudden increase in speech rate.

[0063] As can be seen from the short-time zero-crossing rate (SZCR) reference value, for audio data of a fixed duration, a higher SZCR reference value, generated after in-depth analysis of the rapid changes in the audio signal from positive to negative and from negative to positive within a unit of time, generally indicates a faster speaking speed, and vice versa. This is because the SZCR reference value measures the degree of rapid switching between positive and negative polarities of the audio signal within a fixed duration. A faster speaking speed results in a higher frequency of phoneme transitions during pronunciation, leading to more drastic fluctuations in the zero-crossing rate of the waveform within a short period, thus increasing the cumulative value of the SZCR reference value. Conversely, at a slower speaking speed, users prolong syllables and pause more during pronunciation, resulting in smaller zero-crossing changes in the waveform and smaller differences in the zero-crossing rate between adjacent frames, causing the SZCR reference value to increase more slowly or even plateau. Therefore, the level of the SZCR reference value can be directly used as a characterization of speaking speed, providing a quantitative basis for real-time speech analysis.

[0064] A machine learning model that has been trained and deployed refers to a model whose performance and stability meet design requirements after a series of standardized steps including model building, data training, parameter tuning, and performance validation. It has been deployed in a real-world application environment and can stably process input data in real time, outputting reliable prediction results. Specifically, this process first involves collecting a large amount of clearly labeled historical audio data, especially covering various speech rate scenarios (including normal, slow, medium, fast, and even extremely fast speech rates), and clearly labeling the accurate speech rate level or value corresponding to each audio data segment. Then, based on the collected audio data, specialized speech rate-related features, such as syllable duration shortening reference values ​​and short-term zero-crossing rate mutation reference values, are used as input features to build a model suitable for audio speech rate prediction tasks. This could be a neural network model (such as CNN, LSTM, or Transformer models) or a traditional machine learning model (such as random forest or support vector machine). During the training phase, iterative optimization is needed to gradually correct model weights and parameters, reduce prediction errors, and improve the model's generalization ability and prediction accuracy until the model's performance on the validation dataset reaches or exceeds the expected target.

[0065] Once the training process described above is complete, the model is capable of processing real-world application data and consistently outputting accurate predictions. At this point, the model is considered a "trained and deployed machine learning model." In the actual deployment phase, this model is encapsulated as an independent, callable module and deployed on the actual application system or server. When a user makes a call, the system collects the user's audio signal in real time. It then quickly performs data preprocessing and feature extraction, such as extracting reference values ​​for syllable duration shortening and short-term zero-crossing rate mutations—two feature vectors specifically reflecting speech rate trends—from the audio. These vectors are then input into the trained machine learning model in real time. Based on the data distribution, patterns, and feature representations learned during training, the model rapidly calculates and infers from the currently input feature data, outputting a real-time audio rate evaluation coefficient that quantifies the user's current speech rate. This coefficient allows the model to quickly identify, judge, and predict the user's real-time speech rate. 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 that the model's application capabilities and practical value have been truly realized, becoming an indispensable part of the actual operation of the system, and effectively achieving the goal of real-time detection and intelligent response of user speech rate.

[0066] The machine learning model is not specifically limited here, but it can achieve a reference value for shortening syllable duration. and short-time zero-crossing rate mutation reference value A comprehensive analysis is performed to generate audio rate evaluation coefficients. Any machine learning model can be used. To achieve the technical solution of this invention, this invention provides a specific implementation method; audio rate evaluation coefficient. The generated expression is: In the formula, , These are the reference values ​​for syllable duration shortening. and short-time zero-crossing rate mutation reference value The preset proportional coefficient, and , All are greater than 0. Preset proportionality coefficient (i.e., in the formula) , ) refers to the evaluation coefficient of the generated audio rate ( When ), a reference value for syllable duration shortening is given respectively ( ) and short-time zero-crossing rate mutation reference value ( These coefficients represent the weights or contributions of different features. Their purpose is to adjust and control the degree of influence of each feature on the final generated result, reflecting the differences in importance between features and ensuring that the final audio rate assessment coefficient more accurately reflects the actual speech rate. Typically, these coefficients are obtained through experimental verification, data training, and optimization, representing the importance of specific features in the combined model. For example, if experiments show that the syllable duration shortening reference value contributes more to speech rate judgment, a larger coefficient should be set to increase its influence; conversely, if the short-term zero-crossing rate mutation reference value contributes less, the coefficient can be relatively lower. In short, setting the coefficients is to ensure that the final audio rate assessment coefficient more accurately reflects the actual user's speech rate, improving the accuracy of speech rate judgment and the reliability of model prediction.

[0067] According to the audio rate evaluation coefficient, for audio data of a fixed duration, the larger 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 within a unit time, the larger the performance value of the short-time zero-crossing rate mutation reference value generated after in-depth analysis of the rapid and drastic changes of the audio signal from positive to negative and from negative to positive within a unit time, the larger the performance value of the audio rate evaluation coefficient generated by the pre-trained machine learning model to predict the user's current speech rate in real time, the faster the user's current speech rate, and vice versa.

[0068] The audio rate evaluation coefficient, generated by a pre-trained machine learning model that predicts the user's current speech rate in real time, is compared with a pre-set reference threshold for audio rate evaluation coefficients to identify users whose speech rate exceeds the normal range. The specific steps are as follows:

[0069] If the audio rate evaluation coefficient is greater than the audio rate evaluation coefficient reference threshold, the current user's speech rate is identified as exceeding the normal range, meaning the current user's speech rate is too fast; if the audio rate evaluation coefficient is less than or equal to the audio rate evaluation coefficient reference threshold, the current user's speech rate is identified as being within the normal range, meaning the current user's speech rate is normal.

[0070] Adaptive ASR dynamic adjustment module: If the machine learning model identifies that the current user's speech rate exceeds the normal range, it will automatically trigger the predefined adaptive ASR dynamic adjustment mechanism, activate the real-time adaptive strategy of the backend ASR engine, and automatically adjust the original speech segment duration (such as reducing it from 1 second to 0.5 seconds or even shorter) to perform fine segmentation of audio data.

[0071] The predefined adaptive ASR dynamic adjustment mechanism is automatically triggered, activating the real-time adaptive strategy of the backend ASR engine and automatically adjusting the original speech segment duration. The specific steps are as follows:

[0072] When the audio rate evaluation factor is detected to be greater than the audio rate evaluation factor reference threshold, the system will automatically shorten the segment length of the ASR engine within the next fixed time period. The specific expression for shortening the ASR engine segment length is as follows:

[0073] ,in: The duration of the current segment (e.g., 1 second). The rate coefficient is used to control the degree of segment shortening; This represents the minimum acceptable segment length (to avoid excessive shortening that could impact real-time performance). This refers to the shortened segment slice length of the ASR engine;

[0074] The larger the difference between the detected audio rate evaluation coefficient and the audio rate evaluation coefficient reference threshold, the higher the speech rate, and the more significant the segment shortening; conversely, the shortening is relatively weaker. Through this step, the system can segment 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 a real-time, dynamic processing strategy, ensuring that the system can quickly adapt when the user's speech rate suddenly increases, guaranteeing the timely capture and recognition of key information (such as emergency help keywords), and preventing missed or misjudgments caused by incorrect splicing or ignoring of large segments of speech at high speeds.

[0075] Refined speech slicing and decoding module: Based on the control instructions of the previous step, the backend ASR engine performs refined segmentation and slicing of the user's speech signal in a shorter time window, and then decodes each segmented speech independently. This makes the content of each speech slice clearer and reduces the cross-segment confusion caused by excessively long segments and too much content.

[0076] Keyword detection and speech recognition module: After dynamically segmenting speech into segments, each fine-grained speech segment is independently fed into a real-time ASR decoding unit for speech recognition and conversion into text content. Simultaneously, based on a pre-defined emergency keyword lexicon (such as "help," "fire," "first aid," etc.), keywords are detected and labeled in real time for each speech segment. Once key emergency words are detected, the weight and priority of that segment are immediately increased, triggering the emergency assistance process.

[0077] Semantic understanding and intent analysis module: Although shortening speech segments improves accuracy, the integrity of the contextual semantics still needs to be considered. The system then fuses the text recognition results of multiple consecutive short speech segments in real time, and performs comprehensive semantic understanding analysis through the Natural Language Understanding (NLU) unit to determine the user's overall intent and specific emergency situation. For example, contextual fusion can be used to derive the user's specific request for help, location information, and other important emergency information.

[0078] Emergency Response and Dispatch Module: When it is confirmed that a user's request for help is indeed an emergency, the pre-defined emergency response mechanism is quickly triggered, automatically transferring the call to a live agent or emergency rescue hotline (such as 120, 119, or 110), or automatically pushing alarm information, location, and request details to the appropriate department. Furthermore, to avoid any risks caused by AI misjudgment, the system should establish a fallback mechanism: when the recognition result has low confidence or uncertainty, the user should be automatically and quickly transferred to a live agent to prevent delays in rescue.

[0079] The implementation of the above solution significantly improves the accuracy and timeliness of AI telephone response systems in recognizing emergency calls. By monitoring changes in the user's speech rate in real time and adaptively shortening the segment length of speech segments, the granularity of audio data processing is made more refined. This greatly reduces the problems of cross-segment information confusion, missed recognition, and misrecognition caused by excessively fast speech rates under traditional fixed segment lengths. Specifically, by quickly and accurately capturing the speech characteristics of users at high speech rates and dynamically adjusting the speech segment length of the ASR engine, each speech segment becomes clearer and more accurate, avoiding cross-segment word fragmentation and recognition errors caused by excessively fast speech rates. This effectively avoids the risk of missing or misjudging users' urgent information in emergency call scenarios. Ultimately, the successful application of this method ensures that information in emergency situations can be identified and responded to quickly and reliably, significantly reducing the time it takes for medical personnel, firefighters, or other rescue forces to reach the accident site. It also 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 response delays, and significantly improving the safety and reliability of AI telephone response systems in emergency situations.

[0080] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0081] The foregoing has only described certain exemplary embodiments of the present invention by way of illustration. Undoubtedly, 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 foregoing 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.

[0082] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. An AI-based telephone answering system, characterized in that, It includes a real-time speech acquisition and preprocessing module, a speech rate detection and prediction module, an adaptive ASR dynamic adjustment module, a fine speech slicing and decoding module, a keyword detection and speech recognition module, a semantic understanding and intent analysis module, and an emergency response and dispatch module; The real-time voice acquisition and preprocessing module acquires user voice data from emergency calls in real time, establishes a real-time audio stream transmission channel, and performs rapid preprocessing on the acquired audio signals. The speech rate detection and prediction module quickly extracts audio data features from audio data of a fixed duration, performs in-depth analysis of the extracted speech rate features, and then uses a pre-trained machine learning model to predict the user's current speech rate in real time. The adaptive ASR dynamic adjustment module activates the real-time adaptive strategy of the backend ASR engine, automatically adjusts the original speech segment duration, and performs fine segmentation and cutting of audio data. The fine-grained speech slicing and decoding module receives control commands, performs fine-grained segmentation and slicing of the user's speech signal, and then decodes each segmented speech independently. The keyword detection and speech recognition module independently sends each fine speech slice into the real-time ASR decoding unit for speech recognition and conversion into text content. At the same time, based on a predefined emergency keyword lexicon, it performs real-time keyword detection and labeling on each speech slice content. The semantic understanding and intent analysis module fuses the text recognition results of multiple consecutive short speech segments in real time, and performs comprehensive semantic understanding analysis through the natural language understanding unit to determine the user's overall intent and specific emergency situation; The emergency response and dispatch module, when it is confirmed that the user's request for help is indeed an emergency, quickly triggers the predefined emergency response mechanism, automatically transfers the call to human customer service or emergency rescue hotline, or automatically pushes alarm information, location, and request for help to the appropriate department; Rapid audio data features are extracted from fixed-duration audio data. These features include the compression ratio of each syllable duration relative to the standard duration per unit time and the intensity of rapid changes in the audio signal from positive to negative or from negative to positive per unit time. After in-depth analysis of the extracted speech rate features, syllable duration shortening reference values ​​and short-term zero-crossing rate mutation reference values ​​are generated. These syllable duration shortening reference values ​​and short-term zero-crossing rate mutation reference values ​​are used as feature vectors and input into a trained and deployed machine learning model. The model generates audio rate evaluation coefficients, and the user's current speech rate is predicted in real time based on these coefficients.

2. The AI-based telephone answering system according to claim 1, characterized in that, Audio signals are acquired using high-quality microphone equipment and communication links at a set sampling frequency and bit depth.

3. The AI-based telephone answering system according to claim 1, characterized in that, Fixed-duration 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, The audio rate evaluation coefficient, generated by a pre-trained machine learning model that predicts the user's current speech rate in real time, is compared with a pre-set reference threshold for audio rate evaluation coefficients to identify users whose 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 speech rate is identified as exceeding the normal range; if the audio rate evaluation coefficient is less than or equal to the audio rate evaluation coefficient reference threshold, the current user's speech rate is identified as within the normal range.

5. The AI-based telephone answering system according to claim 1, characterized in that, The specific steps for generating a reference value for syllable duration shortening after conducting in-depth analysis of the compression ratio of each syllable's duration relative to the standard duration within a unit of time for audio data of fixed duration 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. Phoneme boundary detection technology is used to obtain the start and end times at the syllable level. A reference syllable duration is set, and the duration compression ratio for a given syllable is defined as follows: ,in: For the first i The duration compression ratio of a syllable indicates the degree of compression of the pronunciation speed of the current syllable relative to the standard syllable duration. The reference syllable duration is the typical syllable duration under standard pronunciation conditions. The actual duration of pronunciation of the current syllable. To create a smooth term and avoid division by zero errors; Within a fixed-duration audio segment, the duration compression ratios of multiple syllables are aggregated to construct a syllable duration reduction reference value that characterizes the overall speech rate. The expression for constructing the syllable duration reduction reference value is as follows: ,in: This serves as a reference value for syllable duration shortening, representing the overall speech rate of this fixed-duration audio segment. N This represents the total number of syllables in the audio segment. This is a non-linear adjustment factor, with a value range of [value range missing]. This is used to control the sensitivity to exponential growth under high speech rates.

6. The AI-based telephone answering system according to claim 1, characterized in that, The specific steps for generating short-time zero-crossing rate abrupt change reference values ​​after conducting in-depth analysis of the rapid changes in audio signals from positive to negative or from negative to positive within a unit of time for audio data of fixed duration are as follows: A fixed-duration audio signal is represented by a discrete sequence. This is represented and divided into frames according to a pre-set window size, resulting in... M The adjacent but non-overlapping frames; for the first k The sampling points in the frame, starting from the frame start index End of frame index The number of switching between positive and negative polarities is detected point by point, i.e., the number of zero crossings, and the number of zero crossings is defined by the following formula: k The short-time zero-crossing rate of a frame is defined by the following expression: ,in: Indicates the first k The short-time zero-crossing rate of the frame, It is a symbolic function; After obtaining the frame-level short-time zero-crossing rate, the short-time zero-crossing rate abrupt reference value is calculated using the short-time zero-crossing rate. The calculation expression is as follows: ,in: This serves as a short-term reference value for zero-crossing rate mutations. This is a positive amplification factor used to adjust the sensitivity to changes in the zero-crossing rate. Indicates the first k -1 frame short-time zero-crossing rate.

7. The AI-based telephone answering system according to claim 6, characterized in that, The predefined adaptive ASR dynamic adjustment mechanism is automatically triggered, activating the real-time adaptive strategy of the backend ASR engine and automatically adjusting the original speech segment duration. The specific steps are as follows: When the audio rate evaluation coefficient is detected to be greater than the audio rate evaluation coefficient reference threshold, the segment slice length of the ASR engine is automatically shortened within the next fixed duration. The specific expression for shortening the ASR engine segment slice is as follows: ,in: This represents the current segment duration. The rate coefficient is used to control the degree of segment shortening; This represents the minimum acceptable segment length. This refers to the shortened segment slice length of the ASR engine; This is the audio rate evaluation coefficient.

Images