Real-time processing method based on voice SMS

Abstract

This invention discloses a real-time processing method based on voice SMS, relating to the field of communication technology. The specific steps of the method are: voice preprocessing, voice transcription and text standardization, semantic analysis and processing, information processing, and result feedback. This invention achieves dynamic noise reduction and precise segmentation of voice signals through an adaptive filtering formula in voice preprocessing and a segmentation mechanism based on voice pause intervals and intonation changes. It can analyze the frequency and amplitude of noise signals and the corresponding characteristics of voice signals in real time. By dynamically adjusting the filtering coefficients through the adaptive filtering formula, it effectively filters out non-voice signal components while preserving complete voice information. By capturing pause intervals and intonation changes in the voice signal, it accurately segments continuous voice streams, forming standardized voice segments with timestamps and sequence identifiers, providing a high-quality data foundation for subsequent voice transcription and semantic analysis.

Description

Technical Field

[0001] This invention relates to the field of communication technology, specifically to a real-time processing method based on voice text messages. Background Technology

[0002] In today's information society, with the rapid development of mobile communication technology, voice messaging, as an efficient and convenient communication method, has been widely used in people's daily life and work. Users can quickly convey information through voice messaging, especially in scenarios such as driving, exercising, or when hands are busy. Voice input shows greater convenience and security compared to text input. However, the real-time processing and efficient parsing of voice messages has become a major challenge.

[0003] However, traditional methods often use fixed filtering coefficients or preset noise templates, which not only fail to dynamically adjust parameters based on real-time noise characteristics (such as frequency and amplitude), leading to the retention of non-speech signals (such as sudden noise) or the accidental deletion of effective speech components, affecting the accuracy of speech segmentation, but also fail to recognize intonation changes or semantic pauses, resulting in incomplete speech segments after segmentation. Furthermore, traditional methods lack the ability to correlate speech with context, easily producing problems such as typos and incoherent sentences, and cannot correct errors based on context, resulting in poor text coherence. At the same time, the lack of integration of timestamps and sequence identifiers leads to logical chaos in the concatenated text, making it difficult to support subsequent semantic analysis. Moreover, in voice text messages, the timely transmission and processing of emergency information is crucial, but traditional methods often fail to accurately identify emergency keywords in the speech, and cannot trigger corresponding prompt signals according to the degree of urgency, resulting in emergency information not being processed in a timely manner. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of existing technologies and provide a real-time processing method based on voice SMS. This invention achieves dynamic noise reduction and accurate segmentation of voice signals through an adaptive filtering formula in voice preprocessing and a segmentation mechanism based on voice pause intervals and intonation changes. It can analyze the frequency and amplitude of noise signals and the corresponding features of voice signals in real time. By dynamically adjusting the filtering coefficients through the adaptive filtering formula, non-voice signal components are effectively filtered out while retaining complete voice information. By capturing pause intervals and intonation changes in the voice signal, continuous voice streams are accurately segmented to form standardized voice segments with timestamps and sequence identifiers. This provides a high-quality data foundation for subsequent speech transcription and semantic analysis, significantly improving the overall quality of voice processing.

[0005] To solve the above-mentioned technical problems, the present invention provides the following technical solution: a real-time processing method based on voice SMS messages, the specific steps of which are as follows:

[0006] Voice preprocessing: Real-time reception of voice messages, analysis of noise and voice signals and adjustment of filtering coefficients, noise reduction of voice messages, and segmentation according to voice pause intervals and intonation changes to form voice segments with timestamps and sequence identifiers.

[0007] Speech transcription and text standardization: The Transformer architecture deep learning model is called to extract features from speech segments and encode them, then transcribe them into text. The text is compared with the corpus to obtain the matching degree, the context is analyzed to obtain the semantic coherence, and the corrected confidence is calculated. If it is lower than the set value, the text is corrected. Then the text is concatenated according to the sequence identifier and the format is standardized.

[0008] Semantic analysis and processing: Natural language processing technology is used to perform lexical analysis, syntactic analysis and semantic role labeling on the transcribed and corrected text, extract keywords, and use a two-layer processing model to infer and complete the omitted content;

[0009] Information processing: Extract the core information related to time, location and event from the semantically processed text, calculate the urgency level using a preset emergency keyword library and an urgency calculation formula, and trigger the corresponding emergency alert signal;

[0010] Results feedback: The processing results will be fed back to the user in the form of voice broadcast and text, including the completion of annotations and inference information.

[0011] Furthermore, in the speech preprocessing, voice messages are received in real time, and noise and speech signals in the voice messages are detected and analyzed. Based on the frequency and amplitude of the noise and the frequency and amplitude of the speech signal, the filtering coefficients are dynamically adjusted using an adaptive filtering formula to filter out non-speech signal components. At the same time, by capturing pauses and intonation changes in the speech signal, the continuous speech stream is segmented into multiple independent speech segments. Each segment corresponds to a complete semantic unit, and a timestamp and sequence identifier are added to each speech segment to form standardized speech segments that can be directly used for transcription.

[0012] Furthermore, in the speech preprocessing, the filter coefficients are dynamically adjusted using an adaptive filtering formula, which is as follows: ,in, These are the filter coefficients at the current moment. The value is the filter coefficient from the previous time step. This is the filter coefficient adjustment factor. This represents the peak frequency of the current noise signal. The current noise signal amplitude, This is the dominant frequency of the current voice signal. The amplitude of the current speech signal. It is a positive constant.

[0013] Furthermore, in the speech-to-text standardization process, a Transformer architecture deep learning model is used to extract features from speech segments, encode the extracted features, input the encoded features into the deep learning model to generate corresponding text content, and compare the generated text with a pre-set corpus to determine the matching degree between the text and the corpus. Simultaneously, the context of the text is analyzed to determine the semantic coherence of the context. The corrected confidence level is calculated using the text correction confidence level formula. When adjusting the confidence level Below the set value At that time, typos and grammatical errors in the text were corrected, and the text of each segment was spliced ​​together according to the timestamp and sequence identifier of the audio segments to form the first draft of the text. The first draft of the text was then formatted, including adding appropriate punctuation marks, adjusting paragraph spacing, and correcting capitalization.

[0014] Furthermore, in the aforementioned speech-to-text transcription and text standardization, the corrected confidence level is calculated using a text correction confidence level formula, which is as follows: ,in, To adjust the confidence level, For the matching degree between text and corpus, For contextual semantic coherence, These are the weight parameters.

[0015] Furthermore, in the semantic analysis and processing, natural language processing technology is used to perform lexical analysis, syntactic analysis, and semantic role labeling on the transcribed and corrected text. The lexical analysis segments the text into words and labels them with parts of speech. The syntactic analysis clarifies the syntactic relationships between words. The semantic role labeling identifies the semantic roles of the core verbs in the sentence and the agent, patient, time, and place. Keyword scores are calculated using a keyword extraction formula, and the top 10 keywords with the highest scores are extracted. Then, through a two-layer processing model, the omitted content is deduced and completed. The first layer uses keywords as a reference and compares them with the context to identify and correct missing words in the text. The corrected content is marked with a specific color. The second layer constructs an initial semantic framework based on the corrected text. The initial semantic framework includes the basic grammatical relationships of subject-predicate, verb-object, and modifier. The content is deduced and completed, and the deduced and completed content is marked with another specific color.

[0016] Furthermore, in the semantic analysis and processing, keyword scores are calculated using a keyword extraction formula, which is as follows: ,in, For vocabulary Keyword score, The part-of-speech contribution value is determined based on lexical analysis and part-of-speech tagging. The semantic role score is set based on the semantic role annotation results. This represents the frequency of syntactic associations.

[0017] Furthermore, in the information processing, an urgency calculation formula is used to calculate the degree of urgency. The urgency calculation formula is as follows: ,in This is the urgency value, ranging from 0 to 100. For the first The weight of an urgent keyword. The number of times the keyword appears. For the first The degree of relevance between each keyword and the overall semantics of the text;

[0018] When the urgency value When the emergency level is ≥80, it is determined to be extremely urgent, triggering a continuous, high-frequency, rapid ringing sound with the vibration intensity set to maximum. When the value is ≥50 and <80, it is judged as a high urgency level, triggering a medium-frequency ringtone with moderate vibration intensity. When the emergency level is ≥30 and <50, it is judged as medium emergency, triggering a low-frequency ringtone and slight vibration. When the threshold is <30, it is determined to be a low-urgency level, triggering only a slight vibration.

[0019] Compared with existing technologies, this real-time processing method based on voice SMS has the following advantages:

[0020] I. This invention achieves dynamic noise reduction and precise segmentation of speech signals through an adaptive filtering formula in speech preprocessing and a segmentation mechanism based on speech pause intervals and intonation changes. It can analyze the frequency and amplitude of noise signals and the corresponding features of speech signals in real time. By dynamically adjusting the filtering coefficients through the adaptive filtering formula, it effectively filters out non-speech signal components while retaining complete speech information. By capturing pause intervals and intonation changes in the speech signal, it accurately segments continuous speech streams to form standardized speech segments with timestamps and sequence identifiers. This provides a high-quality data foundation for subsequent speech transcription and semantic analysis, significantly improving the overall quality of speech processing.

[0021] Second, this invention utilizes the Transformer architecture deep learning model and natural language processing technology to achieve context-aware correction of transcribed text and real-time identification and processing of emergency information. By using the Transformer architecture deep learning model to extract and encode features from speech segments, combined with contextual semantic analysis, the transcribed text is intelligently corrected, resolving issues such as typos and incoherent sentences, ensuring the accuracy and coherence of the text content. Furthermore, an emergency keyword database is constructed, and through an emergency degree calculation formula, the urgency level can be calculated in real time based on factors such as the frequency of emergency keywords in the text and their relevance to the overall semantics, triggering corresponding emergency alert signals. This ensures that emergency information is processed promptly, greatly enhancing the accuracy and timeliness of information transmission.

[0022] Other advantages, objectives and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination or study, or may be learned from the practice of the invention. Attached Figure Description

[0023] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0024] Figure 1 This is a flowchart of a real-time processing method based on voice SMS.

[0025] Figure 2 This is a framework diagram of a real-time processing method based on voice SMS.

[0026] Figure 3 This is a flowchart illustrating the speech-to-text standardization process for real-time processing of voice SMS messages. Detailed Implementation

[0027] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0028] Example 1

[0029] Voice preprocessing: In an emergency medical assistance scenario, the voice message sent by the user when dialing the emergency hotline is received in real time, and the noise in it (such as the honking of vehicles and the noise of the crowd in the background) and the user's voice signal are detected and analyzed. According to the frequency and amplitude dynamics of the noise and the voice signal, the filtering coefficient is dynamically adjusted through an adaptive filtering formula to filter out these noises. The adaptive filtering formula is: , where is the filtering coefficient at the current moment, is the filtering coefficient at the previous moment, is the filtering coefficient adjustment factor, is the peak frequency of the current noise signal, is the amplitude of the current noise signal, is the main frequency of the current voice signal, is the amplitude of the current voice signal, is a positive constant; at the same time, the pauses (such as the pauses caused by rapid breathing) and intonation changes (such as the intonation rising due to pain) in the user's voice are captured, and the continuous voice stream is segmented into multiple independent voice segments. Each segment corresponds to a complete semantic unit (such as "My chest hurts" and "In a certain community"), and a timestamp and a sequence identifier are added to each segment to form a standardized voice segment.

[0030] Voice transcription and text normalization: Call the deep learning model of the Transformer architecture to extract features and encode the standardized voice segments, and then transcribe them into text. Compare the generated text with a preset corpus (including common diseases, first aid terms, etc.) to determine the matching degree , and analyze the context (such as mentioning symptoms first and then the location) to determine the semantic coherence , and calculate the correction confidence through the text correction confidence formula. The text correction confidence formula is: , where is the correction confidence, is the matching degree between the text and the corpus, is the semantic coherence of the context, is the weight parameter, and the correction confidence lower than the set value , correct the misspelled words (such as correcting "xiong kou" to "chest") and unsmooth sentences in the text, and then splice the text according to the timestamp and sequence identifier. After forming a draft, perform format normalization and add punctuation marks (such as "My chest hurts, in Building 3 of a certain community.").

[0031] Semantic analysis and processing: Natural language processing techniques are used to perform lexical analysis (segmenting and tagging the parts of speech of words such as "chest," "pain," and "community"), syntactic analysis (clarifying the subject-predicate relationship between "chest" and "pain"), and semantic role labeling (identifying the core verb "pain," labeling "I" as the agent and "chest" as the patient). Keyword scores are calculated using a keyword extraction formula, which is as follows: ,in, For vocabulary Keyword score, The part-of-speech contribution value is determined based on lexical analysis and part-of-speech tagging. The semantic role score is set based on the semantic role annotation results. Syntactic association frequency reflects the closeness of the association between words and core syntactic components. The top 10 keywords with the highest scores (such as "chest pain", "a certain community", "difficulty breathing") are extracted and the omitted content is completed through a two-layer processing model. The first layer refers to the keywords and context to identify and correct missing words in the text. The corrected content is marked in red. The second layer constructs an initial semantic framework based on the corrected text and completes the omitted information (such as completing "may require cardiopulmonary resuscitation related first aid measures" based on "chest pain" and "difficulty breathing"), which is marked in blue.

[0032] Information Processing: Extract core information from the semantically processed text (time: real-time reception time; location: Building 3 of a certain community; event: user experiencing chest pain and difficulty breathing). Compare the extracted core information with a pre-defined emergency keyword database (containing emergency keywords such as "chest pain," "difficulty breathing," and "first aid"). Calculate the urgency level using an urgency calculation formula, which is as follows: ,in This is the urgency value, ranging from 0 to 100. For the first The weight of an urgent keyword. The number of times the keyword appears. For the first The degree of relevance between each keyword and the overall semantics of the text is used to derive an urgency value. If the value is greater than 80, it is considered an extremely high level of urgency, triggering a continuous, high-frequency, rapid ringing sound and setting the vibration intensity to maximum.

[0033] Feedback on Results: The processing results will be fed back to the emergency center staff in both voice broadcast (e.g., "Emergency assistance received. The user is in Building 3 of XX Community, experiencing chest pain and difficulty breathing, and may require CPR. The urgency level is extremely high") and text format (including the corrections marked in red and the supplementary information marked in blue).

[0034] In summary, in emergency medical assistance scenarios, by denoising, segmenting, and labeling the received voice requests for help, using a Transformer architecture deep learning model to transcribe the text and standardize its format, then extracting key information and completing the content through semantic analysis, calculating the urgency level based on an emergency keyword database and triggering corresponding prompts, and finally feeding back the results, including corrections and completion information, to emergency personnel in both voice and text formats, this provides highly efficient emergency medical assistance.

[0035] Example 2

[0036] Voice preprocessing: In logistics delivery notification scenarios, voice text messages about delivery information sent by couriers are received in real time. Noise (such as packing noise at the delivery point, vehicle engine noise) and the courier's voice signal are detected and analyzed. Based on the frequency and amplitude dynamics of the noise and voice signal, an adaptive filtering formula is used. The system dynamically adjusts the filtering coefficients to filter out this noise, while capturing pauses (such as pauses when announcing addresses) and tone changes in the courier's speech. It then segments the continuous speech stream into multiple independent speech segments (such as "The package will be delivered this afternoon" and "The recipient is Mr. Zhang"), and adds a timestamp and sequence identifier to each segment to form standardized speech segments.

[0037] Speech-to-text transcription and text standardization: A Transformer architecture deep learning model is used to extract features from standardized speech segments, encode them, and then transcribe them into text. The text is then compared with a pre-set corpus (containing common address expressions, delivery terms, etc.) to determine the matching degree. And analyze the context (e.g., state the delivery time before mentioning the recipient) to determine semantic coherence. The confidence formula is corrected using text: Calculate the adjusted confidence level to obtain the adjusted confidence level. Below the set value Correct typos (e.g., change "Zhang Xiansen" to "Zhang Xiansheng") and awkward sentences, assemble the text according to timestamps and sequence identifiers, and after forming the first draft, standardize the format and add punctuation (e.g., "The package will be delivered in the afternoon, and the recipient is Zhang Xiansheng.").

[0038] Semantic analysis and processing: Natural language processing techniques are used to perform lexical analysis (segmenting and tagging parts of speech for words such as "package," "afternoon," and "Mr. Zhang"), syntactic analysis (clarifying the subject-verb relationship between "package" and "deliver"), and semantic role labeling (identifying the core verb "deliver," labeling "package" as the patient, and "afternoon" as the time). Keyword extraction formulas are then used. The algorithm calculates keyword scores and extracts the top 10 keywords (such as "parcel", "afternoon", "Mr. Zhang", "delivery", etc.). It then uses a two-layer processing model to complete the omitted content. The first layer identifies and corrects missing words in the text by referring to the keywords and context. The corrected content is marked in yellow. The second layer builds an initial semantic framework based on the corrected text and completes the omitted information (such as completing "please keep the recipient's phone accessible" based on "delivery"), which is marked in green.

[0039] Information Processing: Extract core information (time: afternoon; location: delivery address; event: package will be delivered to Mr. Zhang) from the semantically processed text. Compare the extracted core information with a pre-defined emergency keyword database (containing emergency keywords such as "delayed" and "undeliverable"), and calculate the urgency using the following formula: Calculate the urgency level and obtain the urgency value. When the threshold is <30, it is determined to be a low-urgency level, triggering only a slight vibration.

[0040] Feedback on Results: The processing results will be communicated to the recipient, Mr. Zhang, via voice announcement (e.g., "Delivery notification received. The package will be delivered to Mr. Zhang this afternoon. Please keep your phone accessible") and text (including corrections highlighted in yellow and supplementary information highlighted in green).

[0041] In summary, in logistics delivery notification scenarios, after noise reduction, segmentation, and labeling of the courier's voice notification, the text is generated and corrected using a Transformer architecture deep learning model. Semantic analysis is then used to extract key delivery information and complete the content. The urgency level is determined based on an emergency keyword database, triggering corresponding prompts. Finally, the results, including corrections and completion information, are fed back to the recipient in both voice and text formats, effectively improving the processing efficiency and accuracy of logistics delivery notifications.

[0042] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A real-time processing method based on voice SMS, characterized in that, The specific steps of this method are as follows: Voice preprocessing: Real-time reception of voice messages, analysis of noise and voice signals, and noise reduction processing of voice messages. At the same time, the voice messages are segmented according to the pause interval and intonation changes to form voice segments with timestamps and sequence identifiers. Speech transcription and text standardization: The Transformer architecture deep learning model is called to extract features from speech segments and encode them, then transcribe them into text. The text is compared with the corpus to obtain the matching degree, the context is analyzed to obtain the semantic coherence, and the corrected confidence is calculated. If it is lower than the set value, the text is corrected. Then the text is concatenated according to the sequence identifier and the format is standardized. The revised confidence level is calculated using the text revised confidence level formula, which is: ,in, To adjust the confidence level, For the matching degree between text and corpus, For contextual semantic coherence, They are respectively , The weight parameters, and ; Semantic analysis and processing: Natural language processing technology is used to perform lexical analysis, syntactic analysis and semantic role labeling on the transcribed and corrected text, extract keywords, and use a two-layer processing model to infer and complete the omitted content; Information Processing: Extract core information related to time, location, and event from the semantically processed text. Using a pre-defined emergency keyword library and an urgency calculation formula, calculate the urgency level and trigger corresponding emergency alert signals. The urgency calculation formula is as follows: ,in, This is the urgency value, ranging from 0 to 100. For the first The weight of an urgent keyword. The number of times the keyword appears. For the first The degree of relevance between each keyword and the overall semantics of the text; Results feedback: The processing results will be fed back to the user in the form of voice broadcast and text, including the completion of annotations and inference information.

2. The real-time processing method based on voice SMS according to claim 1, characterized in that, In the speech preprocessing, voice messages are received in real time, and noise and speech signals in the voice messages are detected and analyzed. Based on the frequency and amplitude of the noise and the frequency and amplitude of the speech signal, non-speech signal components are filtered out. At the same time, by capturing pauses and intonation changes in the speech signal, the continuous speech stream is segmented into multiple independent speech segments. Each segment corresponds to a complete semantic unit, and a timestamp and sequence identifier are added to each speech segment to form standardized speech segments that can be directly used for transcription.

3. The real-time processing method based on voice SMS according to claim 1, characterized in that, In the aforementioned speech-to-text transcription and text standardization process, a Transformer architecture deep learning model is used to extract features from speech segments. These extracted features are then encoded, and the encoded features are input into the deep learning model to generate corresponding text content. The generated text is then compared with a pre-set corpus to determine the degree of matching between the text and the corpus. Simultaneously, the context of the text is analyzed to determine the semantic coherence of the context. The corrected confidence level is calculated using the text correction confidence level formula. When adjusting the confidence level Below the set value At that time, typos and grammatical errors in the text were corrected, and the text of each segment was spliced ​​together according to the timestamp and sequence identifier of the audio segments to form the first draft of the text. The first draft of the text was then formatted, including adding appropriate punctuation marks, adjusting paragraph spacing, and correcting capitalization.

4. The real-time processing method based on voice SMS according to claim 1, characterized in that, In the semantic analysis and processing, natural language processing technology is used to perform lexical analysis, syntactic analysis, and semantic role labeling on the transcribed and corrected text. The lexical analysis segments the text into words and labels their parts of speech. The syntactic analysis clarifies the syntactic relationships between words. The semantic role labeling identifies the semantic roles of the core verbs in the sentence, as well as the agent, patient, time, and place. Keyword scores are calculated using a keyword extraction formula, and the top 10 keywords with the highest scores are extracted. Then, a two-layer processing model is used to infer and complete the omitted content. The first layer uses keywords as a reference and compares them with the context to identify and correct missing words in the text. The corrected content is marked with a specific color. The second layer constructs an initial semantic framework based on the corrected text. The initial semantic framework includes the basic grammatical relationships of subject-predicate, verb-object, and modifier. The content is then inferred and completed. The inferred and completed content is marked with another specific color.

5. The real-time processing method based on voice SMS according to claim 1, characterized in that, In the semantic analysis and processing, keyword scores are calculated using a keyword extraction formula, which is as follows: ,in, For vocabulary Keyword score, The part-of-speech contribution value is determined based on lexical analysis and part-of-speech tagging. The semantic role score is set based on the semantic role annotation results. This represents the frequency of syntactic associations.

6. The real-time processing method based on voice SMS according to claim 1, characterized in that, In the aforementioned information processing, when the urgency value When the emergency level is ≥80, it is determined to be extremely urgent, triggering a continuous, high-frequency, rapid ringing sound with the vibration intensity set to maximum. When the value is ≥50 and <80, it is judged as a high urgency level, triggering a medium-frequency ringtone with moderate vibration intensity. When the emergency level is ≥30 and <50, it is judged as medium emergency, triggering a low-frequency ringtone and slight vibration. When the threshold is <30, it is determined to be a low-urgency level, triggering only a slight vibration.

Images