Supplier voice communication assistance analysis system and method for logistics

Abstract

This application discloses a supplier voice communication auxiliary analysis system and method for logistics, belonging to the field of voice communication assistance. This application is based on intelligent agents automatically answering calls, acquiring voice signals during voice calls, constructing a noise processing model, importing the original voice signal into the noise processing model to evaluate the noise situation, removing noise other than the driver's voice, constructing a voice-to-text recognition model, importing the denoised driver's voice and historical voice into the driver's voice recognition model to recognize the driver's voice-to-text, constructing a business matching model, and importing the recognized driver's voice-to-text into the business matching model to evaluate the matching of business information, thereby improving the recognition accuracy.

Description

Technical Field

[0001] This application belongs to the field of voice communication assistance, specifically a supplier voice communication assistance analysis system and method for logistics. Background Technology

[0002] In the modern logistics industry, there is a significant amount of voice communication between drivers, transportation management systems, and logistics suppliers, such as exchanging key information like waybill numbers, prices, loading and unloading times, and cargo quantities. Due to the complex in-vehicle environment, with its various sources of interference including engine noise, wind noise, tire noise, conversations among fellow passengers, and walkie-talkie calls, traditional speech recognition methods struggle to accurately capture the driver's spoken information. Furthermore, different drivers exhibit unique characteristics such as fast speaking speed, significant accent differences, and varying pronunciation habits. Traditional general-purpose speech recognition models are prone to errors when identifying key information such as waybill numbers, prices, times, and cargo quantities, leading to processing delays, inaccurate information, and even transportation disputes. Current technologies mostly rely on general-purpose speech recognition models or simple noise reduction algorithms, lacking modeling for drivers' individual characteristics and failing to provide robust automatic verification and anomaly alert mechanisms for critical business information, thus failing to effectively guarantee the accuracy and efficiency of transportation operations.

[0003] This application achieves relatively pure driver speech in complex in-vehicle environments by performing multi-source acoustic acquisition, environmental noise feature extraction, speech activity detection, and sound source separation of driver voice signals. Combined with personalized driver voice profiles and deep learning speech recognition models, it improves the accuracy of key business information recognition and reduces recognition errors caused by accent differences, fast speech speed, and pronunciation deviations. By extracting key business slots and comparing them with historical waybills, negotiation rules, and contract information, it can detect price anomalies, waybill time conflicts, and discrepancies in cargo quantities in real time, and automatically trigger verification or reminder mechanisms to achieve automatic verification of the rationality of business information. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this application proposes a supplier voice communication-assisted analysis system and method for logistics.

[0005] To achieve the above objectives, this application provides the following technical solution:

[0006] A supplier voice communication-assisted analysis system and method for logistics, comprising the following specific steps:

[0007] Based on the intelligent agent to automatically answer calls, the voice signal is collected during the voice call;

[0008] A noise processing model is constructed, and the original speech signal is imported into the noise processing model to evaluate the noise situation and remove noise other than the driver's speech.

[0009] A speech-to-text recognition model was constructed, and the denoised driver's speech and historical speech were imported into the driver's speech recognition model to recognize the driver's speech-to-text.

[0010] A business matching model is constructed, and the recognized driver voice text is imported into the business matching model to evaluate the matching of business information.

[0011] Preferably, the step of automatically answering calls based on an intelligent agent and acquiring voice signals during a voice call includes the following specific steps:

[0012] S11. Automatically answer the driver's call through the intelligent seat. During the voice call, obtain the original voice signal of the call. The original voice signal is the superposition of the driver's real voice and noise. The noise includes environmental acoustic interference and language interference signals. The environmental acoustic interference includes wind noise, tire noise, engine background noise and other broadband random noise. The language interference signals include the voices of passengers in the vehicle, walkie-talkie calls and non-driver voice components.

[0013] Preferably, the construction of the noise processing model, which involves importing the original speech signal into the noise processing model to evaluate the noise situation and removing noise other than the driver's speech, includes the following specific steps:

[0014] S21. Perform a short-time Fourier transform on the environmental acoustic interference signal in the original speech signal to extract the power spectral density, energy envelope and spectral stationarity features of the noise, and obtain the time and frequency characteristics of the environmental noise.

[0015] S22. Perform speech activity detection on the original speech signal, segment effective speech segments, use the pre-recorded driver voiceprint template, and combine it with the sound source separation algorithm to initially separate and mark the driver's speech from the non-driver's speech, and remove the non-driver's speech.

[0016] Preferably, the construction of the speech-to-text recognition model, which involves importing the denoised driver's speech and historical speech into the driver's speech recognition model to recognize the driver's speech-to-text, includes the following specific steps:

[0017] The process of constructing a speech-to-text recognition model, which involves importing the denoised driver's speech and historical speech into the model to recognize the driver's speech-to-text, includes the following specific steps:

[0018] S31. Obtain the driver's place of origin, historical call records, historical waybill information, and historical voice features. The historical voice features include speech rate, accent features, pronunciation shift patterns, feature frequency band offset, common phrases, and negotiation habits. Speech rate is the number of words spoken per minute. Accent features are generated through MFCC clustering. Pronunciation shift patterns include elision and aspiration distortion. Common phrases and negotiation habits are obtained from historical text data through word frequency-inverse document frequency. Generate a voice profile of the driver. Based on the accent features in the voice profile, calibrate the feature mapping, phoneme similarity matrix, and acoustic likelihood score of the acoustic model. Based on the driver's phoneme shift features and common pronunciation habits, construct a driver-specific pronunciation extension database.

[0019] S32. Input the noise-reduced driver's voice and personnel voice profile into the deep learning speech recognition model, and output the preliminary recognition text, which includes the waybill number, price, loading and unloading time and cargo quantity.

[0020] Preferably, the step of constructing a business matching model and importing the recognized driver's voice text into the business matching model to evaluate the matching of business information includes the following specific steps:

[0021] S41. Extract key business slot information from the identified driver's voice through regular expression matching and named entity recognition. The key business slot information includes price, loading and unloading time, waybill number, and the quantity and weight of goods that can be accommodated.

[0022] S42. Compare the extracted key business slot information with the current negotiated waybill business context, compare the extracted price with the historical price and the waybill agreement price. If it exceeds the threshold range, trigger the manual confirmation mechanism. Compare the extracted estimated arrival time with the waybill plan time. If it exceeds the threshold range, issue a waybill mismatch reminder. Compare the extracted quantity and weight with the information of the goods to be transported. If it is less than the quantity and weight of the goods to be transported, issue a waybill mismatch reminder.

[0023] S43. When the identified text contains ambiguous expressions that make it impossible to determine the slot, and when there are ambiguous expressions, easily confused words with accents, and numerical recognition errors in the identified text, semantic verification is performed. A verification inquiry is initiated to the driver in the form of voice. Based on the confirmation result, the final accurate business instruction is generated and the personalized voice recognition model is updated. At the same time, the driver's voice information is automatically structured and written into the transportation management system. The driver's credit and historical price records are updated based on the verification result.

[0024] A supplier voice communication assistance analysis system for logistics, implemented based on the aforementioned supplier voice communication assistance analysis method for logistics, specifically includes:

[0025] The voice acquisition module is used to automatically answer calls through the smart agent and acquire voice signals during the voice call.

[0026] The noise processing module is used to remove noise other than the driver's voice from the original speech signal;

[0027] The speech-to-text recognition module is used to recognize the driver's speech-to-text by comparing the noise-reduced driver's speech with historical speech.

[0028] The business matching module is used to evaluate the matching of business information based on the recognized driver's voice text.

[0029] An electronic device includes: a processor and a memory, wherein the memory stores a computer program that can be called by the processor;

[0030] The processor executes the aforementioned supplier voice communication-assisted analysis method for logistics by calling the computer program stored in the memory.

[0031] A computer-readable storage medium is characterized by storing instructions that, when executed on a computer, cause the computer to perform the aforementioned supplier voice communication-assisted analysis method for logistics.

[0032] Compared with the prior art, the beneficial effects of this application are:

[0033] This application is based on intelligent agents automatically answering calls, acquiring voice signals during voice calls, constructing a noise processing model, importing the original voice signal into the noise processing model to evaluate the noise situation, removing noise other than the driver's voice, constructing a voice-to-text recognition model, importing the noise-reduced driver's voice and historical voice into the driver's voice recognition model to recognize the driver's voice-to-text, constructing a business matching model, importing the recognized driver's voice-to-text into the business matching model to evaluate the matching of business information, thereby improving recognition accuracy. Attached Figure Description

[0034] Figure 1 This is a schematic diagram of the overall process of the supplier voice communication-assisted analysis method for logistics used in this application;

[0035] Figure 2 This is a diagram showing the signal transmission for voice communication between the logistics supplier and the applicant.

[0036] Figure 3 This is a schematic diagram of the overall framework of the supplier voice communication-assisted analysis system used in logistics in this application. Detailed Implementation

[0037] The technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments.

[0038] Example 1

[0039] Please see Figure 1 - Figure 2 One embodiment provided in this application is as follows: Figure 1 As shown, the supplier voice communication-assisted analysis method for logistics includes the following specific steps:

[0040] Based on the intelligent agent to automatically answer calls, the voice signal is collected during the voice call;

[0041] A noise processing model is constructed, and the original speech signal is imported into the noise processing model to evaluate the noise situation and remove noise other than the driver's speech.

[0042] A speech-to-text recognition model was constructed, and the denoised driver's speech and historical speech were imported into the driver's speech recognition model to recognize the driver's speech-to-text.

[0043] A business matching model is constructed, and the recognized driver voice text is imported into the business matching model to evaluate the matching of business information.

[0044] In this embodiment, it should be specifically explained that the acquisition of voice signals during a voice call based on the intelligent agent's automatic call answering includes the following specific steps:

[0045] S11. The intelligent agent automatically answers the driver's calls. During the voice call, the original voice signal is acquired. The original voice signal is a superposition of the driver's actual voice and noise. The noise includes environmental acoustic interference and speech interference signals. The environmental acoustic interference includes wind noise, tire noise, engine background noise and other broadband random noise. The speech interference signals include the voices of passengers in the vehicle, walkie-talkie calls and non-driver voice components. Answering calls through the intelligent agent avoids missed calls. The acquired original voice signal provides raw data for subsequent noise processing and speech recognition. The original signal contains all useful information and noise, which can be used for comprehensive noise analysis and speech enhancement to improve recognition accuracy.

[0046] In this embodiment, it is necessary to specifically explain that constructing a noise processing model, importing the original speech signal into the noise processing model to evaluate the noise situation, and removing noise other than the driver's speech includes the following specific steps:

[0047] S21. Perform a short-time Fourier transform on the environmental acoustic interference signal in the original speech signal to extract the power spectral density, energy envelope, and spectral stationarity features of the noise, and obtain the time and frequency characteristics of the environmental noise. Through the short-time Fourier transform, divide the continuous time domain signal into multiple short-time frames and apply a window function to obtain the spectral information of each frame in the frequency domain. Based on the spectral data, extract the power spectral density of the noise to reflect the energy distribution characteristics of the noise at different frequencies. Calculate the energy envelope to describe the intensity distribution trend of the noise over time. By analyzing the spectral stationarity features, evaluate the stability of the noise spectrum within the short-time frame, distinguish between random broadband noise and periodic sound sources, and obtain the multidimensional characteristics of the environmental noise in the time and frequency domains.

[0048] S22. Perform speech activity detection on the original speech signal, segment effective speech segments, use the pre-recorded driver voiceprint template, and combine the sound source separation algorithm to initially separate and mark the driver's speech from the non-driver's speech, and remove the non-driver's speech, which includes the speech of fellow passengers, walkie-talkie calls and external noise. By removing noise interference, the accuracy of speech recognition is improved and the speech quality is enhanced.

[0049] In this embodiment, the sound source separation algorithm includes the following specific steps: For each time-frequency unit, the extracted acoustic features, including power spectral density, spectral stationarity, and spatial correlation features, are used. A trained deep neural network model (a convolutional network based on time-frequency masking) is used to predict the probability masking matrix for each time-frequency unit, determining whether it belongs to the driver's speech or non-driver's speech. The masking matrix is ​​then applied to the amplitude spectrum for weighted filtering to reconstruct the amplitude spectrum of the driver's speech while maintaining phase information, resulting in the separated driver's speech signal. Simultaneously, Wiener filtering and spectral subtraction are combined to further suppress residual noise. Wiener filtering calculates the filter gain function using the noise power spectrum and the separated speech power spectrum. During vehicle operation, there are environmental noise signals and speech interference signals, requiring the filtering out of these interferences and the extraction of the driver's speech. However, the driver's speech signal and speech interference signals have certain similarities. The amplitude characteristics of the environmental noise signal and speech interference signal are obtained. Based on these signal characteristics, the position with the maximum signal energy kurtosis difference is used as the maximum short-time Wiener filtering window length, effectively processing high-amplitude signals. The purpose of random interference is to weight the amplitude spectrum, adaptively suppress residual noise at each frequency, and preserve the speech signal energy to the maximum extent. The signal duration is denoted as N, and the window length is determined by scanning within the range of [5, N-1] ms with a step size of 2-4 ms. The measurement signal is then sequentially truncated using each window length. Envelope processing is performed on the truncated signals, and the kurtosis value of the envelope is calculated. The difference between the kurtosis values ​​corresponding to adjacent window lengths is calculated, and the window length corresponding to the largest difference is taken as the upper limit of the window length. The window length is then scanned again, and the measurement signal is subjected to short-time Wi-Fi at each window length. Wiener filtering is performed to calculate the signal-to-noise ratio (SNR). The window length corresponding to the maximum SNR is selected as the adaptive window length. Short-time Wiener filtering is then performed based on the adaptive window length to obtain a useful signal with interference removed. Spectral subtraction reduces the interference of broadband random noise and environmental noise on the speech by subtracting the noise amplitude spectrum from the separated speech amplitude spectrum. Smoothing is also applied to avoid music noise artifacts in the speech signal. The processed amplitude spectrum and the original phase spectrum are then subjected to inverse short-time Fourier transform to reconstruct the separated driver speech signal in the time domain, providing input for speech recognition.

[0050] In this embodiment, it is necessary to specifically explain that constructing a speech-to-text recognition model and importing the denoised driver's speech and historical speech into the driver's speech recognition model to recognize the driver's speech-to-text includes the following specific steps:

[0051] The process of constructing a speech-to-text recognition model, which involves importing the denoised driver's speech and historical speech into the model to recognize the driver's speech-to-text, includes the following specific steps:

[0052] S31. Obtain the driver's place of origin, historical call records, historical waybill information, and historical voice features. The historical voice features include speech rate, accent features, pronunciation shift patterns, feature frequency band offset, common phrases, and negotiation habits. Speech rate is the number of words spoken per minute. Accent features are generated through MFCC clustering (Mel frequency cepstral coefficients). Pronunciation shift patterns include elision and aspiration distortion. Common phrases and negotiation habits are obtained from historical text data through word frequency-inverse document frequency. Generate a voice profile of the driver. Based on the accent features in the voice profile, calibrate the feature mapping, phoneme similarity matrix, and acoustic likelihood score of the acoustic model. Based on the driver's phoneme shift features and common pronunciation habits, construct a driver-specific pronunciation extension database. Combine the voice profile of the driver to improve the recognition rate, make the recognition of key slots and personalized phrases more accurate, reduce errors caused by misidentification or accent confusion, and complete personalized recognition.

[0053] In this embodiment, the construction of the driver-specific pronunciation extension database includes the following specific steps: collecting speech samples from the driver's historical call records and historical speech data; preprocessing each speech sample, including noise reduction, endpoint detection, and short-time Fourier transform; extracting amplitude and phase spectra; and calculating Mel-frequency cepstral coefficients and their first and second-order difference features to describe accents, phonology, and pronunciation details. Simultaneously, performing pronunciation shift analysis on historical speech samples, statistically analyzing patterns such as elision, aspiration distortion, vowel and consonant substitution, and feature frequency band offsets during pronunciation; identifying commonly used business phrases and colloquial expressions from historical text data using word frequency-inverse document frequency analysis; analyzing the phoneme sequences corresponding to high-frequency phrases; and extracting driver-specific phoneme shifts. Pronunciation habits and accent characteristics are mapped to a searchable phoneme-pronunciation table, which maps the driver's phoneme offset features, accent patterns, and common phrases. For each key phoneme or phrase, the driver's unique pronunciation pattern is recorded, and a multi-layered index structure is constructed. The dedicated pronunciation database is combined with a deep learning speech recognition model, and the feature mapping, phoneme similarity matrix, and acoustic likelihood score of the acoustic model are calibrated. This allows the recognition model to prioritize the driver's unique pronunciation pattern when recognizing key slots and common phrases, thereby improving personalized recognition accuracy. The speech recognition results and verification information after each call by the driver are used to dynamically update the database, recording new pronunciation offsets, phrase usage frequencies, and accent characteristics. This enables adaptive optimization and continuous expansion of the database, thereby continuously improving recognition accuracy.

[0054] S32. Input the noise-reduced driver's voice and personnel voice profile into the deep learning speech recognition model and output the preliminary recognition text. The preliminary recognition text includes the waybill number, price, loading and unloading time and cargo quantity. The recognition rate can be improved by using the personnel voice profile.

[0055] In this embodiment, the construction of the deep learning speech recognition model includes the following specific steps: The accent features, phoneme shift patterns, common phrases, and pronunciation habits information in the driver's personalized speech profile are fused with acoustic features. Temporal context modeling is performed on the fused feature sequence to capture the long-term and short-term dependencies of the speech signal. An attention mechanism is used to enhance the representation ability of key phonemes and semantic information. The model calculates the attention weights for each time step of the input sequence with all time steps in the sequence. A weighted summation is used to generate a context vector, capturing the long-term and short-term dependencies and cross-frame correlations of the speech signal in time. High weights are assigned to key phonemes or word fragments containing key business information to enhance the effect of key information. The model output is the phoneme or character probability distribution for each time step. The decoder converts the probability distribution into preliminary recognized text. During the decoding process, the driver's common phrases and pronunciation habits are used for language model correction to correct easily confused words or numbers, achieving high-precision recognition of key business information (waybill number, price, loading and unloading time, cargo quantity, etc.). The recurrent neural network and attention mechanism are based on sequence modeling theory; the recurrent neural network captures short-term dependencies, and the attention mechanism captures long-term cross-frame correlations.

[0056] In this embodiment, it is necessary to specifically explain that constructing a business matching model and importing the recognized driver's voice text into the business matching model to evaluate the matching of business information includes the following specific steps:

[0057] S41. Extract key business slot information from the identified driver's voice through regular expression matching and named entity recognition. The key business slot information includes price (freight, loading and unloading fee, surcharge), loading and unloading time (estimated arrival time, loading and unloading time), waybill number, and the quantity and weight of goods that can be accommodated.

[0058] In this embodiment, the following specific steps are taken to extract key business slot information using regular expression matching and named entity recognition: Regular expressions are used to perform pattern matching on numbers, time and unit information that may be contained in the text; price information can be matched with integer values, and freight, loading and unloading fees and surcharges are extracted by combining keywords such as yuan, block, and RMB; time information can be matched with standard time formats and colloquial expressions to extract estimated arrival time and loading and unloading time; waybill numbers are identified by matching specific alphanumeric combination patterns, including fixed length and prefix plus number sequences; cargo quantity and weight are extracted by matching number and unit combinations; the text is input into a named entity recognition model for semantic level recognition; the named entity recognition model is trained to identify entity categories such as price, time, waybill number, quantity and weight in transportation business, and supplements and corrects slots that regular expression matching fails to extract accurately; through the combined use of regular expression matching and named entity recognition, business slots such as freight, loading and unloading fees, surcharges, loading and unloading time, waybill number and cargo quantity and weight can be obtained from the driver's speech recognition text.

[0059] S42. Compare the extracted key business slot information with the current negotiated waybill business context. Compare the extracted price with historical prices and waybill agreement prices. If it exceeds the threshold range, trigger the manual confirmation mechanism. Compare the extracted estimated arrival time with the waybill plan time. If it exceeds the threshold range, issue a waybill mismatch reminder. Compare the extracted quantity and weight with the information of the goods to be transported. If it is less than the quantity and weight of the goods to be transported, issue a waybill mismatch reminder. Through automatic business verification, reduce human error, improve transportation business processing efficiency, and reduce operational risks. Through comparison with historical data, identify abnormal prices and loading and unloading times to reduce business risks.

[0060] In this embodiment, the extracted key business slot information is compared with the current negotiated waybill business context. The threshold acquisition includes the following specific steps: statistically analyzing historical waybill prices, loading and unloading times, cargo quantity and weight, etc., calculating the average value and standard deviation, and setting the threshold as the average value ± several times the standard deviation to reflect the actual business fluctuation range. Based on the company's internal contract agreements and transportation specifications, the allowable price fluctuation ratio, the allowable deviation of the expected arrival time, and the allowable error of cargo quantity and weight are set. Through a dynamic adaptive algorithm, real-time business data is monitored. Combined with sliding window statistics or exponential weighted moving average methods, the threshold is fine-tuned to adapt to changes in business trends. During the comparison process, the extracted price is compared with historical prices and waybill agreement prices. If it exceeds the set threshold range, a manual confirmation mechanism is triggered. The extracted expected arrival time is compared with the waybill plan time. If it exceeds the threshold range, a waybill mismatch reminder is issued. The extracted cargo quantity and weight are compared with the cargo information to be transported. If they are lower than the cargo quantity or weight to be transported, a waybill mismatch reminder is also triggered.

[0061] S43. When the identified text contains ambiguous expressions that make it impossible to determine the slot, and when there are ambiguous expressions, easily confused words with accents, and numerical recognition errors in the identified text, semantic verification is performed. A verification inquiry is initiated to the driver in the form of voice. Based on the confirmation result, the final accurate business instruction is generated and the personalized voice recognition model is updated. At the same time, the driver's voice information is automatically structured and written into the transportation management system. The driver's credit and historical price records are updated based on the verification result.

[0062] The advantages of this embodiment compared to the prior art are:

[0063] This application is based on intelligent agents automatically answering calls, acquiring voice signals during voice calls, constructing a noise processing model, importing the original voice signal into the noise processing model to evaluate the noise situation, removing noise other than the driver's voice, constructing a voice-to-text recognition model, importing the noise-reduced driver's voice and historical voice into the driver's voice recognition model to recognize the driver's voice-to-text, constructing a business matching model, importing the recognized driver's voice-to-text into the business matching model to evaluate the matching of business information, thereby improving recognition accuracy.

[0064] Example 2

[0065] like Figure 3 As shown, the supplier voice communication auxiliary analysis system for logistics is based on the aforementioned supplier voice communication auxiliary analysis method for logistics. Specifically, it includes a voice acquisition module, a noise processing module, a voice-to-text recognition module, and a business matching module. The voice acquisition module is used to automatically answer calls through an intelligent agent and acquire voice signals during the call. The noise processing module is used to remove noise other than the driver's voice from the original voice signal. The voice-to-text recognition module is used to recognize the driver's voice-to-text through the noise-reduced driver's voice and historical voice data. The business matching module is used to evaluate the matching of business information based on the recognized driver's voice-to-text.

[0066] Example 3

[0067] This embodiment provides an electronic device, including: a processor and a memory, wherein the memory stores a computer program that can be called by the processor;

[0068] The processor executes the aforementioned supplier voice communication-assisted analysis method for logistics by calling computer programs stored in memory.

[0069] The electronic device can vary considerably depending on its configuration or performance. It may include one or more Central Processing Units (CPUs) and one or more memories, wherein the memory stores at least one computer program, which is loaded and executed by the processor to implement the supplier voice communication assistance analysis method for logistics provided in the above-described embodiment. The electronic device may also include other components for implementing its functions; for example, it may have wired or wireless network interfaces and input / output interfaces for data input and output. Details will not be elaborated upon in this embodiment.

[0070] Example 4

[0071] This embodiment proposes a computer-readable storage medium on which an erasable and rewritable computer program is stored.

[0072] When the computer program runs on the computer device, it causes the computer device to execute the aforementioned supplier voice communication-assisted analysis method for logistics.

[0073] For example, computer-readable storage media can be read-only memory (ROM), random access memory (RAM), compact disc read-only memory (CD-ROM), magnetic tape, floppy disk, and optical data storage devices.

[0074] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. A computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, all or part of the processes or functions according to the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. Computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via a wired network and / or wireless network. A computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more sets of available media. Available media can be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., DVDs), or semiconductor media. Semiconductor media can be solid-state drives (SSDs).

Claims

1. A method for analyzing voice communication of a supplier for logistics, characterized by, It comprises the following specific steps: Based on the automatic answering of intelligent agents, the voice signal acquisition in the voice call process is obtained; A noise processing model is constructed, and the original voice signal is introduced into the noise processing model to evaluate the noise situation and remove the noise other than the driver's voice; A speech text recognition model is constructed, and the noise-reduced driver's voice and historical voice are introduced into the driver's speech recognition model to recognize the driver's speech text, including the following specific steps: obtaining the driver's driver's hometown and historical call records, historical shipping information and historical voice features, wherein the historical voice features include speech speed, accent features, pronunciation shift pattern, characteristic frequency band bias, commonly used phrases and bargaining habits, generating a personnel voice portrait, calibrating the feature mapping, phoneme similarity matrix and acoustic likelihood score of the acoustic model according to the accent features in the voice portrait, constructing a driver-specific pronunciation expansion database based on the driver's phoneme shift features and common pronunciation habits, wherein the construction of the driver-specific pronunciation expansion database is to collect voice samples from the driver's historical call records and historical voice data, preprocess each voice, including noise reduction, endpoint detection and short-time Fourier transform, extract amplitude spectrum and phase spectrum, and calculate mel frequency cepstral coefficients and their first and second order difference features, simultaneously analyze the pronunciation shift of historical voice, count the driver's elision, aspirated deformation, vowel initial replacement and characteristic frequency band bias pattern, identify the driver's commonly used business phrases and colloquial expressions from the historical text data through term frequency-inverse document frequency, analyze the phoneme sequence corresponding to the high-frequency phrases, extract the driver's specific phoneme shift, pronunciation habit and accent feature, map the driver's phoneme shift feature, accent pattern and commonly used phrases to a searchable phoneme-pronunciation table, record the driver's specific pronunciation pattern for each key phoneme or phrase, and construct a multi-layer index structure, combine the exclusive pronunciation database with the deep learning speech recognition model, calibrate the feature mapping, phoneme similarity matrix and acoustic likelihood score of the acoustic model, input the noise-reduced driver's voice and personnel voice portrait into the deep learning speech recognition model, and output the preliminary recognition text, wherein the preliminary recognition text includes shipping number, price, loading and unloading time and cargo quantity, wherein the construction of the deep learning speech recognition model is to fuse the accent features, phoneme shift pattern, commonly used phrases and pronunciation habits information in the driver's personalized voice portrait with acoustic features, model the time context of the fused feature sequence, capture the long and short term dependence of the voice signal, enhance the representation ability of key phonemes and semantic information through attention mechanism, calculate the attention weight of all time steps in the sequence for each time step of the input sequence, generate a context vector using weighted summation, capture the long and short term dependence of the voice signal in time and the correlation across frames, assign high weight to key phonemes or word fragments containing business key information, and the model output is the phoneme and character probability distribution of each time step; The identified driver voice text is introduced into the business matching model to evaluate the matching of business information.

2. The method for voice communication assistant analysis of suppliers for logistics according to claim 1, wherein, The automatic answering of the call by the intelligent agent, and the acquisition of the voice signal in the voice call process include the following specific steps: The automatic answering of the call by the intelligent agent, and the acquisition of the voice signal in the voice call process include the following specific steps:

3. The method for voice communication assistant analysis of suppliers for logistics according to claim 2, wherein, The automatic answering of the call by the intelligent agent, and the acquisition of the voice signal in the voice call process include the following specific steps: The automatic answering of the call by the intelligent agent, and the acquisition of the voice signal in the voice call process include the following specific steps: S21, short-time Fourier transform is performed on the environmental acoustic interference signal in the original voice signal to extract the power spectral density, energy envelope and spectral stationarity characteristics of the noise, and the time and frequency characteristics of the environmental noise are obtained; 4. The method for voice communication assistant analysis of suppliers for logistics according to claim 3, wherein, S22, the original voice signal is subjected to speech activity detection, and the effective speech segment is segmented, the pre-recorded driver voiceprint template is used, and the sound source separation algorithm is combined to preliminarily separate and mark the driver voice and the non-driver voice, and the non-driver voice is removed. The automatic answering of the call by the intelligent agent, and the acquisition of the voice signal in the voice call process include the following specific steps: S41, the recognized driver voice is extracted through regular matching and named entity recognition to obtain key business slot information, wherein the key business slot information includes price, loading and unloading time, waybill number, and the number and weight of the cargo that can be accommodated; S42, the extracted key business slot information is compared with the current negotiation waybill business context, the extracted price is compared with the historical price and the waybill agreement price, if it exceeds the threshold range, an artificial confirmation mechanism is triggered, the extracted estimated arrival time is compared with the waybill planned time, if it exceeds the threshold range, a waybill mismatch reminder is issued, and the extracted number and weight are compared with the to-be-transported cargo information, and if it is less than the to-be-transported cargo number and weight, a waybill mismatch reminder is issued; 5. A system for assisting analysis of voice communication of a supplier for logistics, which is implemented based on the method for assisting analysis of voice communication of a supplier for logistics according to any one of claims 1 to 4, characterized by, S43, when the recognized text contains ambiguous expressions that cause the slot to be unable to be determined, and the ambiguous expressions, accent confusing words and number recognition errors in the recognized text, semantic verification is performed, verification inquiries are initiated to the driver in the form of voice, and according to the confirmation result, the final accurate business instruction is generated and the voice text recognition model is updated. Specifically, it includes: A voice acquisition module for automatically answering a call by an intelligent agent and acquiring a voice signal in a voice call process; A noise processing module for removing noise other than driver voice in the original voice signal; A voice text recognition module for recognizing driver voice text from the de-noised driver voice and historical voice; A business matching module for evaluating the matching of business information through the recognized driver voice text.

6. An electronic device comprising: a processor and a memory, wherein the memory stores a computer program that can be invoked by the processor; characterized in that the processor, by invoking the computer program stored in the memory, executes the method for analyzing voice communication of a supplier for logistics as claimed in any one of claims 1-4.

7. A computer readable storage medium characterized in that, instructions stored in the memory, when the instructions are run on a computer, cause the computer to execute the method for analyzing voice communication of a supplier for logistics as claimed in any one of claims 1-4.

Images