A method and system for identifying information at least one item category of structured audio data

Abstract

There is provided a system for recognizing information item categories of structured audio data, which comprises an information request component and a recognition server. The service provider configures the information request component to send a recognition request to the recognition server. The recognition request comprises the target audio data file or a storage address of the target audio data file. A corresponding method is also provided.

Description

[0001] A METHOD AND SYSTEM FOR IDENTIFYING INFORMATION AT LEAST ONE ITEM CATEGORY OF STRUCTURED AUDIO DATA

[0002] FIELD OF INVENTION

[0003] The present invention relates to a method and system for identifying at least one item category of structured audio data.

[0004] BACKGROUND

[0005] Structured audio data refers to audio data content with fixed categories of text information items, such as recordings played by navigation software and prompt tones when a call is not connected. In existing technical application scenarios, a large amount of structured audio data needs to be identified and processed to extract the text information for subsequent business use. Structured audio data exhibits the following three characteristics: 1 ) a large number of audio files are generated during business processes, but the categories of text information items contained are relatively few; 2) effective voice information may be incomplete, for example, standard prompts may only contain partial content; 3) the data is filled with a large amount of invalid audio information, such as long periods of silence.

[0006] In the existing technical field, Automatic Speech Recognition (ASR) technology is commonly used to extract text information from audio. ASR technology, based on a deep learning framework, is widely used in business scenarios where audio file text content needs to be recognized. However, the application of ASR technology faces multiple challenges: firstly, it has high computational resource requirements and a relatively high research and development threshold, resulting in significant application costs; secondly, the operation of ASR services relies on the support of a large number of high-performance servers, with significant server resource usage; at the same time, its recognition efficiency is relatively low, making it difficult to meet the demand for efficient processing. The content described in the background technology is only for the purpose of facilitating understanding of related technologies in this field and is not regarded as an acknowledgment of existing technology.

[0007] SUMMARY

[0008] In a first aspect, there is provided a method for identifying information item category of structured audio data, characterized in that the method comprising the following steps: acquiring the target audio data file, wherein the target audio data file is structured audio data; segmenting the target audio data file to obtain multiple target audio data blocks; extracting audio features of the target audio data blocks to obtain target audio features; when preset audio features are stored in the cache, calculating the similarity between each target audio feature and each preset audio feature, wherein the preset audio features have preset information item categories, and when the similarity is greater than or equal to the first threshold, setting the preset information item category of the corresponding preset audio feature as the target information item category of the target audio feature; when no preset audio features are stored in the cache or the similarity between all target audio features and the preset audio features is less than the first threshold, performing speech recognition on the target audio data file to obtain the target information item category of the target audio data blocks of the target audio data file, assigning the target information item category corresponding to the target audio data block to the target audio feature, and storing the target audio feature as a preset audio feature in the cache; acquiring the target information item category of the target audio data file based on the target information item categories of the multiple target audio features.

[0009] In a second aspect, there is provided a system for identifying information item category of structured audio data, characterized in that the system including a first acquisition module, a segmentation module, a feature extraction module, a similarity matching module, a speech recognition module, and a second acquisition module, wherein: the first acquisition module is configured to acquire the target audio data file, wherein the target audio data file is structured audio data; the segmentation module is configured to segment the target audio data file to obtain multiple target audio data blocks; the feature extraction module is configured to extract audio features of the target audio data blocks to obtain target audio features; the similarity matching module is configured to, when preset audio features are stored in the cache, calculate the similarity between the target audio features and each preset audio feature, wherein the preset audio features have preset information item categories, and when the similarity is greater than or equal to the first threshold, set the preset information item category of the corresponding preset audio feature as the target information item category of the target audio feature; the speech recognition module is configured to, when no preset audio features are stored in the cache or the similarity between the target audio features and the preset audio features is less than the first threshold, perform speech recognition on the target audio data file to obtain the target information item category of the target audio data blocks of the target audio data file, assign the target information item category corresponding to the target audio data block to the target audio feature, and store the target audio feature as a preset audio feature in the cache; the second acquisition module is configured to acquire the target information item category of the target audio data file based on the target information item categories of the multiple target audio features.

[0010] In a third aspect, there is provided a computer readable storage medium storing a computer program, wherein the program, when executed by a processor, implements the aforementioned method.

[0011] In a final aspect, there is provided an electronic device, including: a processor and a memory storing computer programs, wherein the processor is configured to execute the aforementioned method when running the computer programs.

[0012] It will be appreciated that the broad forms of the invention and their respective features can be used in conjunction, interchangeably and / or independently, and reference to separate broad forms is not intended to be limiting. DESCRIPTION OF FIGURES

[0013] By reading the detailed description of exemplary embodiments in the following text, those skilled in the art will understand the advantages and benefits described herein, as well as other advantages and benefits. The accompanying drawings are only for the purpose of demonstrating exemplary embodiments and are not considered a limitation on the application. And throughout all drawings, the same components are represented by the same numbers. In the attached drawings:

[0014] FIG. 1 shows an embodiment of a system structure diagram of a system for identifying information item categories of structured audio data;

[0015] FIG. 2 shows an exemplary flowchart of a method for identifying information item categories of structured audio data according to an embodiment of the application;

[0016] FIG. 3 shows an exemplary flowchart of segmentation processing in the method for identifying information item categories of structured audio data according to an embodiment of the application;

[0017] FIG. 4 shows an exemplary flowchart of extracting segmented features in the method for identifying information item categories of structured audio data according to an embodiment of the application;

[0018] FIG. 5 shows an exemplary flowchart of performing speech recognition on the target audio data file in the method for identifying information item categories of structured audio data according to an embodiment of the application;

[0019] FIG. 6 shows an exemplary flowchart of performing speech recognition on a target audio data block in the method for identifying information item categories of structured audio data according to an embodiment of the application;

[0020] FIG. 7A shows an exemplary flowchart of acquiring target information item categories of the target audio data file in the method for identifying information item categories of structured audio data according to an embodiment of the application;

[0021] FIG. 7B shows another exemplary flowchart of acquiring target information item categories of the target audio data file in the method for identifying information item categories of structured audio data according to an embodiment of the application; FIG. 8A shows an exemplary flowchart of a method for identifying information item categories of structured audio data according to another embodiment of the application;

[0022] FIG. 8B shows an exemplary flowchart of extracting segmented features in the method for identifying information item categories of structured audio data according to another embodiment of the application;

[0023] FIG. 80 shows an exemplary flowchart of performing speech recognition in the method for identifying information item categories of structured audio data according to another embodiment of the application;

[0024] FIG. 9 shows an exemplary structural diagram of a device for identifying information item categories of structured audio data according to an embodiment of the application; FIG. 10 shows an exemplary structural diagram of an electronic device capable of implementing the method according to an embodiment of the application.

[0025] DETAILED DESCRIPTION

[0026] To make the objectives, technical solutions, and advantages of the application clearer and more understandable, further detailed descriptions are provided below with specific embodiment and accompanying drawings. The illustrative embodiment and their descriptions are used to explain the application but do not limit it.

[0027] In an embodiment of the application, "ASR" refers to automatic speech recognition technology; "DTW refers to the dynamic time warping algorithm; "VAD (Voice Activity Detection)" refers to voice activity detection technology.

[0028] It is observed that in the current process of recognizing categories of structured audio information items, ASR technology is often used for full-text recognition before making determinations. However, since most of the structured audio consists of invalid information, using full-text recognition is inefficient and increases the usage rate of ASR servers, thereby increasing the cost of recognition.

[0029] To address the issues of existing technologies, an embodiment of the application provides a method for acquiring information item categories based on similarity calculation primarily at scenarios where structured audio contains a large amount of silence, the structured audio is segmented to form target audio data blocks. Audio features of the target audio are extracted and compared with cached target audio features to calculate similarity. When the similarity is greater than or equal to the first threshold, the information item category of the cached target audio features is set as the information item category corresponding to the target audio features. When there are no cached target audio features or the similarity with the cached target audio data blocks is less than the first threshold, the target audio features cannot be matched to a information item category. In such case, speech recognition is performed on the structured audio data to obtain text information of the structured audio, and the information item category of the target audio features is obtained based on the text information. To facilitate the next calculation, the target audio features with acquired information item categories are stored in the cache along with their information item categories. In the process of recognizing information item categories in structured audio, the embodiment of the application reduce the invocation of speech recognition technology, thereby reducing the occupancy of speech recognition servers. Compared to the method of full-text recognition, it has higher recognition efficiency and is suitable for large-scale application scenarios.

[0030] The information item category in an embodiment of the application refers to key information contained in the text information of the audio. For example, for audio with the text information: "The number you dialed is not in service, please check and dial again," its information item category is "not in service." For audio with the text information: "The subscriber you dialed is powered off, please dial again later," its information item category is "powered off."

[0031] Specifically, as shown in FIG 1 , an embodiment of the application provides a system 100 for recognizing information item categories of structured audio data, which comprises an information request component 110 and a recognition server 120. The service provider configures the information request component 110 to send a recognition request to the recognition server 120. The recognition request comprises the target audio data file or a storage address of the target audio data file. The recognition server 120 is configured to perform the following operations: - the recognition server 120 is configured to obtain the target audio data file, which is structured audio data. The recognition server 120 is configured to acquire the target audio data file based on the recognition request.

[0032] - the recognition server 120 is configured to perform segmentation processing on the target audio data file to obtain multiple target audio data blocks. In some embodiment of the application, the recognition server 120 is configured to sample the target audio data file to obtain a first sampled file and call a voice segmentation model to perform segmentation processing on the first sampled file to obtain the multiple target audio data blocks.

[0033] In an embodiment of the application, segmentation processing is performed on the target audio data file, which is structured audio, to eliminate silence signals and noise signals while retaining effective voice signals. Each segmented audio data contains an independent sentence and can form effective audio features. In an embodiment of the application, voice activity detection (VAD) technology can be used for segmentation to accurately locate the start and end points of voice from noisy audio. For example, for the target audio data file M, segmentation is performed to obtain multiple target audio data blocks {ml , m2... mN}, where N is the number of target audio data blocks.

[0034] The recognition server 120 is configured to extract audio features of the target audio data blocks to obtain target audio features. In some embodiment of the application, the recognition server 120 is configured to perform signal preprocessing on the target audio data blocks; perform frame processing on the preprocessed target audio data blocks according to the preset standard frame length to obtain a frame data item set, where the frame data item set comprises multiple frame data items; perform windowing processing on the frame data items to extract frame features of the frame data items; and combine the frame features of all frame data items corresponding to the target audio data blocks to form the target audio features of the target audio data blocks.

[0035] In an embodiment of the application, signal preprocessing includes normalization processing and voice emphasis processing. Voice emphasis processing is performed on the input digital voice signal to enhance the high-frequency part of the voice, remove the influence of lip radiation, and increase the high-frequency resolution of the voice.

[0036] In an embodiment of the application, the framing of the voice signal can be achieved by using a movable finite-length window for weighting. Generally, the number of frames per second is 33 to 100, depending on the actual situation. Although continuous segmentation can be used for framing, overlapping segmentation is generally adopted to ensure smooth transitions between frames and maintain their continuity. The overlapping part between the previous frame and the next frame is called frame shift, and the ratio of frame shift to frame length is generally 0 to 1 / 2. After framing, discontinuities appear at the beginning and end of each frame. Therefore, as the number of segmented frames increases, the error with the original signal becomes larger. Windowing is used to solve this problem, making the framed signal continuous and each frame exhibiting the characteristics of a periodic function. Common window functions include rectangular window, Hamming window, and Hanning window.

[0037] In an embodiment of the application, the data volume of each segment is still large, and it is difficult to obtain audio features that can represent the full data by directly extracting audio features. In an embodiment of the application, through frame processing, the target audio data blocks are divided into multiple audio frames according to the preset standard frame length, audio features of each audio frame are extracted, and the audio features of each audio frame are combined to form target audio features.

[0038] For example, it is generally recommended to divide into one frame every 10ms to 30ms. In an embodiment of the application, one frame is 20ms, and the signal features within this frame are short-term stationary. Assuming the No n target audio data block mN has a length of 10s, then the target audio data block mN contains 500 frames. Each frame is a short-term stationary signal and can extract one feature. Then, the number of features for each segment is 500, and the corresponding target audio feature FmN is {fmN1 , fmN2...fmN500}. The target audio features of multiple target audio data blocks {ml , m2... mN} are {Fm1 , Fm2... FmN}. In an embodiment of the application, the frame-based features include the short-time zero-crossing rate and short-time energy of the frame-based data items. The short- time zero-crossing rate represents the number of times the waveform signal crosses the zero value in a frame of speech. For continuous signals, zero-crossing means that the waveform passes through the time axis, while for discrete signals, zero-crossing means that the sign of adjacent sampling points changes. Short-time energy can be used to distinguish between voiced and unvoiced sounds, between sounds and silence, between initials and finals, and so on. It can be seen as the output of the square of the speech signal passing through a linear filter. In an embodiment of the application, the short-time energy E is calculated as E=x2(n)*w2(n), where x(n) is the speech signal and w(n) is the window function.

[0039] When the recognition server 120 is configured to determine that there are preset audio features stored in the cache, where the preset audio features have preset information item categories, it calculates the similarity between each target audio feature and each of the preset audio features. When the similarity is greater than or equal to the first threshold, the preset information item category of the corresponding preset audio feature is set as the target information item category of the target audio feature. In some embodiment of the application, the DTW (Dynamic Time Warping) algorithm is used for similarity matching. The DTW algorithm is more suitable for the similarity matching needs in this scenario than general distance algorithms because the lengths of the data to be compared and the template data are not consistent, while other distance algorithms require consistent lengths. The DTW algorithm, based on the idea of dynamic programming (DP), solves the problem of template matching with varying lengths of speech and is an earlier and more classic algorithm in speech recognition.

[0040] In an embodiment of the application, the frame-based features include the short-time zero-crossing rate and short-time energy of the frame-based data items. In the process of calculating the similarity of target audio features, the similarity calculation of short- time zero-crossing rate and the similarity calculation of short-time energy can be performed separately. In an embodiment of the application, when both the similarity calculated based on the short-time zero-crossing rate and the similarity calculated based on the short-time energy are greater than or equal to the corresponding thresholds, it is determined that the similarity between the target audio feature and the preset audio feature is greater than or equal to the first threshold.

[0041] In an embodiment of the application, the recognition server 120 is configured to also store the target information item category of the target audio feature in a temporary set of information item categories.

[0042] For example, if there are preset audio features Fca1 and Fca2 stored in the cache, with information item categories "unallocated number" and "phone off," respectively, and the target audio features are {Fm1 , Fm2... FmN}, the similarity between the target audio features and Fca1 , Fca2 is calculated one by one. For instance, when the similarity between Fm2 and Fca2 is calculated to be greater than the first threshold (e.g., 0.7), the information item category of the target audio data block Fm2 is set to "phone off," and the information item category "phone off" is stored in the temporary set of information item categories.

[0043] On the other hand, when the recognition server 120 is configured to determine that there are no preset audio features stored in the cache or that the similarity between all target audio features and the preset audio features is less than the first threshold, it performs speech recognition on the target audio data file to obtain the target information item category of the target audio data block. The obtained target information item category is set for the target audio data block in the target audio feature, and the target audio feature is stored in the cache as a preset audio feature.

[0044] For example, if there are preset audio features Fca1 and Fca2 stored in the cache, with information item categories "unallocated number" and "phone off," respectively, and the target audio features are {Fm1 , Fm2... FmN}, the similarity between the target audio features and Fca1 , Fca2 is calculated one by one, and all similarities are less than the first threshold, then ASR (Automatic Speech Recognition) is performed.

[0045] In an embodiment of the application, the specific steps for the recognition server 120 being configured to perform speech recognition on the target audio data file and obtain the target information item category of the target audio data block are as follows: perform speech recognition on the target audio data file to obtain target text data; if the target text data contains a preset information item category, perform speech recognition on the target audio data block to obtain the target information item category of the target audio data block; if the target text data does not contain a preset information item category, the recognition fails.

[0046] For example, speech recognition is performed on the target audio data file to obtain target text data: "The subscriber you dialed is busy now, please redial later. Sorry, The subscriber you dialed is busy now, please redial later." The text information corresponding to the target audio data blocks is: "The subscriber you dialed is busy now, please redial later" and "Sorry, The subscriber you dialed is busy now, please redial later." The similarity between the target audio features and Fca1 , Fca2 is compared, and all are less than the first threshold. It is determined that the target text data contains the preset information item category "busy," so speech recognition is performed on the target audio data block to obtain the target information item category of the target audio data block.

[0047] In an embodiment of the application, the specific steps for the recognition server 120 being configured to perform speech recognition on the target audio data block and obtain the target information item category of the target audio data block are as follows: take the first target audio data block as the current target audio data block and cyclically perform the following steps: perform speech recognition on the current target audio data block to obtain the text data sub-item of the current target audio data block; if the text data sub-item of the current target audio data block contains a preset information item category, store the information item category of the current target audio data block as the target information item category in the temporary set of information item categories. If the current target audio data block is not the last target audio data block, take the next target audio data block as the current target audio data block; otherwise, end the cycle; if the text data sub-item of the current target audio data block does not contain a preset information item category and the current target audio data block is not the last target audio data block, take the next target audio data block as the current target audio data block. If the current target audio data block is the last target audio data block, end the cycle.

[0048] For example, the target text data is: "The subscriber you dialed is busy now, please redial later. Sorry, The subscriber you dialed is busy now, please redial later." The text information corresponding to the target audio data blocks is: "The subscriber you dialed is busy now, please redial later" and "Sorry, The subscriber you dialed is busy now, please redial later." It is determined that the text information "The subscriber you dialed is busy now, please redial later" of the first target audio data block contains the information item category "busy." The information item category "busy" is taken as the target information item category and stored in the temporary set of information item categories. Then, it is determined that the text information "Sorry, The subscriber you dialed is busy now, please redial later" of the second target audio data block does not contain a preset information item category, so the cycle ends.

[0049] The recognition server 120 is configured to obtain the target information item category of the target audio data file based on the target information item categories of the multiple target audio features.

[0050] In an embodiment of the application, the information item categories have weight values. The steps for the recognition server 120 being configured to obtain the target information item category of the target audio data file based on the target information item categories of the multiple target audio features include: taking the target information item category with the largest weight value in the temporary set of information item categories as the target information item category of the target audio data file. For example, based on statistical experience, the weight value of the information item category "unallocated number" is 3, the weight value of the information item category "phone off" is 5, and the weight value of the information item category "busy" is 7. After the above similarity calculation or speech recognition, the temporary set of information item categories obtained is {"busy", "phone off"}. Then, based on the weight values, the target information item category of the target audio data file is determined to be "busy."

[0051] In an embodiment of the application, the target information item category of the target audio data file can be determined by the number of information item categories. The steps for the recognition server 120 being configured to obtain the target information item category of the target audio data file based on the target information item categories of the multiple target audio features include: counting the number of each target information item category in the temporary set of information item categories; setting the target information item category with the largest number as the target information item category of the target audio data file.

[0052] For example, if the temporary set of information item categories is {"on a call", "powered off", "powered off", "unallocated number"}, then according to the statistical count, the target information item category of the target audio data file is "powered off".

[0053] In an implementation of the application, both the weight value and the number of information item categories can be considered simultaneously. The recognition server 120 is configured to obtain the target information item category of the target audio data file based on the target information item categories of the multiple target audio features include: counting the number of each target information item category in the temporary set of information item categories; obtaining the composite weight value of each target information item category in the temporary set of information item categories, where the composite weight value is the product of the number of the target information item category and the corresponding weight value; setting the target information item category with the largest composite weight value as the target information item category of the target audio data file.

[0054] For example, if the weight value of the information item category "unallocated number" is 3, the weight value of the information item category "powered off" is 5, the weight value of the information item category "on a call" is 7, and the temporary set of information item categories is {"on a call", "powered off", "powered off", "unallocated number"}, then the composite weight value of the target information item category "on a call" is 7*1 =7, the composite weight value of the target information item category "powered off" is 5*2=10, and the composite weight value of the target information item category "unallocated number" is 3*1 =3. Therefore, the composite weight value of the target information item category "powered off' is the largest, and the target information item category of the target audio data file is "powered off".

[0055] The recognition server 120 is configured to send the target information item category of the target audio data file to the information request component 110.

[0056] The system 100 can perform segmented processing on structured audio, generate target audio data blocks, extract target audio features, compare the target audio features with cached target audio features, calculate similarity, and when the similarity is greater than or equal to the first threshold, set the information item category of the cached target audio features as the information item category of the corresponding target audio features. When there are no cached target audio features or the similarity with the cached target audio data blocks is less than the first threshold, the target audio features cannot be matched to an information item category, perform speech recognition on the structured audio data to obtain text information of the structured audio, and obtain the information item category of the target audio features based on the text information. To facilitate the next calculation, store the target audio features with obtained information item categories, along with the information item categories, in the cache. In the process of recognizing the information item categories of structured audio in the system of the embodiment of the application, the invocation of speech recognition technology is reduced, thereby reducing the occupancy of the speech recognition server. Compared with the method of full-text speech recognition, the recognition efficiency is higher, making it suitable for large-scale application scenarios.

[0057] As shown in FIG 2, an embodiment of the application provides a method 200 for recognizing information item categories of structured audio data, which comprises the following steps:

[0058] At step S110, obtaining the target audio data file, where the target audio data file is structured audio data. In an embodiment of the application, the text information item categories contained in the structured audio data are fixed. The target audio data file or a link address of the target audio data file can be obtained from a recognition request, and the target audio data file is obtained based on the link address.

[0059] At step S120, performing segmented processing on the target audio data file to obtain multiple target audio data blocks. In an embodiment of the application, segmented processing is performed on the target audio data file belonging to structured audio, which can eliminate silent signals and noise signals, retain effective speech signals, and each segmented audio data contains an independent sentence that can form effective audio features.

[0060] In an embodiment of the application, as shown in FIG 3, the step of performing segmented processing on the target audio data file to obtain multiple target audio data blocks comprises:

[0061] At step S121 , sampling the target audio data file to obtain a first sampled file. In an embodiment of the application, the sampling rate can be set to 8000 or 16000.

[0062] At step S122, invoking a speech interruption processing model to perform segmented processing on the first sampled file to obtain the multiple target audio data blocks. In an embodiment of the application, speech interruption technology VAD can be used for segmentation to accurately locate the start and end points of speech from noisy speech. For example, for the target audio data file M, segmented processing is performed to obtain multiple target audio data blocks {ml , m2... mN}, where N is the number of target audio data blocks.

[0063] Referring back to FIG 2, at step S130, extracting audio features of the target audio data blocks to obtain target audio features. In an embodiment of the application, the target audio features are used to represent the characteristics of the target audio data blocks, which is conducive to subsequent similarity calculations.

[0064] In an embodiment of the application, as shown in FIG 4, the step of extracting audio features of the target audio data blocks to obtain target audio features comprises:

[0065] At step S131 , performing signal preprocessing on the target audio data blocks. In an embodiment of the application, the signal preprocessing includes normalization processing and voice emphasis processing. Voice emphasis processing is performed on the input digital speech signal to enhance the high-frequency part of the speech, remove the influence of lip radiation, and increase the high-frequency resolution of the speech.

[0066] At step S132, performing frame processing on the preprocessed target audio data blocks according to the preset standard frame length to obtain a set of framed data items, where the set of framed data items comprises multiple framed data items. In an embodiment of the application, framing of the speech signal can be implemented using a method of weighting with a movable finite-length window.

[0067] At step S133, performing windowing processing on the framed data items to extract frame features of the framed data items. Generally, the number of frames per second is 33 to 100, depending on the actual situation. Although continuous segmentation can be used for framing, overlapping segmentation is generally used to smoothly transition between frames and maintain their continuity. The overlapping part between the previous frame and the next frame is called frame shift, and the ratio of the frame shift to the frame length is generally O to 1 / 2. After framing, discontinuities will appear at the beginning and end of each frame. Therefore, as the number of segmented frames increases, the error with the original signal becomes larger. Windowing is used to solve this problem, so that the framed signal becomes continuous, and each frame exhibits the characteristics of a periodic function. Common window functions include rectangular window, Hamming window, Hanning window, etc.

[0068] At step S134, combining the frame features of all the framed data items corresponding to the target audio data block to form the target audio features of the target audio data block.

[0069] In an embodiment of the application, the data volume of each segment is still relatively large, and it is difficult to obtain audio features that can represent the full data by directly extracting audio features. Therefore, an embodiment of the application adopt frame processing, dividing the target audio data block into multiple audio frames based on the preset standard frame length, extracting the audio features of each audio frame, combining the audio features of each audio frame to form the target audio features.

[0070] For example, it is generally recommended to divide the audio into frames with a length of 10ms to 30ms. In some embodiment of the application, a frame length of 20ms is used, and the signal features within this frame are short-term stable. Assuming that the nth target audio data block mN has a length of 10s, then the target audio data block mN contains 500 frames. Each frame is a short-term stable signal, and one feature can be extracted, so the number of features for each segment is 500, and the corresponding target audio feature FmN is {fmN1 , fmN2, ... , fmN500}. The target audio features of multiple target audio data blocks {ml , m2, ... , mN} are {Fm1 , Fm2, ... , FmN}.

[0071] In an embodiment of the application, the frame features include the short-term zerocrossing rate and short-term energy of the frame data items. The short-term zerocrossing rate represents the number of times the waveform signal crosses the zero value in a frame of speech. For continuous signals, zero-crossing means that the waveform passes through the time axis, while for discrete signals, zero-crossing means that the sign of adjacent sampling points changes. Short-term energy can be used to distinguish between voiced and unvoiced sounds, between sounds and silence, between initials and finals, etc. It can be seen as the output of the square of the speech signal passing through a linear filter. In an embodiment of the application, the short-term energy E=x2(n)*w2(n), where x(n) is the speech signal and w(n) is the window function.

[0072] Referring back to FIG 2, at step S140, when there are preset audio features stored in the cache, calculating the similarity between each target audio feature and each preset audio feature. When the similarity is greater than or equal to the first threshold, set the preset information item category of the corresponding preset audio feature as the target information item category of the target audio feature. The preset audio features have preset information item categories.

[0073] In an embodiment of the application, calculating the similarity between the target audio feature and the preset audio feature comprises: using the DTW (Dynamic Time Warping) algorithm to calculate the similarity between the target audio feature and the preset audio feature.

[0074] The DTW algorithm is more suitable for the similarity matching requirements of this scenario than general distance algorithms because the lengths of the data to be compared and the template data are inconsistent, while other distance algorithms require consistent lengths. The DTW algorithm, based on the idea of dynamic programming (DP), solves the problem of template matching with varying lengths of speech and is an earlier and classic algorithm in speech recognition.

[0075] In an embodiment of the application, the frame features include the short-term zerocrossing rate and short-term energy of the frame data items. In the process of calculating the similarity of the target audio features, the similarity calculation of the short-term zero-crossing rate and the similarity calculation of the short-term energy can be performed separately. In some embodiment of the application, when both the similarity calculated based on the short-term zero-crossing rate and the similarity calculated based on the short-term energy are greater than or equal to the corresponding thresholds, it is determined that the similarity between the target audio feature and the preset audio feature is greater than or equal to the first threshold. In an embodiment of the application, the target information item category of the target audio feature is also stored in a temporary set of information item categories.

[0076] For example, if there are preset audio features Fca1 and Fca2 stored in the cache, with information item categories of "unallocated number" and "power off", respectively, and the target audio features are {Fm1 , Fm2, ... , FmN}, the similarity between the target audio features and Fca1 , Fca2 is calculated one by one. For example, when the similarity between Fm2 and Fca2 is greater than the first threshold (e.g., 0.7), the information item category of the target audio data block Fm2 is set to "power off", and the information item category "power off" is stored in the temporary set of information item categories.

[0077] At step S150, when there are no preset audio features stored in the cache or the similarity between all target audio features and the preset audio features is less than the first threshold, performing speech recognition on the target audio data file to obtain the target information item category of the target audio data block of the target audio data file, assigning the target information item category corresponding to the target audio data block to the target audio feature, and storing the target audio feature as a preset audio feature in the cache.

[0078] For example, if there are preset audio features Fca1 and Fca2 stored in the cache, with information item categories of "unallocated number" and "power off", respectively, and the target audio features are {Fm1 , Fm2, ... , FmN}, the similarity between the target audio features and Fca1 , Fca2 is calculated one by one, and if all similarities are less than the first threshold, then ASR (Automatic Speech Recognition) is performed.

[0079] In an embodiment of the application, performing speech recognition on the target audio data file comprises performing speech recognition on the target audio data file itself (the full text) or performing speech recognition on its target audio data blocks.

[0080] In an embodiment, performing speech recognition on the target audio data file comprises performing speech recognition on the full text of the target audio data file and selectively performing speech recognition on its target audio data blocks. As shown in an embodiment of FIG 5, performing speech recognition on the target audio data file to obtain the target information item category of the target audio data block comprises:

[0081] At step S151 , performing full-text speech recognition on the target audio data file to obtain target text data.

[0082] At step S152, when the target text data contains a preset information item category, performing speech recognition on the target audio data block to obtain the target information item category of the target audio data block.

[0083] At step S153, when the target text data does not contain a preset information item category, the recognition fails.

[0084] For example, performing speech recognition on the target audio data file to obtain the target text data: "The subscriber you dialed is busy now, please redial later. Sorry, The subscriber you dialed is busy now, please redial later." The text information corresponding to the target audio data blocks is: "The subscriber you dialed is busy now, please redial later" and "Sorry, The subscriber you dialed is busy now, please redial later." The similarity between the target audio features and Fca1 , Fca2 is compared, and both are less than the first threshold. It is determined that the target text data contains the preset information item category "busy", then speech recognition is performed on the target audio data block to obtain the target information item category of the target audio data block.

[0085] In an embodiment of the application, as shown in FIG 6, performing speech recognition on the target audio data block to obtain the target information item category of the target audio data block comprises:

[0086] At step S1521 , taking the first target audio data block as the current target audio data block and cyclically executing the following steps: At step S1522, perform speech recognition on the current target audio data block to obtain the text data sub-item of the current target audio data block.

[0087] At step S1523, when the text data sub-item of the current target audio data block contains a preset information item category, storing the information item category of the current target audio data block as the target information item category in the temporary set of information item categories. If the current target audio data block is not the last target audio data block, taking the next target audio data block as the current target audio data block; otherwise, ending the loop.

[0088] At step S1524, when the text data sub-item of the current target audio data block does not contain a preset information item category and the current target audio data block is not the last target audio data block, taking the next target audio data block as the current target audio data block. If the current target audio data block is the last target audio data block, ending the loop.

[0089] For example, the target text data is: "The subscriber you dialed is busy now, please redial later. Sorry, The subscriber you dialed is busy now, please redial later." The text information corresponding to the target audio data blocks is: "The subscriber you dialed is busy now, please redial later" and "Sorry The subscriber you dialed is busy now please redial later", respectively. It is determined that the text information "The subscriber you dialed is busy now, please redial later" of the first target audio data block contains the information item category "busy". The information item category "busy" is taken as the target information item category and stored in the temporary set of information item categories, and the target audio features and the corresponding target information item category are stored in the cache. It is determined that the text information "Sorry The subscriber you dialed is busy now please redial later" of the second target audio data block does not contain any preset information item category, so the loop ends.

[0090] Referring back to FIG 2, at step S160, obtaining the target information item category of the target audio data file based on the target information item categories of the multiple target audio features. The target information item category of the target audio features can represent the target information item category of the target audio data file. In an embodiment, the target information item category of the target audio features can be directly used as the target information item category of the target audio data file.

[0091] In an embodiment of the application, to adapt to more complex situations, for example, when multiple target audio data blocks after segmentation of the target audio data file correspond to different target information item categories, further judgment is required to obtain the target information item category of the target audio data file. Specifically, in an embodiment of the application, the information item categories have weight values, obtaining the target information item category of the target audio data file based on the target information item categories of the multiple target audio features comprises: taking the target information item category with the largest weight value in the temporary set of information item categories as the target information item category of the target audio data file.

[0092] For example, based on statistical experience, the weight value of the information item category "non-existent number" is 3, the weight value of the information item category "power off" is 5, and the weight value of the information item category "busy" is 7. Then, after the above-mentioned similarity calculation or speech recognition, the temporary set of information item categories obtained is {"busy", "power off"}. After judging the weight values, the target information item category of the target audio data file is "busy".

[0093] In an embodiment of the application, as shown in FIG 7A, obtaining the target information item category of the target audio data file based on the target information item categories of the multiple target audio features comprises:

[0094] At step S161 a, counting the number of each target information item category in the temporary set of information item categories;

[0095] At step S162a, taking the target information item category with the largest number as the target information item category of the target audio data file. For example, if the temporary set of information item categories is {"busy", "power off", "power off", "non-existent number"}, then based on the counted numbers, the target information item category of the target audio data file is "power off".

[0096] In some embodiment of the application, the information item categories have weight values. As shown in FIG 7B, obtaining the target information item category of the target audio data file based on the target information item categories of the multiple target audio features comprises:

[0097] At step S161 b, counting the number of each target information item category in the temporary set of information item categories;

[0098] At step S162b, obtaining the composite weight value of each target information item category in the temporary set of information item categories, where the composite weight value is the product of the number of the target information item category and the corresponding weight value;

[0099] At step S163b: taking the target information item category with the largest composite weight value as the target information item category of the target audio data file.

[0100] For example, the weight value of the information item category "non-existent number" is 3, the weight value of the information item category "power off" is 5, and the weight value of the information item category "busy" is 7. If the temporary set of information item categories is {"busy", "power off", "power off", "non-existent number"}, then the composite weight value of the target information item category "busy" is 7*1 =7, the composite weight value of the target information item category "power off" is 5*2=10, and the composite weight value of the target information item category "non-existent number" is 3*1 =3. Therefore, the composite weight value of the target information item category "power off' is the largest, and the target information item category of the target audio data file is "power off".

[0101] The method 200 can perform segmentation processing on structured audio to generate target audio data blocks, extract target audio features, compare the target audio features with cached target audio features, calculate similarity, and when the similarity is greater than or equal to the first threshold, set the information item category of the cached target audio features as the information item category of the corresponding target audio features. When there are no cached target audio features or the similarity with the cached target audio data blocks is less than the first threshold, the target audio features cannot be matched to an information item category. Speech recognition is performed on the structured audio data to obtain the text information of the structured audio, and the information item category of the target audio features is obtained based on the text information. To facilitate the next calculation, the target audio features with obtained information item categories are stored in the cache along with the information item categories. In the process of identifying the information item category of structured audio in the method of the embodiment of the application, the invocation of speech recognition technology is reduced, thereby reducing the occupation of speech recognition servers. Compared with the method of full-text recognition, the recognition efficiency is higher, making it suitable for large-scale application scenarios.

[0102] In an embodiment of the application, one characteristic of structured audio is that audio with the same information item category has a high degree of signal feature matching within the normal playback interval.

[0103] An embodiment of the application adopts a segmentation approach, dividing an audio file into multiple segments containing valid information, extracting signal features from each segment, and then performing similarity matching with cached segment features. As shown in FIG 8A, a method for identifying information item categories of structured audio data is provided, with specific steps as follows:

[0104] At step S210, audio file reading is carried out.

[0105] Audio files are generally large and cannot be processed entirely. However, during the initialization phase, the audio file is loaded into memory for subsequent processing.

[0106] Reading the audio file can convert it into a byte array. Assuming a sampling rate of 8000 and each sample is 16 bits, the data volume for 1 second is: 8000 * (16 / 8) = 16000 bytes. Assuming a sampling rate of 16000 and each sample is 16 bits, the data volume for 1 second is: 16000 * (16 / 8) = 32000 bytes. At step S220, segmentation processing is carried out. Voice Activity Detection (VAD) technology is used for segmentation to achieve three goals: eliminating silent signals that lack features and value; removing background noise signals while retaining valid voice signals; and ensuring that segmented audio data (target audio data blocks) contain independent sentences, with each segment's features (target audio features of target audio data blocks) being unique.

[0107] At step S230, extracting all Segment features is carried outas the data volume of each segment is still large, necessitating further frame processing for each segment.

[0108] It is generally recommended to use 10ms-30ms as one frame, with this solution adopting 20ms as one frame, where the signal features within this frame are shortterm stable.

[0109] Assuming one sentence corresponds to one segment and is approximately 10 seconds long, each segment contains 500 frames. Each frame is a short-term stable signal from which one feature can be extracted. Thus, the number of features for each segment is 500.

[0110] In an embodiment of the application, as shown in FIG 8B, the steps for extracting segment features are as follows:

[0111] At step S231 , signal preprocessing is carried out by normalizing the signal and then applying voice emphasis processing.

[0112] At step S232, frame processing is carried out. For example, using 20ms as one frame, at a sampling rate of 8000, each frame contains 8000 * (20 / 1000) = 160 samples.

[0113] At step S233, windowing processing is carried out by applying smoothing processing.

[0114] At step S234, extraction of features from each frame is carried out. An embodiment of the application adopts time-domain features of the voice signal: short-time zerocrossing rate and short-time energy. At step S235, returning each segment feature by merging each frame feature into segment features and returning a feature value array (zero-crossing rate values). For a 10-second segment, returning an array containing 500 feature values.

[0115] Referring to FIG 8A, at step S240, performing similarity matching between segment features and cached segment features. An embodiment of the application uses the Dynamic Time Warping (DTW) algorithm for similarity matching. DTW is more suitable for similarity matching in this scenario than general distance algorithms because the lengths of the data to be compared and the template data are inconsistent, while distance algorithms require consistent lengths. Note that during the system initialization phase, cached segment features are empty and are gradually loaded through subsequent step S260.

[0116] At step S250, if any segment feature matches cached segment features, returning the corresponding information item category of the cached segment features.

[0117] At step S260, identifying segment information item categories. When all segment features do not match cached features, obtain the text information of the segment through ASR and add valid segment features to the cache.

[0118] As shown in FIG 8C, this specifically comprises:

[0119] At step S261 , invoking ASR to recognize the complete audio file and returning the recognized text information.

[0120] At step S262, if the text information does not contain valid information item categories, returning null.

[0121] At step S263, if the text information contains valid information item categories, performing further segmented ASR and adding target audio features containing message information to the cache for subsequent recognition of other files.

[0122] In an embodiment of the application, the cache structure is as follows: key = the sum of all values in the segment feature array; value = a list of segment feature objects, where each segment feature object contains the segment feature array and the information item category.

[0123] At step S264, returning the information item category of the segment.

[0124] Referring to FIG 8A, at step S270, returning the information item category of the audio file.

[0125] In an embodiment of the application, 3325 audio files were randomly selected from the business system to identify information item categories using this technical solution. Among them, 2939 files were identified with information categories, while the remaining 386 files were not identified and still required ASR recognition. 88% of the audio files could be directly identified without requesting ASR. Compared to the time spent on full ASR requests, which was approximately 82 minutes, the time was reduced to 9 minutes using this solution. ASR server resources were greatly released, and time efficiency was significantly improved.

[0126] An embodiment of the application can be applied to call status refinement recognition in intelligent outbound call scenarios. In intelligent outbound call scenarios, there are a large number of unanswered calls, and the call status needs to be identified. The specific process is as follows: initiate robot outbound calls, record the "unanswered" flag for unanswered calls, and save the media audio file to the file system. A scheduled task scans "unanswered" call records and retrieves associated media audio files. The method in the embodiment of the application is called to obtain the information item category, such as "unallocated number" or "out of service." The information item category is then updated in the "unanswered" call record.

[0127] An embodiment of the application can identify the text information item categories contained of structured audio data. Here are two examples:

[0128] Text information: "The number you dialed is unallocated. Please check and dial again."; Information item category: "Unallocated number";

[0129] Text information: "The subscriber you dialed is out of service. Please dial again later.";

[0130] Information item category: "Out of service."

[0131] An embodiment of the application can reduce the consumption of ASR resources and improve recognition efficiency.

[0132] In an embodiment of the application, as shown in FIG 9, a system 900 for identifying information item categories of structured audio data is illustrated, comprising a first acquisition module 910, a segmentation module 920, a feature extraction module 930, a similarity matching module 940, a speech recognition module 950, and a second acquisition module 960, wherein: the first acquisition module 910 is configured to acquire the target audio data file, which is structured audio data; the segmentation module 920 is configured to segment the target audio data file to obtain multiple target audio data blocks; the feature extraction module 930 is configured to extract audio features of the target audio data blocks to obtain target audio features; the similarity matching module 940 is configured to, when preset audio features are stored in a cache, calculate the similarity between the target audio features and each of the preset audio features, wherein the preset audio features have preset information item categories, and when the similarity is greater than or equal to the first threshold, set the preset information item category of the corresponding preset audio feature as the target information item category of the target audio feature; the speech recognition module 950 is configured to, when no preset audio features are stored in the cache or the similarity between the target audio features and the preset audio features is less than the first threshold, perform speech recognition on the target audio data file to obtain the target information item category of the target audio data blocks of the target audio data file, assign the target information item category corresponding to the target audio data blocks to the target audio features, and store the target audio features as preset audio features in the cache; the second acquisition module 960 is configured to obtain the target information item category of the target audio data file based on the target information item categories of the multiple target audio features.

[0133] In an embodiment of the application, the segmentation module 920 is specifically configured to: sample the target audio data file to obtain a first sampled file; invoke a speech segmentation processing model to segment the first sampled file to obtain the multiple target audio data blocks.

[0134] In an embodiment of the application, the feature extraction module 930 is specifically configured to: perform signal preprocessing on the target audio data blocks; according to the preset standard frame length, perform frame processing on the preprocessed target audio data blocks to obtain a frame data item set, wherein the frame data item set comprises multiple frame data items; perform windowing processing on the frame data items to extract frame features of the frame data items; combine the frame features of all frame data items corresponding to the target audio data blocks to form the target audio features of the target audio data blocks. In an embodiment of the application, the frame features include the short-time zerocrossing rate and short-time energy of the frame data items.

[0135] In an embodiment of the application, the similarity matching module 940 is specifically configured to: invoke dynamic time warping model to calculate the similarity between the target audio features and the preset audio features.

[0136] In an embodiment of the application, the speech recognition module 950 is specifically configured to: perform speech recognition on the target audio data file to obtain target text data; when the target text data contains a preset information item category, perform speech recognition on the target audio data blocks to obtain the target information item category of the target audio data blocks; when the target text data does not contain a preset information item category, the recognition fails.

[0137] In an embodiment of the application, the speech recognition module 950 is specifically further configured to: take the first target audio data block as the current target audio data block, and cyclically perform the following steps: perform speech recognition on the current target audio data block to obtain text data sub-items of the current target audio data block; when the text data sub-items of the current target audio data block contain a preset information item category, store the information item category of the current target audio data block as a target information item category in a temporary set of information item categories, and when the current target audio data block is not the last target audio data block, take the next target audio data block as the current target audio data block, otherwise, end the cycle; when the text data sub-items of the current target audio data block do not contain a preset information item category and the current target audio data block is not the last target audio data block, take the next target audio data block as the current target audio data block, and when the current target audio data block is the last target audio data block, end the cycle.

[0138] In an embodiment of the application, the information item categories have weight values, and the second acquisition module 960 is configured to: take the target information item category with the largest weight value in the temporary set of information item categories as the target information item category of the target audio data file.

[0139] In an embodiment of the application, the second acquisition module 960 is configured to: count the number of each target information item category in the temporary set of information item categories; take the target information item category with the largest number as the target information item category of the target audio data file.

[0140] In an embodiment of the application, the information item categories have weight values, and the second acquisition module 960 is configured to: count the number of each target information item category in the temporary set of information item categories; obtain the composite weight value of each target information item category in the temporary set of information item categories, where the composite weight value is the product of the number of target information item categories and the corresponding weight values; take the target information item category with the largest composite weight value as the target information item category of the target audio data file.

[0141] An embodiment of the application proposes a system for identifying information item categories of structured audio data, which can segment structured audio to generate target audio data blocks, extract target audio features, compare the target audio features with cached target audio features, calculate similarity, and when the similarity is greater than or equal to the first threshold, set the information item category of the cached target audio features as the information item category of the corresponding target audio features; when there are no cached target audio features in the cache or the similarity with the cached target audio data blocks is less than the first threshold, the target audio features cannot be matched to an information item category, so speech recognition is performed on the structured audio data to obtain text information of the structured audio, and the information item category of the target audio features is obtained based on the text information; in order to facilitate the next calculation, the target audio features with the obtained information item categories are stored in the cache along with the information item categories. In the process of identifying information item categories of structured audio in the system of the embodiment of the application, the invocation of speech recognition technology is reduced, thereby reducing the occupation of speech recognition servers. Compared with the method of full-text recognition, the recognition efficiency is high, and it is suitable for large-scale application scenarios.

[0142] In an embodiment, the system can be combined with the features of the method and system for identifying information item categories of structured audio data in any embodiment, and vice versa, which will not be repeated here.

[0143] In an embodiment of the application, an electronic device is provided, comprising: a processor and a memory storing computer programs, wherein the processor is configured to execute any method for identifying the information item category of structured audio data according to the embodiment of the application when running the computer programs.

[0144] FIG 10 illustrates a schematic diagram of an electronic device 1000 that can be configured to implement the method of an embodiment of the application or realize the electronic device of an embodiment of the application. In an embodiment, it may include more or fewer electronic devices than those shown in the figure. In an embodiment, implementation may utilize a single or multiple electronic devices. In an embodiment, implementation may utilize cloud-based or distributed electronic devices.

[0145] As shown in FIG 10, the electronic device 1000 comprises a processor 1001 , which can execute various appropriate operations and processing based on programs and / or data stored in a read-only memory (ROM) 1002 or programs and / or data loaded into a random access memory (RAM) 1003 from a storage section 1008. In some embodiment, the processor may include a central processing unit (CPU) and / or other processing units. The processor 1001 may be a multi-core processor or may include multiple processors. In some embodiment, the processor 1001 may include a general- purpose main processor and one or more special coprocessors, such as a graphics processing unit (GPU), a neural network processing unit (NPU), a digital signal processor (DSP), and the like. The RAM 1003 also stores various programs and data required for the operation of the electronic device 1000. The processor 1001 , the ROM 1002, and the RAM 1003 are connected to each other through a bus 1004. An input / output (I / O) interface 1005 is also connected to the bus 1004.

[0146] The aforementioned processor and memory are collectively used to execute programs stored in the memory. When executed by a computer, these programs can implement the steps or functions of the method for identifying the information item category of structured audio data described in various embodiment.

[0147] The following components are connected to the I / O interface 1005: an input section 1006 including a keyboard, a mouse, a touch screen, etc.; an output section 1007 including devices such as a cathode ray tube (CRT), a liquid crystal display (LCD), speakers, etc.; a storage section 1008 including hard disks, etc.; and a communication section 1009 including network interface cards such as LAN cards, modems, etc. The communication section 1009 performs communication processing via a network such as the Internet. A drive 1010 is also connected to the I / O interface 1005 as needed. Removable media 1011 , such as magnetic disks, optical disks, magneto-optical disks, semiconductor memories, etc., are installed on the drive 1010 as needed, so that computer programs read from them can be installed into the storage section 1008 as needed. Only some components are schematically shown in Figure 10, which does not mean that the computer system 1000 only comprises the components shown in FIG 10.

[0148] In some embodiment, the aforementioned electronic device 1000 refers to a mobile terminal, including a mobile phone, an in-vehicle terminal, a smart TV, etc. taking a mobile phone as an example, the electronic device 1000 also comprises a touch- enabled display screen, external speakers, a gyroscope, a camera, 4G / 5G antennas, and other device modules.

[0149] The systems, modules, or units illustrated in the above embodiment can be implemented by a computer or its associated components. The computer may be, for example, a mobile terminal, a smartphone, a personal computer, a laptop computer, an in-vehicle human-computer interaction device, a personal digital assistant, a media player, a navigation device, a game console, a tablet computer, a wearable device, a smart TV, an loT system, a smart home, an industrial computer, a server, or a combination of them.

[0150] Although not shown, in the embodiment of the application, a storage medium is provided, which stores computer programs configured to execute any method for task scheduling and monitoring according to the embodiment of the application when run.

[0151] In the embodiment of the application, the storage medium comprises permanent and non-permanent, removable and non-removable items that can implement information storage by any method or technology. Examples of storage medium include, but are not limited to, Phase-change Random Access Memory (PRAM), Static Random- Access Memory (SRAM), Dynamic Random Access Memory (DRAM), other types of Random Access Memory (RAM), Read-Only Memory (ROM), Electrically Erasable Programmable Read-Only Memory (EEPROM), flash memory or other memory technologies, Compact Disc Read-Only Memory (CD-ROM), Digital Versatile Disc (DVD) or other optical storage, magnetic tape cassettes, magnetic tapes, magnetic disk storage or other magnetic storage devices, or any other non-transmission medium that can be used to store information accessible by computing devices.

[0152] The methods, programs, systems, etc., in the embodiment of the application can be executed or implemented in a single or multiple networked computers and can also be practiced in a distributed computing environment. In the distributed computing environment of these embodiment, tasks can be performed by remote processing devices connected through a communication network.

[0153] Those skilled in the art should understand that the embodiment of the present specification can be provided as methods, systems, or computer program products. Therefore, it is conceivable to those skilled in the art that the implementation of the functional modules / units or controllers and the related method steps illustrated in the above embodiment can be achieved through software, hardware, or a combination of software and hardware.

[0154] Unless explicitly indicated, the actions or steps of the methods and programs according to the embodiment of the application do not necessarily need to be executed in a specific order and can still achieve the desired results. In some embodiment, multitasking and parallel processing are also possible or may be advantageous.

[0155] For the sake of brevity, the descriptions of the embodiment are not exhaustive, and features or parts that are the same or similar between the embodiment may be omitted. In this document, "an embodiment," "some embodiment," "example," "specific example," or "some examples" refer to being applicable to at least one embodiment or example according to the application, rather than all embodiment. The aforementioned terms do not necessarily mean referring to the same embodiment or example. Without contradiction, those skilled in the art can combine and integrate different embodiment or examples described in this specification and the features of different embodiment or examples. The exemplary systems and methods of the application have been specifically shown and described with reference to the above embodiment, which are only examples of the best mode for implementing the systems and methods. Those skilled in the art can understand that various changes can be made to the embodiment of the systems and / or methods described here when implementing the systems and / or methods without departing from the spirit and scope of the application defined in the appended claims.

[0156] Throughout this specification and claims which follow, unless the context requires otherwise, the word “comprise", and variations such as “comprises” or “comprising”, will be understood to imply the inclusion of a stated integer or group of integers or steps but not the exclusion of any other integer or group of integers.

[0157] Persons skilled in the art will appreciate that numerous variations and modifications will become apparent. All such variations and modifications which become apparent to persons skilled in the art, should be considered to fall within the spirit and scope that the invention broadly appearing before described.

Claims

CLAIMS1.A method for identifying information item category of structured audio data, characterized in that the method comprising the following steps: acquiring the target audio data file, wherein the target audio data file is structured audio data; segmenting the target audio data file to obtain multiple target audio data blocks; extracting audio features of the target audio data blocks to obtain target audio features; when preset audio features are stored in the cache, calculating the similarity between each target audio feature and each preset audio feature, wherein the preset audio features have preset information item categories, and when the similarity is greater than or equal to the first threshold, setting the preset information item category of the corresponding preset audio feature as the target information item category of the target audio feature; when no preset audio features are stored in the cache or the similarity between all target audio features and the preset audio features is less than the first threshold, performing speech recognition on the target audio data file to obtain the target information item category of the target audio data blocks of the target audio data file, assigning the target information item category corresponding to the target audio data block to the target audio feature, and storing the target audio feature as a preset audio feature in the cache; acquiring the target information item category of the target audio data file based on the target information item categories of the multiple target audio features.

2. The method according to claim 1 , characterized in that segmenting the target audio data file to obtain multiple target audio data blocks comprises: sampling the target audio data file to obtain a first sampled file; invoking speech segmentation model to segment the first sampled file to obtain the multiple target audio data blocks.

3. The method according to claim 1 , characterized in that extracting audio features of the target audio data blocks to obtain target audio features comprises: performing signal preprocessing on the target audio data blocks;performing frame segmentation on the preprocessed target audio data blocks according to the preset standard frame length to obtain a frame data item set, wherein the frame data item set comprises multiple frame data items; performing windowing processing on the frame data items to extract frame features of the frame data items; combining the frame features of all frame data items corresponding to the target audio data block to form the target audio feature of the target audio data block.

4. The method according to claim 3, characterized in that the frame features include the short-time zero-crossing rate and short-time energy of the frame data items.

5. The method according to claim 1 , characterized in that calculating the similarity between the target audio feature and the preset audio feature comprises: invoking dynamic time warping model to calculate the similarity between the target audio feature and the preset audio feature.

6. The method according to claim 1 , characterized in that performing speech recognition on the target audio data file to obtain the target information item category of the target audio data block comprises: performing full-text speech recognition on the target audio data file to obtain target text data; when the target text data contains a preset information item category, performing speech recognition on the target audio data block of the target audio data file to obtain the target information item category of the target audio data block; when the target text data does not contain a preset information item category, the recognition fails.

7. The method according to claim 6, characterized in that performing speech recognition on the target audio data block of the target audio data file to obtain the target information item category of the target audio data block comprises: taking the first target audio data block as the current target audio data block and cyclically executing the following steps: performing speech recognition on the current target audio data block to obtain text data sub-items of the current target audio data block;when the text data sub-items of the current target audio data block contain a preset information item category, storing the information item category of the current target audio data block as a target information item category in a temporary information item category set; when the current target audio data block is not the last target audio data block, taking the next target audio data block as the current target audio data block, otherwise ending the cycle; when the text data sub-items of the current target audio data block do not contain a preset information item category and the current target audio data block is not the last target audio data block, taking the next target audio data block as the current target audio data block; when the current target audio data block is the last target audio data block, ending the cycle.

8. The method according to claim 7, characterized in that the information item categories have weight values, and acquiring the target information item category of the target audio data file based on the target information item categories of the multiple target audio features comprises: taking the target information item category with the largest weight value in the temporary information item category set as the target information item category of the target audio data file.

9. The method according to claim 7, characterized in that acquiring the target information item category of the target audio data file based on the target information item categories of the multiple target audio features comprises: counting the number of each target information item category in the temporary information item category set; taking the target information item category with the largest number as the target information item category of the target audio data file.

10. The method according to claim 7, characterized in that the information item categories have weight values, and acquiring the target information item category of the target audio data file based on the target information item categories of the multiple target audio features comprises: counting the number of each target information item category in the temporary information item category set;obtaining the composite weight value of each target information item category in the temporary information item category set, wherein the composite weight value is the product of the number of the target information item category and the corresponding weight value; taking the target information item category with the largest composite weight value as the target information item category of the target audio data file.

11. A system for identifying information item category of structured audio data, characterized in that the system including a first acquisition module, a segmentation module, a feature extraction module, a similarity matching module, a speech recognition module, and a second acquisition module, wherein: the first acquisition module is configured to acquire the target audio data file, wherein the target audio data file is structured audio data; the segmentation module is configured to segment the target audio data file to obtain multiple target audio data blocks; the feature extraction module is configured to extract audio features of the target audio data blocks to obtain target audio features; the similarity matching module is configured to, when preset audio features are stored in the cache, calculate the similarity between the target audio features and each preset audio feature, wherein the preset audio features have preset information item categories, and when the similarity is greater than or equal to the first threshold, set the preset information item category of the corresponding preset audio feature as the target information item category of the target audio feature; the speech recognition module is configured to, when no preset audio features are stored in the cache or the similarity between the target audio features and the preset audio features is less than the first threshold, perform speech recognition on the target audio data file to obtain the target information item category of the target audio data blocks of the target audio data file, assign the target information item category corresponding to the target audio data block to the target audio feature, and store the target audio feature as a preset audio feature in the cache; the second acquisition module is configured to acquire the target information item category of the target audio data file based on the target information item categories of the multiple target audio features.

12. A computer readable storage medium storing a computer program, wherein the program, when executed by a processor, implements any one of the methods according to any of claims 1 -10.

13. An electronic device, including: a processor and a memory storing computer programs, wherein the processor is configured to execute any one of the methods according to any of claims 1 -10 when running the computer programs.

Images