A method and system for intelligent sound recognition in voice-assisted examination paper marking

Abstract

The present application relates to the field of speech processing technology, and specifically to a method and system for intelligent sound recognition in speech-assisted paper marking, the method comprising: collecting speech signals in speech-assisted paper marking; obtaining all connected domains of the pitch period spectrogram of the speech signal; determining the falseness of the pitch period trajectory corresponding to each connected domain in the pitch period spectrogram to obtain a voiced segment speech signal set and a non-voiced segment speech signal set of the speech signal; determining the unvoiced sound suspicion of each frame of speech signal in the non-voiced segment speech signal set to obtain a voiced segment speech signal set of the speech signal; determining the short-time energy high threshold value, short-time energy low threshold value and short-time average zero-crossing rate threshold value of the dual threshold endpoint detection algorithm to determine the speech recognition result. The present application can improve the accuracy of speech recognition in speech-assisted paper marking.

Description

Technical Field

[0001] The present application relates to the field of speech processing technology, and in particular to a method and system for intelligent sound recognition for speech-assisted examination paper marking. Background Art

[0002] When teachers are marking students' assignments online, they can use the voice recognition system to give comments and suggestions directly using voice. The system will convert the voice into text for students to view. However, teachers may pause or include non-verbal sounds when giving comments. The endpoint detection algorithm can help the system more accurately identify the beginning and end of the teacher's speech, avoiding the teacher's thinking time or irrelevant background sounds as valid voice input and affecting the voice recognition results.

[0003] The double threshold endpoint detection algorithm is a commonly used endpoint detection algorithm. The algorithm is relatively simple, easy to implement, and has low computational complexity, making it suitable for real-time systems. However, the performance of the algorithm depends to a large extent on the selection of threshold parameters. Different teachers' speaking habits, such as speaking speed and pause frequency, make it difficult to determine and adjust the value of the threshold parameters in the algorithm. Improper threshold parameter selection may cause the voiced segment in the speech signal to be misjudged as a silent segment, resulting in the loss of useful information, which in turn affects the final speech recognition result. Summary of the invention

[0004] In order to solve the above technical problems, a method and system for intelligent sound recognition for voice-assisted marking are provided to solve the existing problems.

[0005] The solution to the technical problem of the present application is to provide a method and system for intelligent voice recognition of voice-assisted examination papers, including the following steps:

[0006] In a first aspect, an embodiment of the present application provides a method for intelligent voice recognition of voice-assisted examination paper marking, the method comprising the following steps:

[0007] Collect voice signals in voice-assisted marking;

[0008] Based on the speech signal, all connected domains of the pitch period spectrogram of the speech signal are obtained;

[0009] Each connected domain corresponds to a pitch period track of the pitch period spectrogram, and the false degree of the pitch period track corresponding to each connected domain in the pitch period spectrogram is determined according to the gray value of the pixel points in each connected domain of the pitch period spectrogram and the distance of the horizontal coordinates of the pixel points in each connected domain, so as to obtain a voiced segment speech signal set and a non-voiced segment speech signal set of the speech signal;

[0010] Determining the unvoiced sound suspicion degree of each frame of speech signal in the non-voiced sound segment speech signal set according to the energy change and discreteness of each frame of speech signal in the non-voiced sound segment speech signal set, so as to obtain the unvoiced sound segment speech signal set of the speech signal;

[0011] Based on the voiced segment speech signal set and the unvoiced segment speech signal set, the short-time energy high threshold value, the short-time energy low threshold value and the short-time average zero-crossing rate threshold value of the double threshold endpoint detection algorithm are determined to determine the speech recognition result.

[0012] Preferably, the step of acquiring all connected domains of the pitch period spectrogram of the speech signal includes:

[0013] The speech signal is divided into frames and windows are added to obtain each frame of speech signal and its corresponding sampling time interval;

[0014] The pitch period spectrogram of the speech signal after frame division and windowing is grayed out, a threshold segmentation algorithm is used for the pitch period spectrogram after graying, and connected domain extraction is performed to obtain all connected domains of the pitch period spectrogram.

[0015] Preferably, the step of determining the falseness degree of the pitch period trajectory corresponding to each connected domain in the pitch period spectrum graph comprises:

[0016] The maximum value of the difference between the horizontal coordinates of all two arbitrary pixel points in each connected domain of the pitch period spectrum is used as the trajectory length of the pitch period trajectory corresponding to each connected domain of the pitch period spectrum;

[0017] The mean of the gray values ​​of all pixels in each connected domain of the pitch period spectrum is used as the color darkness of the pitch period trajectory corresponding to each connected domain of the pitch period spectrum;

[0018] The ratio of the trajectory length to the color dimness is used as the falseness degree of the pitch period trajectory corresponding to each connected domain in the pitch period spectrum.

[0019] Preferably, the acquiring of the voiced segment speech signal set and the unvoiced segment speech signal set of the speech signal includes:

[0020] Based on the false degree, obtaining a connected domain set;

[0021] Obtain the minimum and maximum values ​​of the horizontal coordinates of all pixels in each connected domain in the connected domain set to form the interval range of each connected domain;

[0022] Recording the union of the interval ranges of all connected domains in the connected domain set as the union interval of the connected domain set;

[0023] All the frame speech signals corresponding to the intersection between the union interval and the sampling time interval are selected to form a voiced segment speech signal set; and all the remaining frame speech signals are selected to form a non-voiced segment speech signal set.

[0024] Preferably, the acquiring the connected domain set includes:

[0025] A threshold segmentation algorithm is adopted to obtain a segmentation threshold of the falsehood degree of the pitch period trajectory corresponding to all connected domains in the pitch period spectrogram of the speech signal; all connected domains with a falsehood degree greater than the segmentation threshold are combined into a connected domain set.

[0026] Preferably, the determining of the unvoiced sound suspicion degree of each frame of speech signal in the non-voiced sound segment speech signal set includes:

[0027] Obtaining a spectrogram of each frame of speech signal in a non-voiced segment speech signal set;

[0028] The product of the mean value and the discrete degree of all energies in the spectrum of each frame of speech signal is taken as the unvoiced sound suspicion degree of each frame of speech signal in the non-voiced segment speech signal set.

[0029] Preferably, the step of acquiring a speech signal set of unvoiced segments of a speech signal includes:

[0030] A threshold segmentation algorithm is adopted to obtain the segmentation threshold of the unvoiced sound suspicion degree of all frame speech signals in the non-voiced sound segment speech signal set; and all frame speech signals whose unvoiced sound suspicion degree is greater than the segmentation threshold are combined into an unvoiced sound segment speech signal set.

[0031] Preferably, the determination of the short-time energy high threshold value, the short-time energy low threshold value and the short-time average zero-crossing rate threshold value of the dual-threshold endpoint detection algorithm includes:

[0032] Calculate the mean of the short-time average energy of all frame speech signals in the voiced segment speech signal set as the short-time energy high threshold value in the double threshold endpoint detection algorithm;

[0033] Calculate the short-time average energy and the short-time average zero-crossing rate of each frame of speech signal in the unvoiced segment speech signal set respectively;

[0034] The minimum values ​​of the short-time average energy and the short-time average zero-crossing rate are respectively used as the short-time energy low threshold value and the short-time average zero-crossing rate threshold value in the dual-threshold endpoint detection algorithm.

[0035] Preferably, the determining of the speech recognition result includes:

[0036] A double-threshold endpoint detection algorithm is used on the speech signal to obtain each speech signal segment;

[0037] All speech signal segments are arranged in chronological order, and all the sorted speech signal segments are converted into text in sequence for output to obtain speech recognition results.

[0038] In a second aspect, an embodiment of the present application further provides a sound intelligent recognition system for voice-assisted marking, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, wherein when the processor executes the computer program, the steps of any one of the above-mentioned sound intelligent recognition methods for voice-assisted marking are implemented.

[0039] This application has at least the following beneficial effects:

[0040] The present application analyzes the speech signal, obtains the pitch period spectrum of the speech signal, and performs grayscale processing on it to obtain all connected domains of the pitch period spectrum. The beneficial effect of the application is that multiple pitch period trajectories formed in the pitch period spectrum are extracted and analyzed, which is beneficial to the subsequent analysis of the change characteristics of the pitch period trajectory corresponding to each connected domain; the false degree of the pitch period trajectory corresponding to each connected domain in the pitch period spectrum is determined to obtain the voiced segment speech signal set and the non-voiced segment speech signal set of the speech signal. The beneficial effect of the application is that the pitch period trajectory characteristics of the voiced segment speech signal generated by the Chinese character finals in the speech signal when the user pronounces the speech signal are taken into account, the possibility of the speech signal corresponding to the false pitch trajectory in the pitch period spectrum being identified as the voiced segment speech signal is reduced, and the speech signal carrying important The discrimination between the voiced segment speech signal with language information and the rest of the speech signals; determining the unvoiced sound suspicion of each frame of speech signal in the non-voiced segment speech signal set to obtain the unvoiced segment speech signal set of the speech signal, which has the beneficial effect of improving the discrimination between the unvoiced segment speech signal carrying less language information in the speech signal and the silent segment speech signal carrying no language information; determining the short-time energy high threshold value, the short-time energy low threshold value and the short-time average zero-crossing rate threshold value of the double threshold endpoint detection algorithm to obtain the speech recognition result, which has the beneficial effect of improving the retention degree of the voiced segment speech signal and the unvoiced speech signal carrying language information in the speech signal, reducing the possibility of the voiced segment speech signal in the speech signal being misjudged as the silent segment and causing the loss of useful information, thereby improving the accuracy of speech recognition in speech-assisted marking. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] The following is a further detailed description of a method for intelligent sound recognition for voice-assisted examination paper marking in accordance with the present application in conjunction with the accompanying drawings.

[0042] Figure 1 A flowchart of the steps of a method for intelligent voice recognition of voice-assisted examination paper marking provided in an embodiment of the present application;

[0043] Figure 2A flowchart of the steps of a method for acquiring a voiced segment speech signal set and a non-voiced segment speech signal set of a speech signal provided in an embodiment of the present application;

[0044] Figure 3 A flowchart of the steps of a method for acquiring a set of unvoiced segment speech signals of a speech signal provided in an embodiment of the present application. DETAILED DESCRIPTION

[0045] In order to make the purpose, technical solution and advantages of the present application more clear, the following is a further detailed description of a method and system for intelligent voice recognition for voice-assisted examination paper marking proposed in the present application in combination with the accompanying drawings and implementation examples. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0046] Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs.

[0047] See also Figure 1 , which shows a flowchart of a method for intelligent voice recognition of voice-assisted examination paper marking provided by an embodiment of the present application, the method comprising the following steps:

[0048] Step 1: Collect voice signals in voice-assisted marking.

[0049] When the user clicks the start button of the voice comment in the online scoring system, the built-in voice sensor of the system starts to collect the user's voice signal, and when the user clicks the end button, the collection of the voice signal stops. Secondly, the collected voice signal is denoised to reduce the impact of the noise introduced during the collection and transmission of the voice signal on the subsequent processing process.

[0050] Furthermore, the denoising speech signal is pre-emphasized to compensate for the high-frequency portion of the sound emitted by the user in the collected speech signal and to reduce the effect of the attenuation of the high-frequency energy in the sound emitted by the user on subsequent processing processes, thereby obtaining a pre-emphasized speech signal. The pre-emphasis process is a well-known technology and the specific process will not be repeated here.

[0051] Preferably, in this embodiment, the sampling frequency and quantization value of the speech sensor are set to 15kHz and 16bit respectively, which can be set by the implementer. The spectral subtraction method is used to denoise the speech signal, wherein the spectral subtraction method is a well-known technology and the specific process will not be repeated. As other implementation methods, the implementer can adopt other methods of the prior art, such as adaptive filters and Wiener filtering for denoising, and this embodiment does not impose any special restrictions on this.

[0052] At this point, the voice signal is obtained.

[0053] Step 2: Based on the speech signal, all connected domains of the pitch period of the speech signal are obtained.

[0054] When the user makes a speech comment, the speech signal collected by the speech sensor usually includes the voiced speech signal emitted by the user and the silent speech signal generated by the background noise. The voiced speech signal usually includes the voiced speech signal and the unvoiced speech signal. The sound produced by the vibration of the vocal cords when the user pronounces is called voiced sound, and the sound produced by the non-vibration of the vocal cords when the user pronounces is called unvoiced sound. The voiced sound will produce a certain vibration period, and the corresponding period is called the fundamental pitch period.

[0055] The finals of Chinese characters usually carry very important semantic information, and the finals part of the Mandarin spoken by the user usually exists in the voiced part of the speech signal collected by the speech sensor. When the user is reviewing a test paper, the user's emotion is usually relatively calm, and the tone of voice usually does not fluctuate greatly, and will remain in a relatively stable state. The vibration frequency of the vocal cords does not change much, and the fundamental frequency of the finals of Chinese characters in the Mandarin spoken by the user will maintain a relatively stable change, so that the fundamental frequency of a continuous voiced segment of the collected speech signal will change smoothly and continuously over time, causing the voiced segment speech signal to form a fundamental frequency track in the fundamental frequency spectrum of the speech signal, while the unvoiced segment speech signal and the silent segment speech signal usually do not have a fundamental frequency, so the unvoiced segment speech signal and the silent segment speech signal do not have a fundamental frequency track in the fundamental frequency spectrum of the collected speech signal.

[0056] Based on the above analysis, multiple pitch period tracks will appear in the pitch period spectrogram of the voiced speech signal. Figure 1 Generally, different colors are set to represent the energy level, and the grayscale image is used to reflect the energy level between pixels in the pitch period spectrum. The larger the energy value of the pixel, the closer the color of the pixel is to black, that is, the grayscale value of the pixel is closer to 0. Therefore, all connected domains in the pitch period spectrum can be extracted through the grayscale value of the pixel in the pitch period spectrum, specifically:

[0057] Performing frame-by-frame windowing processing on the speech signal to obtain each frame of speech signal and the sampling time interval of each frame of speech signal;

[0058] Preferably, in this embodiment, during the frame segmentation and windowing process, the frame length, frame shift, window function, and the length of the window function are set to 30ms, 15ms, Hamming window, and 30ms, respectively. As other implementation methods, the implementer can set them according to the actual situation. Among them, the frame segmentation and windowing process is a well-known technology and will not be repeated here.

[0059] The cepstrum method is used on the speech signal after frame division and windowing to obtain the pitch period spectrum of the speech signal;

[0060] It should be noted that, in the pitch period spectrum, the horizontal axis is time and the vertical axis is cepstrum frequency, wherein the cepstrum method is a well-known technology and the specific process will not be described in detail.

[0061] The pitch period spectrum is grayed out, and the threshold segmentation algorithm is used to perform threshold segmentation on the gray values ​​of all pixels in the pitch period spectrum after graying to obtain the optimal segmentation threshold;

[0062] The pixel points in the pitch period spectrum whose grayscale value is greater than the optimal segmentation threshold are recorded as foreground parts;

[0063] Performing connected domain extraction on the foreground part to obtain all connected domains of the pitch period spectrum, wherein connected domain extraction is a well-known technology and the specific process will not be described in detail;

[0064] Preferably, in this embodiment, the Otsu threshold segmentation algorithm is used to obtain the optimal segmentation threshold. As other implementation methods, the implementer can adopt other methods of the prior art, for example, constructing a histogram method. This embodiment does not impose any special restrictions on this. The Otsu threshold algorithm is a well-known technology, and the specific process will not be repeated.

[0065] It should be noted that each connected domain represents a pitch period trajectory formed by a segment of a speech signal suspected to be a voiced segment in the speech signal in the pitch period spectrogram.

[0066] At this point, all connected domains of the pitch period spectrogram of the speech signal are obtained.

[0067] Step 3, each connected domain corresponds to a pitch period trajectory of the pitch period spectrogram, and the falsehood degree of the pitch period trajectory corresponding to each connected domain in the pitch period spectrogram is determined according to the grayscale value of the pixel points in each connected domain of the pitch period spectrogram and the distance between the horizontal coordinates of the pixel points in each connected domain, so as to obtain a voiced segment speech signal set and a non-voiced segment speech signal set of the speech signal.

[0068] Furthermore, coughs, pauses, and mouse clicks produced by the user will cause false peaks that do not contain language information to appear in the spectral envelope of the speech signal. This is because a cough is a sudden, strong, non-periodic sound, and its spectral components may be distributed in a wide frequency range, which will produce significant peaks in the spectrum; pauses and intonations usually contain irregular breathing sounds, swallowing sounds, etc., and the non-stationarity of these sounds will produce discontinuous peaks in the spectrum; mouse clicks have obvious transient characteristics, which will produce spikes in the spectrum, which may cause the cepstrum curve of a certain frame of speech signal to have false peaks, and the cepstrum curve The peak represents the pitch period of the speech signal, which will cause a false pitch period trajectory to appear in the pitch period spectrogram. However, in a multi-frame speech signal that is continuous in time, false peaks usually do not appear at the same inverse frequency position at the same time, so that the false pitch period trajectory in the pitch period spectrogram usually has a very short length and a darker color. The voiced segment speech signal is usually produced by the finals with tones in Mandarin uttered by the user, and the frequency of the tone changes continuously in time, so that the pitch period trajectory formed by a continuous voiced segment speech signal generated by the same final in the pitch period spectrogram usually has a longer length and a darker color. Therefore, the false degree is constructed to reflect the possibility of false pitch period trajectory, specifically:

[0069] The maximum value of the difference between the horizontal coordinates of all two arbitrary pixel points in each connected domain of the pitch period spectrum is used as the trajectory length of the pitch period trajectory corresponding to each connected domain of the pitch period spectrum;

[0070] The mean of the gray values ​​of all pixels in each connected domain of the pitch period spectrum is used as the color darkness of the pitch period trajectory corresponding to each connected domain of the pitch period spectrum;

[0071] The ratio of the trajectory length to the color dimness is used as the falseness degree of the pitch period trajectory corresponding to each connected domain in the pitch period spectrum;

[0072] In this embodiment, the method for calculating the false degree of the pitch period trajectory corresponding to each connected domain in the pitch period spectrum is: Among them, A k is the false degree of the pitch period trajectory corresponding to the kth connected domain in the pitch period spectrum, L k is the length of the pitch period trajectory corresponding to the kth connected domain in the pitch period spectrum, μ k is the mean grayscale value of all pixels in the kth connected domain in the pitch period spectrum, ε is a preset value greater than 0 to avoid the denominator being 0. In this embodiment, ε is set to 0.01.

[0073] It should be noted that the darker the color, the closer the overall color of the corresponding connected domain is to white, and the darker the color of the corresponding pitch period trajectory; the smaller the trajectory length and the darker the color, the greater the false degree, which means that the pitch period trajectory is more likely to be a false pitch period trajectory formed in the pitch period spectrogram by a false peak existing in the inverse spectrum curve of a certain frame of speech signal.

[0074] Further, based on the false degree, the speech signal is divided into a voiced segment speech signal and a non-voiced segment speech signal, specifically:

[0075] A threshold segmentation algorithm is used to obtain a segmentation threshold of the false degree of the pitch period trajectory corresponding to all connected domains in the pitch period spectrum of the speech signal; all connected domains with a false degree greater than the segmentation threshold are combined into a connected domain set;

[0076] Preferably, in this embodiment, Otsu threshold segmentation algorithm is used for threshold segmentation, wherein Otsu threshold segmentation algorithm is a well-known technology and will not be described in detail here. As other implementation methods, implementers can use other methods of the prior art to perform threshold segmentation, and this embodiment does not impose any special restrictions on this.

[0077] It should be noted that the connected domain set excludes the connected domains corresponding to the false pitch period trajectory in the pitch period spectrogram, and the connected domain set is composed of the connected domains corresponding to the pitch period trajectory formed by all the suspected voiced segment speech signals in the pitch period spectrogram.

[0078] Obtain the minimum and maximum values ​​of the horizontal coordinates of all pixels in each connected domain in the connected domain set to form the interval range of each connected domain;

[0079] Record the union of the interval ranges of all connected domains in the connected domain set as the union interval of the connected domain set;

[0080] Selecting all the frame speech signals corresponding to the intersection between the union interval and the sampling time interval to form a voiced segment speech signal set; selecting all the remaining frame speech signals to form a non-voiced segment speech signal set;

[0081] It should be noted that in the pitch period spectrogram, the horizontal axis is time and the vertical axis is the cepstrum frequency. The obtained union interval is also an interval consisting of the time of collecting the speech signal, and the sampling time interval of each frame of the speech signal is also an interval consisting of time. Therefore, the two intervals have an intersecting part, thereby obtaining a set of voiced segment speech signals and a set of unvoiced segment speech signals.

[0082] It should be noted that the voiced segment speech signal set reflects the set of all frame speech signals suspected to be voiced segment speech signals in all frame speech signals; the unvoiced segment speech signal set reflects the set of all frame speech signals except the voiced segment speech signals, that is, the set of all frame speech signals suspected to be unvoiced segment speech signals and silent segment speech signals in the speech signals.

[0083] Further, the steps of the method for obtaining the voiced segment speech signal set and the unvoiced segment speech signal set of the speech signal provided in the embodiment of the present application are as follows: Figure 2 shown.

[0084] At this point, a voiced segment speech signal set and a non-voiced segment speech signal set of the speech signal are obtained.

[0085] Step 4: determine the unvoiced sound suspicion of each frame of speech signal in the non-voiced sound segment speech signal set according to the energy change and discreteness of each frame of speech signal in the non-voiced sound segment speech signal set to obtain the unvoiced sound segment speech signal set of the speech signal.

[0086] Furthermore, since the unvoiced speech signals in the speech signals also carry a small amount of semantic information in the user's pronunciation, and the silent speech signals in the speech signals generally do not contain the sounds made by the user, the suspected unvoiced speech signals in the non-voiced speech signal set usually have greater energy, while the suspected silent speech signals in the non-voiced speech signal set usually have extremely low energy or even no energy, and the energy changes between the frequencies in the silent speech signals are relatively gentle, while the energy changes between the frequencies in the unvoiced speech signals are relatively large.

[0087] Based on the above analysis, the non-voiced segment speech signal set is analyzed to construct the unvoiced segment suspicion degree to reflect the possibility that each frame of speech signal in the non-voiced segment speech signal set is an unvoiced segment speech signal, specifically:

[0088] Obtaining a spectrogram of each frame of speech signal in a non-voiced segment speech signal set;

[0089] Preferably, in this embodiment, Fourier transform is used to obtain the spectrum diagram of each frame of speech signal. It should be noted that the horizontal axis of the spectrum diagram represents frequency and the vertical axis represents energy. Among them, Fourier transform is a well-known technology and will not be repeated here. As other implementation methods, the implementer can adopt other methods of the prior art, such as fast Fourier transform, Hilbert-Huang transform, etc., and this embodiment does not impose any special restrictions on this.

[0090] Calculate the mean and discreteness of all energies in the spectrogram of each frame of speech signal in the non-voiced segment speech signal set respectively;

[0091] The product of the mean and the discrete degree is used as the unvoiced sound suspicion degree of each frame of speech signal in the non-voiced segment speech signal set;

[0092] Preferably, in this embodiment, the degree of discreteness is measured by calculating the standard deviation of all energies in the spectrogram of each frame of speech signal in the non-voiced segment speech signal set. As other implementation methods, the implementer may adopt other methods of the prior art, such as variance, coefficient of variation, etc., and this embodiment does not impose any special restrictions on this.

[0093] In this embodiment, the calculation method of the unvoiced sound suspicion degree of each frame of speech signal in the non-voiced segment speech signal set is: n =μ n ×σ n , where C n is the unvoiced suspicion of the nth frame of speech signal in the non-voiced segment speech signal set, μ n is the mean value of all energies in the spectrogram of the nth frame of speech signal in the non-voiced segment speech signal set, σ n is the standard deviation of all energies in the spectrogram of the nth frame of speech signal in the non-voiced segment speech signal set.

[0094] It should be noted that the larger the mean value and the larger the discreteness, the more likely the corresponding frame speech signal is to be an unvoiced segment speech signal, and the larger the unvoiced suspicion, indicating that the possibility that each frame speech signal is suspected to be an unvoiced segment speech signal in the speech signal is greater.

[0095] Further, based on the unvoiced sound suspicion degree, the unvoiced sound segment speech signal is distinguished, specifically:

[0096] Using a threshold segmentation algorithm, obtaining a segmentation threshold of the unvoiced sound suspicion degree of all frame speech signals in the non-voiced sound segment speech signal set; forming an unvoiced sound segment speech signal set with all frame speech signals whose unvoiced sound suspicion degree is greater than the segmentation threshold;

[0097] Preferably, in this embodiment, Otsu threshold segmentation algorithm is used to perform threshold segmentation to obtain a set of unvoiced segment speech signals.

[0098] It should be noted that the unvoiced segment speech signal set reflects a set of all frame speech signals suspected to be unvoiced segment speech signals in the speech signal.

[0099] Further, the steps of the method for obtaining the unvoiced segment speech signal set of the speech signal provided in the embodiment of the present application are as follows: Figure 3 shown.

[0100] At this point, a speech signal set of unvoiced segments of the speech signal is obtained.

[0101] Step 5: Based on the voiced segment speech signal set and the unvoiced segment speech signal set, determine the short-time energy high threshold value, the short-time energy low threshold value and the short-time average zero-crossing rate threshold value in the dual threshold endpoint detection algorithm to determine the speech recognition result.

[0102] Furthermore, the short-time energy high threshold in the dual-threshold endpoint detection algorithm is used to determine the part with higher energy in the speech signal. In the sound emitted by the user, the voiced part usually has higher energy than the unvoiced part and the silent part, and the short-time energy low threshold and the short-time average zero-crossing rate threshold are used to distinguish the silent segment speech signal and the voiced speech signal in the speech signal. Therefore, in order to retain the language information expressed by the user in the process of speaking as much as possible, the short-time energy high threshold, the short-time energy low threshold and the short-time average zero-crossing rate threshold in the dual-threshold endpoint detection algorithm are determined by the voiced segment speech signal set and the unvoiced segment speech signal set, specifically:

[0103] Calculate the mean of the short-time average energy of all frame speech signals in the voiced segment speech signal set as the short-time energy high threshold value in the double threshold endpoint detection algorithm;

[0104] Calculate the short-time average energy and the short-time average zero-crossing rate of each frame of speech signal in the unvoiced segment speech signal set respectively;

[0105] Respectively taking the minimum values ​​of the short-time average energy and the short-time average zero-crossing rate as the short-time energy low threshold value and the short-time average zero-crossing rate threshold value in the dual-threshold endpoint detection algorithm;

[0106] It should be noted that the calculation of the short-time average energy and the short-time average zero-crossing rate is a well-known technique, and the specific process will not be described in detail.

[0107] Furthermore, based on the double threshold endpoint detection algorithm, the speech signal is recognized, specifically:

[0108] A double-threshold endpoint detection algorithm is used on the speech signal to obtain each speech signal segment;

[0109] Arrange all voice signal segments in chronological order, convert all sorted voice signal segments into text in sequence through voice transcription software for output, obtain voice recognition results, and complete the voice recognition of voice comments in the online marking system.

[0110] It should be noted that the dual threshold endpoint detection algorithm is a well-known technology and will not be described in detail here. Secondly, the speech transcription software can adopt iFlytek Hearing software, etc. Among them, the use of speech transcription software is a well-known technology and will not be described in detail here.

[0111] Based on the same inventive concept as the above method, an embodiment of the present application also provides a sound intelligent recognition system for voice-assisted marking, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, wherein when the processor executes the computer program, the steps of any one of the above-mentioned sound intelligent recognition methods for voice-assisted marking are implemented.

[0112] It should be understood that although Figure 1 The steps in the flowchart are shown in sequence as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified in this document, there is no strict order restriction for the execution of these steps, and these steps can be executed in other orders. Moreover, Figure 1 At least part of the steps may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily executed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least part of the sub-steps or stages of other steps.

[0113] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0114] The above-mentioned embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the present application. It should be pointed out that, for those of ordinary skill in the art, several modifications and improvements can be made without departing from the concept of the present application. Therefore, any simple modification, equivalent change and modification made to the above embodiments according to the technical essence of the present application without departing from the content of the technical solution of the present application, shall fall within the protection scope of the technical solution of the present application.

Claims

1. A method for intelligent voice recognition in voice-assisted examination paper marking, characterized in that: The method comprises the following steps: Collect voice signals in voice-assisted marking; Based on the speech signal, obtaining all connected domains of the pitch period spectrogram of the speech signal; Each connected domain corresponds to a pitch period track of the pitch period spectrogram, and the false degree of the pitch period track corresponding to each connected domain in the pitch period spectrogram is determined according to the gray value of the pixel points in each connected domain of the pitch period spectrogram and the distance of the horizontal coordinates of the pixel points in each connected domain, so as to obtain a voiced segment speech signal set and a non-voiced segment speech signal set of the speech signal; Determining the unvoiced sound suspicion degree of each frame of speech signal in the non-voiced sound segment speech signal set according to the energy change and discreteness of each frame of speech signal in the non-voiced sound segment speech signal set, so as to obtain the unvoiced sound segment speech signal set of the speech signal; Based on the voiced segment speech signal set and the unvoiced segment speech signal set, the short-time energy high threshold value, the short-time energy low threshold value and the short-time average zero-crossing rate threshold value of the double threshold endpoint detection algorithm are determined to determine the speech recognition result.

2. A method for intelligent voice recognition for voice-assisted examination paper marking as claimed in claim 1, characterized in that: The method of obtaining all connected domains of the pitch period spectrogram of the speech signal includes: The speech signal is divided into frames and windows are added to obtain each frame of speech signal and its corresponding sampling time interval; The pitch period spectrogram of the speech signal after frame division and windowing is grayed out, a threshold segmentation algorithm is used for the pitch period spectrogram after graying, and connected domain extraction is performed to obtain all connected domains of the pitch period spectrogram.

3. The method for intelligent voice recognition for voice-assisted examination paper marking according to claim 1, characterized in that: Determining the falseness degree of the pitch period trajectory corresponding to each connected domain in the pitch period spectrum graph includes: The maximum value of the difference between the horizontal coordinates of all two arbitrary pixel points in each connected domain of the pitch period spectrum is used as the trajectory length of the pitch period trajectory corresponding to each connected domain of the pitch period spectrum; The mean of the gray values ​​of all pixels in each connected domain of the pitch period spectrum is used as the color darkness of the pitch period trajectory corresponding to each connected domain of the pitch period spectrum; The ratio of the trajectory length to the color dimness is used as the falseness degree of the pitch period trajectory corresponding to each connected domain in the pitch period spectrum.

4. A method for intelligent voice recognition for voice-assisted examination paper marking as claimed in claim 2, characterized in that: The acquiring of the voiced segment speech signal set and the unvoiced segment speech signal set of the speech signal comprises: Based on the false degree, obtaining a connected domain set; Obtain the minimum and maximum values ​​of the horizontal coordinates of all pixels in each connected domain in the connected domain set to form the interval range of each connected domain; Recording the union of the interval ranges of all connected domains in the connected domain set as the union interval of the connected domain set; All the frame speech signals corresponding to the intersection between the union interval and the sampling time interval are selected to form a voiced segment speech signal set; and all the remaining frame speech signals are selected to form a non-voiced segment speech signal set.

5. A method for intelligent voice recognition for voice-assisted examination paper marking as claimed in claim 4, characterized in that: The step of obtaining a connected domain set includes: A threshold segmentation algorithm is adopted to obtain a segmentation threshold of the falsehood degree of the pitch period trajectory corresponding to all connected domains in the pitch period spectrogram of the speech signal; all connected domains with a falsehood degree greater than the segmentation threshold are combined into a connected domain set.

6. The method for intelligent voice recognition for voice-assisted examination paper marking according to claim 1, characterized in that: The determining of the unvoiced sound suspicion degree of each frame of speech signal in the non-voiced sound segment speech signal set includes: Obtaining a spectrogram of each frame of speech signal in a non-voiced segment speech signal set; The product of the mean value and the discrete degree of all energies in the spectrum of each frame of speech signal is taken as the unvoiced sound suspicion degree of each frame of speech signal in the non-voiced segment speech signal set.

7. The method for intelligent voice recognition for voice-assisted examination paper marking according to claim 1, characterized in that: The step of acquiring a voice signal set of unvoiced segments of a voice signal comprises: A threshold segmentation algorithm is adopted to obtain the segmentation threshold of the unvoiced sound suspicion degree of all frame speech signals in the non-voiced sound segment speech signal set; and all frame speech signals whose unvoiced sound suspicion degree is greater than the segmentation threshold are combined into an unvoiced sound segment speech signal set.

8. The method for intelligent voice recognition for voice-assisted examination paper marking according to claim 1, characterized in that: The method of determining the short-time energy high threshold value, the short-time energy low threshold value and the short-time average zero-crossing rate threshold value of the dual-threshold endpoint detection algorithm includes: Calculate the mean of the short-time average energy of all frame speech signals in the voiced segment speech signal set as the short-time energy high threshold value in the double threshold endpoint detection algorithm; Calculate the short-time average energy and the short-time average zero-crossing rate of each frame of speech signal in the unvoiced segment speech signal set respectively; The minimum values ​​of the short-time average energy and the short-time average zero-crossing rate are respectively used as the short-time energy low threshold value and the short-time average zero-crossing rate threshold value in the dual-threshold endpoint detection algorithm.

9. The method for intelligent voice recognition for voice-assisted examination paper marking according to claim 1, characterized in that: The determining of the speech recognition result comprises: A double-threshold endpoint detection algorithm is used on the speech signal to obtain each speech signal segment; All speech signal segments are arranged in chronological order, and all the sorted speech signal segments are converted into text in sequence for output to obtain speech recognition results.

10. A voice intelligent recognition system for voice-assisted examination paper marking, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that: When the processor executes the computer program, the steps of the method for intelligent sound recognition for voice-assisted paper marking as described in any one of claims 1 to 9 are implemented.

Images