A speech transcription method, device, electronic device and storage medium

By separating the vocal signal from the original speech signal and performing error correction processing, the problem of poor speech transcription quality in the prior art is solved, and higher transcription accuracy and quality are achieved.

CN115331671BActive Publication Date: 2024-11-08BEIJING BOHUI TECH CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202210963749.0
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2022-08-11
Publication Date
2024-11-08
Estimated Expiration
2042-08-11

AI Technical Summary

Technical Problem

The existing phonological transcription technology has poor quality and low transcription accuracy due to complex application scenarios, limited model accuracy and insufficient generalization ability.

Method used

By separating the original vocal signal from the original voice signal, performing voice endpoint detection, speech recognition, punctuation prediction, word correction and word correction, combining the background sound prediction model and the voice prediction model, eliminating the influence of background sound and integrating correction Transcription results.

Benefits of technology

It improves the accuracy and quality of speech transcription and improves the accuracy of transcription results.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN115331671B_ABST
    Figure CN115331671B_ABST
Patent Text Reader

Abstract

The present application provides a speech transcription method, device, electronic device and storage medium, including: separating the original human voice signal from the original speech signal; performing speech endpoint detection on the original human voice signal to split the original human voice signal into at least one segment of human voice signal according to the detected speech endpoint; performing speech recognition on each segment of human voice signal to obtain a segment of initial speech text corresponding to each segment of human voice signal; performing punctuation prediction on each segment of initial speech text, adding punctuation to each segment of initial speech text to split each segment of initial speech text into at least one initial sentence; for each initial sentence, determining the word correction result of the entity word in the initial sentence and the character correction result of the single word; according to the word correction result and the character correction result of each initial sentence, the initial sentence is integrated and corrected to obtain the transcription result of the original speech signal. In this way, the accuracy of the transcription result can be improved and the quality of speech transcription can be improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present application relates to the field of speech technology, and in particular to a speech transcription method, device, electronic device and storage medium. Background Art

[0002] With the development of speech technology, speech transcription technology from speech to corresponding text has gradually become popular in daily life. Speech transcription technology mainly performs speech-to-text processing on longer speech to obtain the text content of the speech. At present, speech transcription technology has been widely used in various application scenarios including content recognition of voice recorders and subtitle generation.

[0003] In the existing technology, traditional acoustic models or deep learning models are often used to recognize the text content of speech. However, due to the complex application scenarios of speech transcription, limited accuracy of the model, insufficient generalization ability and other reasons, the quality of speech transcription is currently poor and the accuracy of transcription is low. Summary of the invention

[0004] In view of this, the purpose of this application is to provide a speech transcription method, device, electronic device and storage medium, which can separate the original human voice signal from the original speech signal to eliminate the influence of background sound on speech transcription; determine the word correction result and character correction result of the initial speech text, and integrate and correct the initial speech text based on the word correction result and character correction result to obtain the transcription result of the original speech signal. In this way, the accuracy of the transcription result can be improved and the quality of speech transcription can be improved.

[0005] The present application provides a method for speech transcription, the method comprising:

[0006] Separating an original human voice signal from an original speech signal;

[0007] Performing voice endpoint detection on the original human voice signal to split the original human voice signal into at least one segment of human voice signal according to the detected voice endpoints;

[0008] Perform speech recognition on each human voice signal to obtain an initial speech text corresponding to each human voice signal;

[0009] Performing punctuation prediction on each segment of initial speech text, and adding punctuation to each segment of initial speech text, so as to segment each segment of initial speech text into at least one initial sentence;

[0010] For each initial sentence, determining a word correction result for an entity word in the initial sentence and determining a character correction result for a single character in the initial sentence;

[0011] According to the word error correction result and the character error correction result of each initial sentence, the initial sentence is integrated and corrected to obtain the transcription result of the original speech signal.

[0012] Furthermore, the separating the original human voice signal from the original speech signal comprises:

[0013] Acquire the frequency spectrum of the original speech signal;

[0014] Inputting the spectrum into a pre-trained background sound prediction model to obtain a background sound spectrum predicted by the background sound prediction model;

[0015] Inputting the spectrum into a pre-trained human voice prediction model to obtain a human voice spectrum predicted by the human voice prediction model;

[0016] Determine a background sound energy sequence according to the background sound spectrum; wherein the background sound energy sequence includes a plurality of background sound energy values, and the plurality of background sound energy values ​​are energy values ​​at different frequencies in a frequency band of the background sound spectrum;

[0017] Determine a human voice energy sequence according to the human voice spectrum; wherein the human voice energy sequence includes a plurality of human voice energy values, and the plurality of human voice energy values ​​are energy values ​​at different frequencies in the frequency band of the human voice spectrum;

[0018] Determine, according to the background sound energy sequence and the human voice energy sequence, a sequence of energy proportions of the human voice signal in the original speech signal;

[0019] Obtaining the spectrum of the human voice signal according to the energy proportion sequence of the human voice signal in the original voice signal and the spectrum of the original voice signal;

[0020] Performing an inverse short-time Fourier transform on the frequency spectrum of the human voice signal to obtain the human voice signal.

[0021] Furthermore, for each initial sentence, determining the word correction result of the entity words in the initial sentence includes:

[0022] Perform entity recognition on the initial sentence to obtain entity words included in the initial sentence;

[0023] For each entity word, calculating the minimum edit distance between the pinyin of the entity word and the pinyin of the central word of each vocabulary category in the pre-constructed standard vocabulary;

[0024] The vocabulary category to which the central word with the smallest minimum edit distance belongs is determined as the target category;

[0025] Calculating the minimum edit distance between the pinyin of the entity word and the pinyin of each standard word in the target category of the vocabulary; wherein the standard words in the target category include the central word of the target category;

[0026] If the minimum edit distance between the pinyin of the entity word and the pinyin of any standard word in the target category is zero, the standard word is used as the error correction result corresponding to the entity word;

[0027] If the minimum edit distance between the pinyin of the entity word and the pinyin of each word in the target category is not zero, the entity word is added to the target category of the vocabulary as a standard word.

[0028] Furthermore, for each initial sentence, determining the word error correction result for a single word in the initial sentence includes:

[0029] Starting from the first word of the initial sentence, each word in the initial sentence is taken as a target word in turn, and a word correction process is performed on the target word;

[0030] The step of performing word correction processing on the target word includes:

[0031] Obtain the mask representation of the target word and the feature vector representation of other words in the initial sentence;

[0032] Predicting multiple optional words corresponding to the position of the target word according to the feature vector representations of other words located before the target word, the feature vector representations of other words located after the target word, and the mask representation of the target word in the initial sentence;

[0033] If the target word is not included in the multiple optional words corresponding to the target word position, the optional word with the highest probability of appearing in the target word position among the multiple optional words is used as the word error correction result corresponding to the target word, and the target word is replaced by the word error correction result corresponding to the target word;

[0034] If the target word is included in the multiple optional words corresponding to the target word position, the target word is not replaced.

[0035] Furthermore, the initial sentence is integrated and corrected according to the word correction result and the character correction result of each initial sentence to obtain the transcription result of the original speech signal, including:

[0036] For each entity word in each initial sentence, determine whether the entity word has a corresponding word correction result;

[0037] If the entity word has a corresponding word correction result, the entity word is replaced with the word correction result corresponding to the entity word;

[0038] If the entity word has no corresponding word error correction result, determine whether each character in the entity word has a corresponding word error correction result;

[0039] If any character in the entity word has a corresponding character correction result, the character is replaced with the corresponding character correction result;

[0040] For each character except entity words in each initial sentence, determine whether the character has a corresponding character error correction result;

[0041] If the character has a corresponding word correction result, the word correction result corresponding to the character is used to replace the character;

[0042] Each initial sentence after entity word replacement and character replacement is determined as a transcription result of the original speech signal.

[0043] Furthermore, the background sound prediction model and the human voice prediction model are trained in the following manner:

[0044] Acquire multiple sections of sample speech signals and a sample background sound signal and a sample human voice signal included in each section of the sample speech signal;

[0045] For each sample speech signal, respectively obtain a true frequency spectrum of the sample speech signal, a true background sound frequency spectrum of the sample background sound signal included in the sample speech signal, and a true human voice frequency spectrum of the sample human voice signal included in the sample speech signal;

[0046] Inputting the real frequency spectrum of the sample speech signal into a pre-built initial background sound prediction model and a pre-built initial human voice prediction model respectively, to obtain a predicted background sound frequency spectrum output by the initial background sound prediction model and a predicted human voice frequency spectrum output by the initial human voice prediction model;

[0047] Determining a first loss function value according to a real background audio spectrum and a predicted background audio spectrum of the sample background audio signal;

[0048] Determining a second loss function value according to a real human voice spectrum and a predicted human voice spectrum of the sample human voice signal;

[0049] The initial background sound prediction model and the initial human voice prediction model are trained respectively according to the average value of the first loss function value and the second loss function value to obtain the background sound prediction model and the human voice prediction model.

[0050] Furthermore, the standard vocabulary is constructed in the following way:

[0051] Get multiple standard words;

[0052] randomly dividing the plurality of standard words into a plurality of vocabulary categories;

[0053] Selecting standard words from the plurality of standard words in turn as the words to be classified, calculating the minimum edit distance between the words to be classified and each vocabulary category, and classifying the words to be classified into the vocabulary category with the smallest minimum edit distance; wherein, for each vocabulary category, the minimum edit distance between the words to be classified and the vocabulary category refers to the average value of the minimum edit distances between the pinyin of the words to be classified and the pinyin of each standard word in the vocabulary category;

[0054] Repeat the previous step until the division result of each standard word in the plurality of standard words does not change, thereby obtaining a plurality of vocabulary categories divided by the plurality of standard words;

[0055] For each vocabulary category of the multiple standard words, the minimum edit distance between the pinyin of each standard word in the vocabulary category and the pinyin of other standard words in the vocabulary category is calculated respectively, and the standard word with the smallest average value of the minimum edit distance to the pinyin of other standard words in the vocabulary category is taken as the central word of the vocabulary category.

[0056] The present application also provides a speech transcription device, the device comprising:

[0057] A separation module, used for separating an original human voice signal from an original speech signal;

[0058] A detection module, configured to perform voice endpoint detection on the original human voice signal, so as to split the original human voice signal into at least one segment of human voice signal according to the detected voice endpoints;

[0059] A recognition module is used to perform speech recognition on each human voice signal to obtain an initial speech text corresponding to each human voice signal;

[0060] A segmentation module, used for predicting punctuation marks for each segment of initial speech text, adding punctuation marks to each segment of initial speech text, so as to segment each segment of initial speech text into at least one initial sentence;

[0061] An error correction module, for determining, for each initial sentence, word error correction results for entity words in the initial sentence and word error correction results for individual words in the initial sentence;

[0062] The integration module is used to integrate and correct each initial sentence according to the word correction result and the character correction result of the initial sentence to obtain the transcription result of the original speech signal.

[0063] An embodiment of the present application also provides an electronic device, including: a processor, a memory and a bus, wherein the memory stores machine-readable instructions executable by the processor, and when the electronic device is running, the processor and the memory communicate through the bus, and when the machine-readable instructions are executed by the processor, the steps of a speech transcription method as described above are performed.

[0064] An embodiment of the present application also provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the steps of the speech transcription method as described above are executed.

[0065] The embodiments of the present application provide a speech transcription method, device, electronic device and storage medium, including: separating an original human voice signal from an original speech signal; performing speech endpoint detection on the original human voice signal to split the original human voice signal into at least one segment of human voice signal according to the detected speech endpoint; performing speech recognition on each segment of human voice signal to obtain an initial speech text corresponding to each segment of human voice signal; performing punctuation prediction on each segment of initial speech text, adding punctuation to each segment of initial speech text, so as to split each segment of initial speech text into at least one initial sentence; for each initial sentence, determining a word correction result for an entity word in the initial sentence and determining a character correction result for a single character in the initial sentence; and integrating and correcting the initial sentence according to the word correction result and the character correction result of each initial sentence to obtain a transcription result of the original speech signal.

[0066] By separating the original human voice signal from the original speech signal, the influence of background sound on speech transcription is eliminated; by determining the word correction result and character correction result of the initial speech text, and integrating and correcting the initial speech text based on the word correction result and character correction result, the transcription result of the original speech signal is obtained. In this way, the accuracy of the transcription result can be improved, and the quality of speech transcription can be improved.

[0067] In order to make the above-mentioned objects, features and advantages of the present application more obvious and easy to understand, preferred embodiments are specifically cited below and described in detail with reference to the attached drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings required for use in the embodiments will be briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present application and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without paying creative work.

[0069] Figure 1 A flow chart of a speech transcription method provided in an embodiment of the present application is shown;

[0070] Figure 2 One of the structural schematic diagrams of a speech transcription device provided in an embodiment of the present application is shown;

[0071] Figure 3 A second structural schematic diagram of a speech transcription device provided in an embodiment of the present application is shown;

[0072] Figure 4 A schematic structural diagram of an electronic device provided in an embodiment of the present application is shown. DETAILED DESCRIPTION

[0073] To make the purpose, technical scheme and advantages of the embodiments of the present application clearer, the technical scheme in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all of the embodiments. The components of the embodiments of the present application usually described and shown in the drawings here can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present application provided in the drawings is not intended to limit the scope of the application claimed for protection, but merely represents the selected embodiments of the present application. Based on the embodiments of the present application, each other embodiment obtained by those skilled in the art without making creative work belongs to the scope of protection of the present application.

[0074] With the development of speech technology, speech transcription technology from speech to corresponding text has gradually become popular in daily life. Speech transcription technology mainly performs speech-to-text processing on longer speech to obtain the text content of the speech. At present, speech transcription technology has been widely used in various application scenarios including content recognition of voice recorders and subtitle generation.

[0075] In the existing technology, traditional acoustic models or deep learning models are often used to recognize the text content of speech. However, due to the complex application scenarios of speech transcription, limited accuracy of the model, insufficient generalization ability and other reasons, the quality of speech transcription is currently poor and the accuracy of transcription is low.

[0076] Based on this, the embodiments of the present application provide a speech transcription method, device, electronic device and storage medium to improve the accuracy of transcription results and improve the quality of speech transcription.

[0077] See also Figure 1 , Figure 1 A flow chart of a speech transcription method provided in an embodiment of the present application. Figure 1 As shown in, the method provided in the embodiment of the present application includes:

[0078] S101, separating an original human voice signal from an original speech signal.

[0079] Here, the original speech signal includes the original human voice signal and the background sound signal. The speech transcription technology needs to transcribe the speech text embodied by the original human voice signal from the original speech signal. During the transcription process, the background sound signal will affect the speech transcription. Taking the application scenario of generating subtitles for film and television dramas as an example, the speech of film and television dramas may include the human voice and background sound effects emitted by the characters; taking the application scenario of recording content recognition of a voice recorder as an example, the audio recorded by the voice recorder at the meeting site includes not only the speeches of the participants, but also the noise at the meeting site.

[0080] In a possible implementation, step S101 may include:

[0081] S1011. Acquire the frequency spectrum of the original speech signal.

[0082] In a specific implementation, the spectrum of the original speech signal can be obtained through Fourier transform.

[0083] S1012: Input the frequency spectrum into a pre-trained background sound prediction model to obtain a background sound frequency spectrum predicted by the background sound prediction model.

[0084] S1013: Input the spectrum into a pre-trained human voice prediction model to obtain a human voice spectrum predicted by the human voice prediction model.

[0085] In steps S1012 and S1013, the spectrum of the original speech signal may be input into the background sound prediction model and the human voice prediction model respectively to obtain the background sound spectrum output by the background sound prediction model and the human voice spectrum output by the human voice prediction model.

[0086] S1014: Determine a background sound energy sequence according to the background sound spectrum.

[0087] The background sound energy sequence includes multiple background sound energy values, which are energy values ​​at different frequencies in the frequency band of the background sound spectrum. Here, the spectrum can be expressed as a function f(t) with frequency as the independent variable; the frequency band is the bandwidth, which refers to the width of the frequency band occupied by the signal; the energy value is the square of the modulus of the instantaneous power at a certain frequency in the frequency band. Specifically, the following formula can be used to determine a certain frequency t in the frequency band of the background sound spectrum: i The energy value at (i is a positive integer):

[0088]

[0089] In the formula, a_engry i Represents a frequency t in the frequency band of the background sound spectrum i The energy value at f a (ti ) represents the frequency t i The background audio spectrum when

[0090] Thus, we get the background sound energy sequence a={a_engry 0 , ..., a_engry i , ..., a_engry n}.

[0091] S1015: Determine a human voice energy sequence according to the human voice spectrum.

[0092] The human voice energy sequence includes a plurality of human voice energy values, and the plurality of human voice energy values ​​are energy values ​​at different frequencies in the frequency band of the human voice spectrum.

[0093] In this step, a certain frequency t in the frequency band of the human voice spectrum can be determined by the following formula: i The energy value at (i is a natural number):

[0094]

[0095] In the formula, v_engry i Represents a frequency t in the frequency band of the human voice spectrum i The energy value at f v (t i ) represents the frequency t i The human voice spectrum.

[0096] Thus, we get the vocal energy sequence v = {v_engry 0 , ..., v_engry i , ..., v_engry n}.

[0097] S1016. Determine a sequence of energy proportions of the human voice signal in the original speech signal according to the background sound energy sequence and the human voice energy sequence.

[0098] The energy ratio sequence includes energy ratios corresponding to multiple frequencies in the frequency band. Specifically, the energy ratio of the human voice signal at a certain frequency t in the frequency band of the original speech signal can be determined by the following formula: i The corresponding energy ratio is:

[0099]

[0100] In the formula, Indicates the frequency t of the human voice signal in the frequency band of the original speech signal i The energy proportion at.

[0101] Similarly, for each frequency band, the energy proportion of the background sound signal at a certain frequency in the frequency band of the original speech signal can be determined by the following formula:

[0102]

[0103] In the formula, Indicates the frequency t of the background sound signal in the frequency band of the original speech signal i The energy proportion at.

[0104] Thus, the energy ratio sequence of the human voice signal in the original speech signal can be obtained:

[0105] S1017: Obtain a spectrum of the human voice signal according to a sequence of energy proportions of the human voice signal in the original voice signal and a spectrum of the original voice signal.

[0106] In this step, the spectrum of the original speech signal may be multiplied by the energy proportion sequence of the human voice signal in the original speech signal to obtain the actual spectrum of the human voice signal.

[0107] For example, the human voice signal at frequency t i The actual spectrum corresponding to can be expressed as:

[0108]

[0109] In the formula, f v_s (t i ) represents the human voice signal at frequency t i The actual spectrum at t, that is, the human voice signal at frequency t i The corresponding spectrum at f(t i ) represents the original speech signal at frequency t i The corresponding spectrum.

[0110] S1018. Perform an inverse short-time Fourier transform on the frequency spectrum of the human voice signal to obtain the human voice signal.

[0111] In a possible implementation manner, the background sound prediction model and the human voice prediction model are trained in the following manner:

[0112] Step 1: Acquire multiple sample speech signals and a sample background sound signal and a sample human voice signal included in each sample speech signal.

[0113] Step 2: For each sample speech signal, respectively obtain the real spectrum of the sample speech signal, the real background sound spectrum of the sample background sound signal included in the sample speech signal, and the real human voice spectrum of the sample human voice signal included in the sample speech signal.

[0114] Step 3: The real frequency spectrum of the sample speech signal is respectively input into a pre-built initial background sound prediction model and a pre-built initial human voice prediction model to obtain a predicted background sound frequency spectrum output by the initial background sound prediction model and a predicted human voice frequency spectrum output by the initial human voice prediction model.

[0115] Step 4: Determine a first loss function value according to the real background sound spectrum and the predicted background sound spectrum of the sample background sound signal.

[0116] Step 5: Determine the value of the second loss function according to the actual vocal spectrum and the predicted vocal spectrum of the sample vocal signal.

[0117] Step 6: According to the average value of the first loss function value and the second loss function value, the initial background sound prediction model and the initial human voice prediction model are trained respectively to obtain the background sound prediction model and the human voice prediction model.

[0118] In a specific implementation, the background sound prediction model and the human voice prediction model may be a Transformer model, etc.; the first loss function value and the second loss function value may be determined by an L1 norm loss function.

[0119] S102: Perform voice endpoint detection on the original human voice signal to split the original human voice signal into at least one segment of human voice signal according to the detected voice endpoints.

[0120] In this step, by performing speech endpoint detection on the original human voice signal, the speech endpoints in the speech signal, that is, at least one group of speech start points and speech end points, can be detected; through at least one group of speech start points and speech end points, the original human voice signal can be split into at least one segment of human voice signal, thereby removing the silent part in the original human voice signal.

[0121] In specific implementation, the libfvad algorithm can be used to perform voice endpoint detection on the original human voice signal in wav format, find the silent part therein, and divide it into at least one shorter human voice signal according to the detected voice endpoint.

[0122] S103: Perform speech recognition on each segment of human voice signal to obtain an initial speech text corresponding to each segment of human voice signal.

[0123] In this step, 10,000+ hours of Chinese speech data can be used in advance to train a speech recognition model with high recognition accuracy; then the speech recognition model is used to perform speech recognition on each human voice signal obtained after speech endpoint detection to obtain an initial speech text corresponding to each human voice signal.

[0124] S104: perform punctuation prediction on each segment of initial speech text, and add punctuation to each segment of initial speech text, so as to segment each segment of initial speech text into at least one initial sentence.

[0125] Here, by adding punctuation to each segment of initial voice text, each segment of initial voice text can be segmented into at least one initial sentence according to the added punctuation, and the text can also be made more in line with natural expression and more readable.

[0126] S105 . For each initial sentence, determine a word correction result for the entity words in the initial sentence and determine a character correction result for the single character in the initial sentence.

[0127] In a possible implementation, for each initial sentence, determining a word error correction result for an entity word in the initial sentence may include:

[0128] S1051. Perform entity recognition on the initial sentence to obtain entity words included in the initial sentence.

[0129] Among them, entity words are words with specific meanings in the text, including names of people, places, institutions, proper nouns, etc. By performing entity recognition on each initial sentence, the entity words included in the initial sentence can be obtained. It should be noted that some initial sentences may not include entity words, and some initial sentences may include one or more entity words.

[0130] S1052: For each entity word, calculate the minimum edit distance between the pinyin of the entity word and the pinyin of the central word of each vocabulary category in the pre-constructed standard vocabulary.

[0131] The pre-constructed standard vocabulary table divides a plurality of standard words into a plurality of vocabulary categories, and selects a standard word from a plurality of standard words corresponding to each vocabulary category as the central word of the vocabulary category.

[0132] In this step, for each entity word, the minimum edit distance between the entity word and the central word of each vocabulary category in the standard vocabulary based on pinyin can be determined.

[0133] Here, the minimum edit distance refers to the minimum number of single-character edit operations required to convert one string into the other string between two strings.

[0134] S1053: Determine the vocabulary category to which the central word with the smallest minimum edit distance belongs as the target category.

[0135] S1054. Calculate the minimum edit distance between the pinyin of the entity word and the pinyin of each standard word in the target category of the vocabulary; wherein the standard words in the target category include the central word of the target category.

[0136] In this step, the minimum edit distance between the entity word and each standard word in the target category based on pinyin is calculated. Here, the standard words in the target category also include the central word of the category.

[0137] S1055. If the minimum edit distance between the pinyin of the entity word and the pinyin of any standard word in the target category is zero, the standard word is used as the error correction result corresponding to the entity word.

[0138] S1056: If the minimum edit distance between the pinyin of the entity word and the pinyin of each word in the target category is not zero, then add the entity word as a standard word to the target category of the vocabulary.

[0139] Among them, if the minimum edit distance between the pinyin of the entity word and the pinyin of any standard word in the target category is zero, it means that the entity word needs to be corrected, and the standard word with the minimum edit distance of zero is used as the correction result corresponding to the entity word; if the minimum edit distance between the pinyin of the entity word and the pinyin of each word in the target category is not zero, it means that the entity word does not need to be corrected, and the entity word is added to the target category of the vocabulary as a standard word to realize continuous updating and expansion of the vocabulary.

[0140] Here, the standard vocabulary can be constructed in the following way:

[0141] Step 1: Get multiple standard words.

[0142] Step 2: Randomly divide the plurality of standard words into a plurality of vocabulary categories.

[0143] Step 3: Select standard words from the multiple standard words in turn as the words to be classified, calculate the minimum edit distance between the words to be classified and each vocabulary category, and classify the words to be classified into the vocabulary category with the smallest minimum edit distance.

[0144] For each vocabulary category, the minimum edit distance between the word to be classified and the vocabulary category refers to the average value of the minimum edit distances between the pinyin of the word to be classified and the pinyin of each standard word in the vocabulary category.

[0145] Step 3 is repeatedly performed until the division result of each standard word in the plurality of standard words no longer changes, thereby obtaining a plurality of vocabulary categories into which the plurality of standard words are divided.

[0146] Step 4: For each vocabulary category of the multiple standard words, calculate the minimum edit distance between the pinyin of each standard word in the vocabulary category and the pinyin of other standard words in the vocabulary category, and take the standard word with the smallest average value of the minimum edit distance to the pinyin of other standard words in the vocabulary category as the central word of the vocabulary category.

[0147] In another possible implementation, for each initial sentence, a word correction result for a single word in the initial sentence is determined, including: starting from the first word of the initial sentence, taking each word in the initial sentence as a target word in turn, and performing word correction processing on the target word.

[0148] The word correction processing for the target word may include:

[0149] Step 1: Get the mask representation of the target word and the feature vector representation of other words in the initial sentence.

[0150] Step 2: Predict multiple optional words corresponding to the position of the target word based on the feature vector representations of other words located before the target word, the feature vector representations of other words located after the target word, and the mask representation of the target word in the initial sentence.

[0151] Step 3: If the target word is not included in the multiple optional words corresponding to the target word position, the optional word with the highest probability of appearing in the target word position among the multiple optional words is used as the word correction result corresponding to the target word, and the target word is replaced using the word correction result corresponding to the target word.

[0152] Step 4: If the target word is included in the multiple optional words corresponding to the target word position, the target word is not replaced.

[0153] In specific implementation, the probability distribution of optional words corresponding to the target word position can be obtained by using MLM (Mask Language Model) through cloze test. Specifically, first, each word in the initial sentence is input into the encoder in turn, and the encoder vectorizes the input to obtain a mask representation of the target word of the initial sentence and a feature vector representation of other words in the initial sentence. For example, the encoder can perform a mask representation on the target word; secondly, the mask representation of the target word of the initial sentence and the feature vector representation of other words in the initial sentence are input into the decoder, and the decoder predicts multiple optional words corresponding to the target word position according to the target word context (that is, the feature vector representation of other words before the target word and the feature vector representation of other words after the target word in the initial sentence); thirdly, it is determined whether the target word is included in the multiple optional words corresponding to the target word position. If the target word is included in the multiple optional words, it is determined that the target word does not need to be corrected, that is, the target word is not replaced; if the target word is not included in the multiple optional words, it is determined that the target word needs to be corrected, and the optional word with the highest probability of appearing at the target word position among the multiple optional words is used as the word correction result corresponding to the target word, and the word correction result corresponding to the target word is used to replace the target word.

[0154] After the word correction process for the current target word is completed, the next word of the current target word in the initial sentence is taken as the target word again, and the above word correction process is performed on the target word. It is worth noting that if the current target word is replaced, when the word correction process is performed on the next target word of the current target word, the feature vector representation obtained again is the feature vector representation of the corresponding word after the current target word is replaced.

[0155] S106: According to the word error correction result and the character error correction result of each initial sentence, the initial sentence is integrated and corrected to obtain the transcription result of the original speech signal.

[0156] In this step, the word correction result and the character correction result of each initial sentence may overlap, so the word correction result and the character correction result of each initial sentence need to be integrated.

[0157] In a possible implementation, step S106 may include:

[0158] S1061. For each entity word in each initial sentence, determine whether the entity word has a corresponding word correction result.

[0159] S1062. If the entity word has a corresponding word correction result, the entity word is replaced with the word correction result corresponding to the entity word.

[0160] S1063. If the entity word has no corresponding word correction result, determine whether each character in the entity word has a corresponding word correction result.

[0161] S1064. If any character in the entity word has a corresponding character correction result, the character is replaced with the corresponding character correction result.

[0162] S1065. For each character except the entity words in each initial sentence, determine whether the character has a corresponding character error correction result.

[0163] S1066. If the character has a corresponding character error correction result, the character is replaced with the corresponding character error correction result.

[0164] Here, when an entity word has a corresponding word correction result, the word correction result is used to replace the entity word first; when an entity word does not have a corresponding word correction result, but one or more characters in the entity word have corresponding character correction results, the corresponding character correction results are used to replace the corresponding characters; for each character in the sentence except the entity word, the corresponding character correction result is used to replace the character.

[0165] S1067: Determine each initial sentence after entity word replacement and character replacement as a transcription result of the original speech signal.

[0166] In this way, the word correction result in the speech transcription method provided in the embodiment of the present application performs text correction from the perspective of traditional phonetics, focusing on text correction of homophone errors of entity words; while the character correction result takes into account the contextual relationship of each word from the perspective of deep learning speech features, focusing on text correction of non-entity words. By combining the two correction results to obtain the final speech transcription result, the accuracy and quality of speech transcription can be improved.

[0167] A speech transcription method provided by an embodiment of the present application includes: separating an original human voice signal from an original speech signal; performing speech endpoint detection on the original human voice signal to split the original human voice signal into at least one segment of human voice signal according to the detected speech endpoint; performing speech recognition on each segment of human voice signal to obtain a segment of initial speech text corresponding to each segment of human voice signal; performing punctuation prediction on each segment of initial speech text, adding punctuation to each segment of initial speech text, to split each segment of initial speech text into at least one initial sentence; for each initial sentence, determining a word correction result for an entity word in the initial sentence and determining a character correction result for a single character in the initial sentence; and integrating and correcting the initial sentence according to the word correction result and the character correction result of each initial sentence to obtain a transcription result of the original speech signal.

[0168] By separating the original human voice signal from the original speech signal, the influence of background sound on speech transcription is eliminated; by determining the word correction result and character correction result of the initial speech text, and integrating and correcting the initial speech text based on the word correction result and character correction result, the transcription result of the original speech signal is obtained. In this way, the accuracy of the transcription result can be improved, and the quality of speech transcription can be improved.

[0169] See also Figure 2 , Figure 3 , Figure 2 This is one of the structural schematic diagrams of a speech transcription device provided in an embodiment of the present application. Figure 3 This is a second structural diagram of a speech transcription device provided in an embodiment of the present application. Figure 2 As shown in , the speech transcription device 200 comprises:

[0170] The separation module 210 is used to separate the original human voice signal from the original speech signal;

[0171] The detection module 220 is used to perform voice endpoint detection on the original human voice signal to split the original human voice signal into at least one segment of human voice signal according to the detected voice endpoints;

[0172] The recognition module 230 is used to perform speech recognition on each segment of human voice signal to obtain an initial speech text corresponding to each segment of human voice signal;

[0173] The segmentation module 240 is used to predict punctuation marks for each segment of initial speech text and add punctuation marks to each segment of initial speech text so as to segment each segment of initial speech text into at least one initial sentence;

[0174] The error correction module 250 is used to determine, for each initial sentence, a word error correction result for an entity word in the initial sentence and a character error correction result for a single character in the initial sentence;

[0175] The integration module 260 is used to integrate and correct each initial sentence according to the word correction result and character correction result of the initial sentence to obtain the transcription result of the original speech signal.

[0176] Furthermore, when the separation module 210 is used to separate the original human voice signal from the original speech signal, the separation module 210 is used to:

[0177] Acquire the frequency spectrum of the original speech signal;

[0178] Inputting the spectrum into a pre-trained background sound prediction model to obtain a background sound spectrum predicted by the background sound prediction model;

[0179] Inputting the spectrum into a pre-trained human voice prediction model to obtain a human voice spectrum predicted by the human voice prediction model;

[0180] Determine a background sound energy sequence according to the background sound spectrum; wherein the background sound energy sequence includes a plurality of background sound energy values, and the plurality of background sound energy values ​​are energy values ​​at different frequencies in a frequency band of the background sound spectrum;

[0181] Determine a human voice energy sequence according to the human voice spectrum; wherein the human voice energy sequence includes a plurality of human voice energy values, and the plurality of human voice energy values ​​are energy values ​​at different frequencies in the frequency band of the human voice spectrum;

[0182] Determine, according to the background sound energy sequence and the human voice energy sequence, a sequence of energy proportions of the human voice signal in the original speech signal;

[0183] Obtaining the spectrum of the human voice signal according to the energy proportion sequence of the human voice signal in the original voice signal and the spectrum of the original voice signal;

[0184] Performing an inverse short-time Fourier transform on the frequency spectrum of the human voice signal to obtain the human voice signal.

[0185] Furthermore, when the error correction module 250 is used to determine the word error correction result of the entity word in each initial sentence, the error correction module 250 is used to:

[0186] Perform entity recognition on the initial sentence to obtain entity words included in the initial sentence;

[0187] For each entity word, calculating the minimum edit distance between the pinyin of the entity word and the pinyin of the central word of each vocabulary category in the pre-constructed standard vocabulary;

[0188] The vocabulary category to which the central word with the smallest minimum edit distance belongs is determined as the target category;

[0189] Calculating the minimum edit distance between the pinyin of the entity word and the pinyin of each standard word in the target category of the vocabulary; wherein the standard words in the target category include the central word of the target category;

[0190] If the minimum edit distance between the pinyin of the entity word and the pinyin of any standard word in the target category is zero, the standard word is used as the error correction result corresponding to the entity word;

[0191] If the minimum edit distance between the pinyin of the entity word and the pinyin of each word in the target category is not zero, the entity word is added to the target category of the vocabulary as a standard word.

[0192] Furthermore, when the error correction module 250 is used to determine the word error correction result of a single word in each initial sentence, the error correction module 250 is used to:

[0193] Starting from the first word of the initial sentence, each word in the initial sentence is taken as a target word in turn, and a word correction process is performed on the target word;

[0194] The step of performing word correction processing on the target word includes:

[0195] Obtain the mask representation of the target word and the feature vector representation of other words in the initial sentence;

[0196] Predicting multiple optional words corresponding to the position of the target word according to the feature vector representations of other words located before the target word, the feature vector representations of other words located after the target word, and the mask representation of the target word in the initial sentence;

[0197] If the target word is not included in the multiple optional words corresponding to the target word position, the optional word with the highest probability of appearing in the target word position among the multiple optional words is used as the word error correction result corresponding to the target word, and the target word is replaced by the word error correction result corresponding to the target word;

[0198] If the target word is included in the multiple optional words corresponding to the target word position, the target word is not replaced.

[0199] Furthermore, when the integration module 260 is used to integrate and correct the initial sentence according to the word correction result and the character correction result of each initial sentence to obtain the transcription result of the original speech signal, the integration module 260 is used to:

[0200] For each entity word in each initial sentence, determine whether the entity word has a corresponding word correction result;

[0201] If the entity word has a corresponding word correction result, the entity word is replaced with the word correction result corresponding to the entity word;

[0202] If the entity word has no corresponding word error correction result, determine whether each character in the entity word has a corresponding word error correction result;

[0203] If any character in the entity word has a corresponding character correction result, the character is replaced with the corresponding character correction result;

[0204] For each character except entity words in each initial sentence, determine whether the character has a corresponding character error correction result;

[0205] If the character has a corresponding word correction result, the word correction result corresponding to the character is used to replace the character;

[0206] Each initial sentence after entity word replacement and character replacement is determined as a transcription result of the original speech signal.

[0207] Further, such as Figure 3 As shown, the speech transcription device 200 further includes: a training module 270; the training module 270 is used to train the background sound prediction model and the human voice prediction model in the following manner:

[0208] Acquire multiple sections of sample speech signals and a sample background sound signal and a sample human voice signal included in each section of the sample speech signal;

[0209] For each sample speech signal, respectively obtain a true frequency spectrum of the sample speech signal, a true background sound frequency spectrum of the sample background sound signal included in the sample speech signal, and a true human voice frequency spectrum of the sample human voice signal included in the sample speech signal;

[0210] Inputting the real frequency spectrum of the sample speech signal into a pre-built initial background sound prediction model and a pre-built initial human voice prediction model respectively, to obtain a predicted background sound frequency spectrum output by the initial background sound prediction model and a predicted human voice frequency spectrum output by the initial human voice prediction model;

[0211] Determining a first loss function value according to a real background audio spectrum and a predicted background audio spectrum of the sample background audio signal;

[0212] Determining a second loss function value according to a real human voice spectrum and a predicted human voice spectrum of the sample human voice signal;

[0213] The initial background sound prediction model and the initial human voice prediction model are trained respectively according to the average value of the first loss function value and the second loss function value to obtain the background sound prediction model and the human voice prediction model.

[0214] Further, such as Figure 3 As shown, the speech transcription device 200 further includes: a construction module 280; the construction module 280 is used to construct the standard vocabulary in the following manner:

[0215] Get multiple standard words;

[0216] randomly dividing the plurality of standard words into a plurality of vocabulary categories;

[0217] Selecting standard words from the plurality of standard words in turn as the words to be classified, calculating the minimum edit distance between the words to be classified and each vocabulary category, and classifying the words to be classified into the vocabulary category with the smallest minimum edit distance; wherein, for each vocabulary category, the minimum edit distance between the words to be classified and the vocabulary category refers to the average value of the minimum edit distances between the pinyin of the words to be classified and the pinyin of each standard word in the vocabulary category;

[0218] Repeat the previous step until the division result of each standard word in the plurality of standard words does not change, thereby obtaining a plurality of vocabulary categories divided by the plurality of standard words;

[0219] For each vocabulary category of the multiple standard words, the minimum edit distance between the pinyin of each standard word in the vocabulary category and the pinyin of other standard words in the vocabulary category is calculated respectively, and the standard word with the smallest average value of the minimum edit distance to the pinyin of other standard words in the vocabulary category is taken as the central word of the vocabulary category.

[0220] A speech transcription device provided by an embodiment of the present application includes: separating an original human voice signal from an original speech signal; performing speech endpoint detection on the original human voice signal to split the original human voice signal into at least one segment of human voice signal according to the detected speech endpoint; performing speech recognition on each segment of human voice signal to obtain a segment of initial speech text corresponding to each segment of human voice signal; performing punctuation prediction on each segment of initial speech text, adding punctuation to each segment of initial speech text, so as to split each segment of initial speech text into at least one initial sentence; for each initial sentence, determining a word correction result for an entity word in the initial sentence and determining a character correction result for a single character in the initial sentence; and integrating and correcting the initial sentence according to the word correction result and the character correction result of each initial sentence to obtain a transcription result of the original speech signal.

[0221] By separating the original human voice signal from the original speech signal, the influence of background sound on speech transcription is eliminated; by determining the word correction result and character correction result of the initial speech text, and integrating and correcting the initial speech text based on the word correction result and character correction result, the transcription result of the original speech signal is obtained. In this way, the accuracy of the transcription result can be improved, and the quality of speech transcription can be improved.

[0222] See also Figure 4 , Figure 4 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present application. Figure 4 As shown in , the electronic device 400 includes a processor 410 , a memory 420 and a bus 430 .

[0223] The memory 420 stores machine-readable instructions executable by the processor 410. When the electronic device 400 is running, the processor 410 communicates with the memory 420 via the bus 430. When the machine-readable instructions are executed by the processor 410, the above-mentioned Figure 1 The steps of a speech transcription method in the method embodiment shown are specifically implemented in accordance with the method embodiment and will not be described in detail here.

[0224] The present application also provides a computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, the computer program can execute the above-mentioned Figure 1 The steps of a speech transcription method in the method embodiment shown are specifically implemented in accordance with the method embodiment and will not be described in detail here.

[0225] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0226] In the several embodiments provided in the present application, it should be understood that the disclosed systems, devices and methods can be implemented in other ways. The device embodiments described above are merely schematic. For example, the division of the units is only a logical function division. There may be other division methods in actual implementation. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some communication interfaces, and the indirect coupling or communication connection of devices or units can be electrical, mechanical or other forms.

[0227] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0228] In addition, each functional unit in each embodiment of the present application may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit.

[0229] If the functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a non-volatile computer-readable storage medium that is executable by a processor. Based on this understanding, the technical solution of the present application can essentially be embodied in the form of a software product, or in other words, the part that contributes to the prior art or the part of the technical solution. The computer software product is stored in a storage medium, including several instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk.

[0230] Finally, it should be noted that the above-described embodiments are only specific implementation methods of the present application, which are used to illustrate the technical solutions of the present application, rather than to limit them. The protection scope of the present application is not limited thereto. Although the present application is described in detail with reference to the above-mentioned embodiments, ordinary technicians in the field should understand that any technician familiar with the technical field can still modify the technical solutions recorded in the above-mentioned embodiments within the technical scope disclosed in the present application, or can easily think of changes, or make equivalent replacements for some of the technical features therein; and these modifications, changes or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application, and should be included in the protection scope of the present application. Therefore, the protection scope of the present application shall be based on the protection scope of the claims.

Claims

1. A speech transcription method, characterized in that: The method comprises: Separating an original human voice signal from an original speech signal; Performing voice endpoint detection on the original human voice signal to split the original human voice signal into at least one segment of human voice signal according to the detected voice endpoints; Perform speech recognition on each human voice signal to obtain an initial speech text corresponding to each human voice signal; Performing punctuation prediction on each segment of initial speech text, and adding punctuation to each segment of initial speech text, so as to segment each segment of initial speech text into at least one initial sentence; For each initial sentence, determining a word correction result for an entity word in the initial sentence and determining a character correction result for a single character in the initial sentence; According to the word error correction result and the character error correction result of each initial sentence, the initial sentence is integrated and corrected to obtain the transcription result of the original speech signal; The step of separating the original human voice signal from the original speech signal comprises: Acquire the frequency spectrum of the original speech signal; Inputting the spectrum into a pre-trained background sound prediction model to obtain a background sound spectrum predicted by the background sound prediction model; Inputting the spectrum into a pre-trained human voice prediction model to obtain a human voice spectrum predicted by the human voice prediction model; Determine a background sound energy sequence according to the background sound spectrum; wherein the background sound energy sequence includes a plurality of background sound energy values, and the plurality of background sound energy values ​​are energy values ​​at different frequencies in a frequency band of the background sound spectrum; Determine a human voice energy sequence according to the human voice spectrum; wherein the human voice energy sequence includes a plurality of human voice energy values, and the plurality of human voice energy values ​​are energy values ​​at different frequencies in the frequency band of the human voice spectrum; Determine, according to the background sound energy sequence and the human voice energy sequence, a sequence of energy proportions of the human voice signal in the original speech signal; Obtaining the spectrum of the human voice signal according to the energy proportion sequence of the human voice signal in the original voice signal and the spectrum of the original voice signal; Performing an inverse short-time Fourier transform on the frequency spectrum of the human voice signal to obtain the human voice signal.

2. The method according to claim 1, characterized in that The step of determining, for each initial sentence, a word correction result for an entity word in the initial sentence includes: Perform entity recognition on the initial sentence to obtain entity words included in the initial sentence; For each entity word, calculating the minimum edit distance between the pinyin of the entity word and the pinyin of the central word of each vocabulary category in the pre-constructed standard vocabulary; The vocabulary category to which the central word with the smallest minimum edit distance belongs is determined as the target category; Calculating the minimum edit distance between the pinyin of the entity word and the pinyin of each standard word in the target category of the vocabulary; wherein the standard words in the target category include the central word of the target category; If the minimum edit distance between the pinyin of the entity word and the pinyin of any standard word in the target category is zero, the standard word is used as the error correction result corresponding to the entity word; If the minimum edit distance between the pinyin of the entity word and the pinyin of each word in the target category is not zero, the entity word is added to the target category of the vocabulary as a standard word.

3. The method according to claim 1, characterized in that The step of determining, for each initial sentence, a word error correction result for a single word in the initial sentence comprises: Starting from the first word of the initial sentence, each word in the initial sentence is taken as a target word in turn, and a word correction process is performed on the target word; The step of performing word correction processing on the target word includes: Obtain the mask representation of the target word and the feature vector representation of other words in the initial sentence; Predicting multiple optional words corresponding to the position of the target word according to the feature vector representations of other words located before the target word, the feature vector representations of other words located after the target word, and the mask representation of the target word in the initial sentence; If the target word is not included in the multiple optional words corresponding to the target word position, the optional word with the highest probability of appearing in the target word position among the multiple optional words is used as the word error correction result corresponding to the target word, and the target word is replaced by the word error correction result corresponding to the target word; If the target word is included in the multiple optional words corresponding to the target word position, the target word is not replaced.

4. The method according to claim 1, characterized in that The step of integrating and correcting each initial sentence according to the word error correction result and the character error correction result of each initial sentence to obtain the transcription result of the original speech signal includes: For each entity word in each initial sentence, determine whether the entity word has a corresponding word correction result; If the entity word has a corresponding word correction result, the entity word is replaced with the word correction result corresponding to the entity word; If the entity word has no corresponding word error correction result, determine whether each character in the entity word has a corresponding word error correction result; If any character in the entity word has a corresponding character correction result, the character is replaced with the corresponding character correction result; For each character except entity words in each initial sentence, determine whether the character has a corresponding character error correction result; If the character has a corresponding word correction result, the word correction result corresponding to the character is used to replace the character; Each initial sentence after entity word replacement and character replacement is determined as a transcription result of the original speech signal.

5. The method according to claim 1, characterized in that The background sound prediction model and the human voice prediction model are trained in the following manner: Acquire multiple sections of sample speech signals and a sample background sound signal and a sample human voice signal included in each section of the sample speech signal; For each sample speech signal, respectively obtain a true frequency spectrum of the sample speech signal, a true background sound frequency spectrum of the sample background sound signal included in the sample speech signal, and a true human voice frequency spectrum of the sample human voice signal included in the sample speech signal; Inputting the real frequency spectrum of the sample speech signal into a pre-built initial background sound prediction model and a pre-built initial human voice prediction model respectively, to obtain a predicted background sound frequency spectrum output by the initial background sound prediction model and a predicted human voice frequency spectrum output by the initial human voice prediction model; Determining a first loss function value according to a real background audio spectrum and a predicted background audio spectrum of the sample background audio signal; Determining a second loss function value according to a real human voice spectrum and a predicted human voice spectrum of the sample human voice signal; The initial background sound prediction model and the initial human voice prediction model are trained respectively according to the average value of the first loss function value and the second loss function value to obtain the background sound prediction model and the human voice prediction model.

6. The method according to claim 2, characterized in that The standard vocabulary is constructed in the following way: Get multiple standard words; randomly dividing the plurality of standard words into a plurality of vocabulary categories; Selecting standard words from the plurality of standard words in turn as the words to be classified, calculating the minimum edit distance between the words to be classified and each vocabulary category, and classifying the words to be classified into the vocabulary category with the smallest minimum edit distance; wherein, for each vocabulary category, the minimum edit distance between the words to be classified and the vocabulary category refers to the average value of the minimum edit distances between the pinyin of the words to be classified and the pinyin of each standard word in the vocabulary category; Repeat the previous step until the division result of each standard word in the plurality of standard words does not change, thereby obtaining a plurality of vocabulary categories divided by the plurality of standard words; For each vocabulary category of the multiple standard words, the minimum edit distance between the pinyin of each standard word in the vocabulary category and the pinyin of other standard words in the vocabulary category is calculated respectively, and the standard word with the smallest average value of the minimum edit distance to the pinyin of other standard words in the vocabulary category is taken as the central word of the vocabulary category.

7. A speech transcription device, characterized in that: The device comprises: A separation module, used for separating an original human voice signal from an original speech signal; A detection module, configured to perform voice endpoint detection on the original human voice signal, so as to split the original human voice signal into at least one segment of human voice signal according to the detected voice endpoints; A recognition module is used to perform speech recognition on each human voice signal to obtain an initial speech text corresponding to each human voice signal; A segmentation module, used for predicting punctuation marks for each segment of initial speech text, adding punctuation marks to each segment of initial speech text, so as to segment each segment of initial speech text into at least one initial sentence; An error correction module, for determining, for each initial sentence, word error correction results for entity words in the initial sentence and word error correction results for individual words in the initial sentence; An integration module, used for integrating and correcting each initial sentence according to the word correction result and the character correction result of the initial sentence to obtain the transcription result of the original speech signal; When the separation module is used to separate the original human voice signal from the original speech signal, the separation module is used to: Acquire the frequency spectrum of the original speech signal; Inputting the spectrum into a pre-trained background sound prediction model to obtain a background sound spectrum predicted by the background sound prediction model; Inputting the spectrum into a pre-trained human voice prediction model to obtain a human voice spectrum predicted by the human voice prediction model; Determine a background sound energy sequence according to the background sound spectrum; wherein the background sound energy sequence includes a plurality of background sound energy values, and the plurality of background sound energy values ​​are energy values ​​at different frequencies in a frequency band of the background sound spectrum; Determine a human voice energy sequence according to the human voice spectrum; wherein the human voice energy sequence includes a plurality of human voice energy values, and the plurality of human voice energy values ​​are energy values ​​at different frequencies in the frequency band of the human voice spectrum; Determine, according to the background sound energy sequence and the human voice energy sequence, a sequence of energy proportions of the human voice signal in the original speech signal; Obtaining the spectrum of the human voice signal according to the energy proportion sequence of the human voice signal in the original voice signal and the spectrum of the original voice signal; Performing an inverse short-time Fourier transform on the frequency spectrum of the human voice signal to obtain the human voice signal.

8. An electronic device, characterized in that: include: A processor, a memory and a bus, wherein the memory stores machine-readable instructions executable by the processor, and when the electronic device is running, the processor and the memory communicate through the bus, and the machine-readable instructions are executed by the processor to execute the steps of a speech transcription method as described in any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of a speech transcription method as claimed in any one of claims 1 to 6 are executed.