Voice quality inspection method and device, storage medium and electronic device
Through automated voice quality inspection methods, natural language processing technology and multiple matching modes are used to mark Q&A that meets the standards, solving the problem of insufficient labor and accuracy of traditional artificial quality inspection, and achieving efficient and accurate voice quality inspection results .
Patent Information
- Application Number
- CN202111601368.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2021-12-24
- Publication Date
- 2024-11-29
- Estimated Expiration
- 2041-12-24
AI Technical Summary
Traditional voice quality inspection methods rely on manual random inspections, consume manpower and material resources, and are difficult to ensure accuracy and stability, making it easy to miss inspections.
By determining the collection of short speech sentences of the questioner and answerer in the voice data to be quality-tested, we can obtain standard question-and-answer pairs and multiple matching patterns, and use natural language processing technology to automate quality inspection, select matching patterns in preset order, and mark them to meet the standards. , make sure all standard Q&A pairs are marked to pass quality inspection.
It realizes the automation and rapid quality inspection of voice quality inspection, overcomes the shortcomings of manual quality inspection, improves matching accuracy, prevents missed inspections, and ensures the accuracy and stability of quality inspection results.
Smart Images

Figure CN114242114B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of data processing technology, and in particular to a voice quality inspection method and device, a storage medium and an electronic device. Background Art
[0002] Nowadays, there are many industries that sell business through phone calls, such as the banking industry. In order to regulate the sales behavior of banking financial institutions' wealth management and agency sales products, the China Banking and Insurance Regulatory Commission requires that phone calls during the sales process be recorded. In order to ensure the compliance of the recorded data, early detection of illegal operations in the sales process, and reduce the litigation risks that banks may face, it is necessary to conduct voice quality inspection on the recorded data. In the recorded data, the account manager is the questioner and the customer is the respondent. The purpose of the voice quality inspection is to check whether the questioner in the recorded data has asked multiple target questions in accordance with regulations and whether the respondent has given clear answers to all target questions.
[0003] The traditional method of voice quality inspection is to conduct random inspections manually. This method not only consumes manpower and material resources, but is also prone to missed inspections. The accuracy and stability of the quality inspection results cannot be guaranteed. Summary of the invention
[0004] In view of the above problems, the present invention provides a method for overcoming the above problems or at least partially solving the above problems. The technical solution is as follows:
[0005] A voice quality inspection method, comprising:
[0006] Determine a first short sentence set and a second short sentence set in the voice data to be quality checked, wherein the first short sentence set includes a plurality of voice short sentences corresponding to the questioner, and the second short sentence set includes a plurality of voice short sentences corresponding to the answerer;
[0007] Obtain a set of standard question-answer pairs corresponding to a target voice service, wherein the set of standard question-answer pairs includes a plurality of standard question-answer pairs, each of which includes a standard question and a standard answer corresponding to the standard question; the target voice service is a voice service corresponding to the voice data to be inspected;
[0008] Determine a matching set corresponding to the voice data to be quality-checked, wherein the matching set includes a plurality of matching patterns;
[0009] Select matching modes in the matching set in sequence according to a preset selection order, and determine a target question-answer pair set corresponding to the standard question-answer pair set under the currently selected matching mode, wherein the target question-answer pair set is a subset of the standard question-answer pair set;
[0010] In the currently selected matching mode, for each standard question-answer pair in the target question-answer pair set, if there is a speech short sentence corresponding to the standard question in the standard question-answer pair in the first short sentence set, and there is a speech short sentence corresponding to the standard answer in the standard question-answer pair in the second short sentence set, then the standard question-answer pair is marked;
[0011] If in the currently selected matching mode, all standard question-answer pairs in the target question-answer pair set are marked, and the quality inspection of the speech data to be inspected is passed.
[0012] In the above method, optionally, the determining of the first set of short sentences and the second set of short sentences in the voice data to be quality inspected comprises:
[0013] Extracting the voice data of the questioner and the voice data of the answerer from the voice data to be inspected;
[0014] Segmenting the questioner's voice data into a plurality of questioner's voice short sentences to form the first short sentence set;
[0015] The respondent's voice data is segmented into a plurality of respondent's voice short sentences to form the second short sentence set.
[0016] In the above method, optionally, the extracting the voice data of the questioner and the voice data of the answerer from the voice data to be inspected includes:
[0017] Obtaining the preset voiceprint features of the questioner and the answerer;
[0018] The questioner's voice data and the answerer's voice data are extracted from the voice data to be inspected based on the questioner's voiceprint features and the answerer's voiceprint features.
[0019] The above method may optionally determine the target question-answer pair set corresponding to the standard question-answer pair set under the currently selected matching mode, including:
[0020] If the currently selected matching pattern is the first matching pattern selected in the matching set, determining the standard question-answer pair set as the target question-answer pair set;
[0021] If the currently selected matching pattern is not the first matching pattern selected in the matching set, the unmarked standard question and answer pairs in the target question and answer set corresponding to the matching pattern previously selected before the currently selected matching pattern are combined into a new target question and answer pair set as the target question and answer pair set corresponding to the currently selected matching pattern.
[0022] In the above method, optionally, determining whether there is a speech sentence corresponding to the standard question in the standard question-answer pair in the first sentence set includes:
[0023] Determine the feature extraction rule corresponding to the currently selected matching mode;
[0024] Applying the feature extraction rule to extract the target matching feature corresponding to the standard question in the standard question-answer pair;
[0025] Matching the target matching feature with each feature in the acquired first feature set corresponding to the first short sentence set;
[0026] If there is a feature matching the target matching feature in the first feature set, it is determined that there is a speech sentence corresponding to the standard question in the standard question-answer pair in the first sentence set.
[0027] In the above method, optionally, the process of acquiring the first feature set includes:
[0028] A preset sliding box is applied to traverse each speech short sentence in the first short sentence set in turn, and during the traversal process, a feature extraction rule corresponding to the currently selected matching mode is applied to extract features from each speech short sentence in the first short sentence set to form the first feature set.
[0029] The above method may optionally further include:
[0030] If, in the last matching mode in the matching set, all standard questions and answers in the target question and answer pair set corresponding to the last matching mode are not all marked, it is determined that the quality inspection of the speech data to be inspected fails.
[0031] A speech quality inspection device, comprising:
[0032] A first determining unit, configured to determine a first short sentence set and a second short sentence set in the voice data to be quality checked, wherein the first short sentence set includes a plurality of voice short sentences corresponding to the questioner, and the second short sentence set includes a plurality of voice short sentences corresponding to the answerer;
[0033] A second determining unit is configured to obtain a set of standard question-answer pairs corresponding to a target voice service, wherein the set of standard question-answer pairs includes a plurality of standard question-answer pairs, each of which includes a standard question and a standard answer corresponding to the standard question; the target voice service is a voice service corresponding to the voice data to be quality inspected;
[0034] A third determining unit, configured to determine a matching set corresponding to the voice data to be quality-checked, wherein the matching set includes a plurality of matching patterns;
[0035] An execution unit is used to select matching modes in the matching set in sequence according to a preset selection order, and determine a target question-answer pair set corresponding to the standard question-answer pair set under the currently selected matching mode, wherein the target question-answer pair set is a subset of the standard question-answer pair set; under the currently selected matching mode, the execution unit determines whether to mark each standard question-answer pair in the target question-answer pair set, and if there is a speech short sentence corresponding to the standard question in the standard question-answer pair in the first short sentence set, and there is a speech short sentence corresponding to the standard answer in the standard question-answer pair in the second short sentence set, then the standard question-answer pair is marked;
[0036] A judgment unit is used to judge whether all standard question-answer pairs in the target question-answer pair set are marked under the currently selected matching mode, and if so, pass the quality inspection of the voice data to be inspected.
[0037] A storage medium includes stored instructions, wherein when the instructions are executed, the device where the storage medium is located is controlled to execute the above-mentioned voice quality inspection method.
[0038] An electronic device includes at least one processor, and at least one memory and a bus connected to the processor; wherein the processor and the memory communicate with each other via the bus; and the processor is used to call program instructions in the memory to execute the above-mentioned voice quality inspection method.
[0039] Compared with the prior art, the present invention has the following advantages: a speech quality inspection method and device, a storage medium and an electronic device, specifically comprising: determining a first short sentence set and a second short sentence set in the speech data to be inspected, wherein the first short sentence set includes a plurality of speech short sentences corresponding to the questioner, and the second short sentence set includes a plurality of speech short sentences corresponding to the answerer; obtaining a standard question-answer pair set corresponding to a target speech service, wherein the standard question-answer pair set includes a plurality of standard question-answer pairs, each of the standard question-answer pairs includes a standard question and a standard answer corresponding to the standard question; the target speech service is a speech service corresponding to the speech data to be inspected; determining a matching set corresponding to the speech data to be inspected, wherein the matching set includes a plurality of matching patterns. Formula; select matching modes in the matching set in turn according to the preset selection order, and determine the target question-answer pair set corresponding to the standard question-answer pair set under the currently selected matching mode, and the target question-answer pair set is a subset of the standard question-answer pair set; under the currently selected matching mode, for each standard question-answer pair in the target question-answer pair set, if there is a speech short sentence corresponding to the standard question in the standard question-answer pair in the first short sentence set, and there is a speech short sentence corresponding to the standard answer in the standard question-answer pair in the second short sentence set, then the standard question-answer pair is marked; if all standard question-answer pairs in the target question-answer pair set are marked under the currently selected matching mode, the quality inspection of the speech data to be inspected is performed. In the scheme provided by the embodiment of the present invention, natural language processing technology is used to realize the automatic and rapid quality inspection of the speech data to be inspected, overcoming the shortcomings of manual quality inspection methods that consume manpower and material resources and cannot guarantee the accuracy and stability; and in the scheme provided by the embodiment of the present invention, multiple matching modes are designed to prevent missed inspections and improve matching accuracy.
[0040] The above description is only an overview of the technical solution of the present invention. In order to more clearly understand the technical means of the present invention, it can be implemented according to the contents of the specification. In order to make the above and other purposes, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are listed below. BRIEF DESCRIPTION OF THE DRAWINGS
[0041] Various other advantages and benefits will become apparent to those of ordinary skill in the art by reading the detailed description of the preferred embodiments below. The accompanying drawings are only for the purpose of illustrating the preferred embodiments and are not to be considered as limiting the present invention. Moreover, the same reference symbols are used throughout the accompanying drawings to represent the same components. In the accompanying drawings:
[0042] Figure 1 A method flow chart of a voice quality inspection method provided by an embodiment of the present invention;
[0043] Figure 2A flow chart of performing polling matching in a voice quality inspection method provided by an embodiment of the present invention;
[0044] Figure 3 A schematic diagram of the structure of a speech quality inspection device provided by an embodiment of the present invention;
[0045] Figure 4 A structural diagram of an electronic device provided by an embodiment of the present invention. DETAILED DESCRIPTION
[0046] The exemplary embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. Although the exemplary embodiments of the present disclosure are shown in the accompanying drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments set forth herein. On the contrary, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the present disclosure to those skilled in the art.
[0047] refer to Figure 1 FIG. 4 shows a method flow chart of a voice quality inspection method provided by an embodiment of the present invention. Figure 1 The method execution process shown is a feasible implementation scheme of the voice quality inspection method provided by the embodiment of the present invention. The method can be applied to various system platforms, and its execution subject is a server set in the system platform. The voice quality inspection method specifically includes:
[0048] S101: Determine a first short sentence set and a second short sentence set in the voice data to be quality checked, wherein the first short sentence set includes a plurality of voice short sentences corresponding to the questioner, and the second short sentence set includes a plurality of voice short sentences corresponding to the answerer;
[0049] In the method provided by the embodiment of the present invention, the voice data to be inspected is a dialogue in the form of question and answer. For example, when an account manager sells a product, in order to ensure compliance with sales behavior, the account manager is required to ask the customer multiple different target questions as a questioner in accordance with regulations, and the customer is required to give a clear and compliant answer to each target question as a respondent. Therefore, the acquired voice data to be inspected includes the voice data of the questioner and the voice data of the respondent. For the purpose of matching questions and answers, it is necessary to perform speaker diversion on the voice data to be inspected, and divide the voice data of the questioner and the respondent. This behavior of dividing the voice of the questioner and the voice of the respondent is called speaker diversion.
[0050] In the method provided by the embodiment of the present invention, the extracting of the questioner's voice data and the answerer's voice data from the voice data to be inspected, that is, performing speaker splitting, includes: obtaining a preset voiceprint feature of the questioner and a voiceprint feature of the answerer; and extracting the questioner's voice data and the answerer's voice data from the voice data to be inspected based on the voiceprint feature of the questioner and the voiceprint feature of the answerer.
[0051] In the method provided by the embodiment of the present invention, the speaker diversion can be performed by device diversion. When device diversion is used, two voice acquisition modules can be used to respectively record the voice data of the questioner and the voice data of the answerer, wherein the voice acquisition module can use a microphone. Then, the voice data acquired by the two voice acquisition modules are respectively subjected to noise reduction, such as reducing or even eliminating the voice data of the answerer in acquiring the voice data of the questioner, or reducing or even eliminating the voice data of the questioner in acquiring the voice data of the answerer. Finally, the voice data of the questioner and the voice data of the answerer after collection and noise reduction are stored for use in subsequent processes.
[0052] In the method provided by the embodiment of the present invention, the speaker splitting can also be performed by software splitting. When software splitting is used, a voice acquisition module such as a microphone can be used to record the voice of the questioner and the voice of the answerer. At this time, the voice data acquired by the voice acquisition module contains both the voice data of the questioner and the voice data of the answerer. The software splitting method can segment the voice data of the questioner and the answerer through natural language processing technology. The segmentation principle adopted in the method provided by the embodiment of the present invention can be mainly segmented according to the voiceprint information of different speakers, and is implemented through model training. The voice acquisition module acquires the voice data of the questioner and the answerer and stores them. The stored voice data of the questioner and the answerer are called during model training and application. During the model training and application process, the voiceprint information of different speakers is calculated according to the input voice data, and the voiceprint information of the questioner and the answerer are obtained. Then, the stored voice data containing the questioner and the answerer are segmented according to the voiceprint information of the questioner and the answerer, and the segmented voice data of the questioner and the answerer are obtained, and stored for use in subsequent processes.
[0053] In the method provided by the embodiment of the present invention, the voice data of the questioner and the voice data of the answerer are usually composed of long sentences. The calculation amount of long sentences is large when matching. Therefore, the method provided by the embodiment of the present invention divides the voice data of the questioner after speaker diversion into the first short sentence set, and divides the voice data of the answerer into the second short sentence set, so that when actually matching after segmentation, several independent short sentences are matched.
[0054] In the method provided by the embodiment of the present invention, short sentence segmentation can be implemented by using speech segmentation in natural language processing technology. Short sentence segmentation can ensure that the obtained speech data is a number of short sentences, so as to facilitate matching calculation.
[0055] In the method provided by an embodiment of the present invention, the determining of the first short sentence set and the second short sentence set in the voice data to be inspected includes: extracting the questioner voice data and the answerer voice data in the voice data to be inspected; segmenting the questioner voice data, dividing the questioner voice data into multiple questioner voice short sentences to form the first short sentence set; segmenting the answerer voice data, dividing the answerer voice data into multiple answerer voice short sentences to form the second short sentence set.
[0056] S102: Acquire a set of standard question-answer pairs corresponding to a target voice service, wherein the set of standard question-answer pairs includes a plurality of standard question-answer pairs, each of which includes a standard question and a standard answer corresponding to the standard question; the target voice service is a voice service corresponding to the voice data to be inspected;
[0057] In the method provided in the embodiment of the present invention, the purpose of voice quality inspection is to check whether the questioner and the answerer ask questions and give compliant answers in accordance with regulations. For the voice data to be quality inspected obtained under different voice services, the target questions to be inspected are also different. Therefore, a set of standard question-answer pairs corresponding to the target voice service is first obtained.
[0058] In the method provided by the embodiment of the present invention, each of the standard question-answer pairs includes a standard question and a standard answer corresponding to the standard question. Generally speaking, the standard answer corresponding to the standard question should be an affirmative answer, but in actual situations, when it is determined that the respondent has not answered a preset negative answer, it is considered that the respondent has given a standard answer corresponding to the standard question.
[0059] S103: Determine a matching set corresponding to the voice data to be quality-checked, wherein the matching set includes a plurality of matching patterns;
[0060] In the method provided by the embodiment of the present invention, multiple matching modes are used to perform voice quality inspection on the voice data to be inspected in turn. Each matching mode has different matching rules. Some standard question and answer pairs may not be matched when matched according to the matching rules of one of the matching modes, but may be matched in the next matching mode according to the matching rules of the next matching mode. Therefore, each matching mode can check and fill in the gaps with each other, which can improve the detection accuracy of voice quality inspection compared to the case where only one matching module is used.
[0061] S104: selecting matching patterns in the matching set in sequence according to a preset selection order;
[0062] Each time you select a matching mode, it includes:
[0063] S1041: Determine a target question-answer pair set corresponding to the standard question-answer pair set under the currently selected matching mode, wherein the target question-answer pair set is a subset of the standard question-answer pair set;
[0064] In the method provided by an embodiment of the present invention, in order to avoid repeated matching, after each matching pattern is completed, the standard question and answer pairs that have been matched in each matching pattern are removed from the target question and answer pair set used by the currently selected matching pattern and used as the target question and answer pair set for the next matching pattern; that is, from the target question and answer pair set used by each matching pattern, all standard question and answer pairs marked in the matching pattern are removed and used as the target question and answer pair set for the next matching pattern.
[0065] In the method provided by an embodiment of the present invention, the target question and answer pair set corresponding to the standard question and answer pair set under the currently selected matching mode is determined, including: if the currently selected matching mode is the first selected matching mode in the matching set, the standard question and answer pair set is determined as the target question and answer pair set; if the currently selected matching mode is not the first selected matching mode in the matching set, the unmarked standard question and answer pairs in the target question and answer set corresponding to the matching mode previously selected before the currently selected matching mode are combined into a new target question and answer pair set as the target question and answer pair set corresponding to the currently selected matching mode.
[0066] In the method provided by the embodiment of the present invention, the target question-answer pair set used in the first selected matching mode is the standard question-answer pair set. If there are n standard question-answer pairs in the standard question-answer pair set, it means that voice quality inspection is to be performed on n specified different questions and corresponding answers. If x1 questions and corresponding answers are matched in the first selected matching mode, the corresponding x1 standard question-answer pairs in the target question-answer pair set used by the first selected matching mode are marked, and the remaining n-x1 standard question-answer pairs in the target question-answer pair set used by the first selected matching mode are used to form the target question-answer pair set used by the second selected matching mode. If x2 questions and corresponding answers are matched in the second selected matching mode, the corresponding x2 standard question-answer pairs in the second selected matching mode are marked, and the remaining n-x1-x2 standard question-answer pairs in the target question-answer pair set used by the second selected matching mode are used to form the target question-answer pair set used by the third selected matching mode. By analogy, as long as the target question-answer pair set used by the currently selected matching mode has remaining unlabeled standard question-answer pairs after completing the current matching mode, and the currently selected matching mode is not the last matching mode, the remaining unlabeled standard question-answer pairs will be formed into the target question-answer pair set used by the next selected matching mode.
[0067] S1042: In the currently selected matching mode, for each standard question-answer pair in the target question-answer pair set, if there is a speech short sentence corresponding to the standard question in the standard question-answer pair in the first short sentence set, and there is a speech short sentence corresponding to the standard answer in the standard question-answer pair in the second short sentence set, then the standard question-answer pair is marked;
[0068] In the method provided by the embodiment of the present invention, each standard question and answer pair includes a standard question and a standard answer corresponding to the standard question. In each selected matching mode, only when the standard question and the corresponding standard answer of a standard question and answer pair are successfully matched, it means that the standard question and answer pair is successfully matched and can be marked.
[0069] In the method provided by the embodiment of the present invention, since the first short sentence set consisting of multiple short voice sentences corresponding to the questioner is extracted from the voice data to be inspected, the standard question of each standard question and answer pair can be matched and compared with each short voice sentence in the first short sentence set to determine whether there is a short voice sentence corresponding to the standard question in the standard question and answer pair in the first short sentence set. Similarly, since the second short sentence set consisting of multiple short voice sentences corresponding to the respondent is extracted from the voice data to be inspected, the standard answer of each standard question and answer pair can be matched and compared with each short voice sentence in the second short sentence set to determine whether there is a short voice sentence corresponding to the standard answer in the standard question and answer pair in the second short sentence set. In this way, classification before matching can improve matching efficiency and matching accuracy, and prevent standard questions from being matched in the respondent's voice data, resulting in waste of resources and mismatching.
[0070] In the method provided by the embodiment of the present invention, matching is performed in each matching mode according to the matching rules corresponding to the matching mode. For example, when a matching mode uses keywords as matching rules, it is necessary to extract the keywords of the standard question and standard answer in the standard question and answer pair, and compare them with the keywords extracted from the first short sentence set and the keywords extracted from the second short sentence set, respectively, to determine whether there is a match.
[0071] In the method provided by an embodiment of the present invention, determining whether there is a speech short sentence corresponding to the standard question in the standard question and answer pair in the first short sentence set includes: determining a feature extraction rule corresponding to the currently selected matching mode; applying the feature extraction rule to extract a target matching feature corresponding to the standard question in the standard question and answer pair; matching the target matching feature with each feature in the first feature set corresponding to the first short sentence set that has been acquired; if there is a feature in the first feature set that matches the target matching feature, determining that there is a speech short sentence corresponding to the standard question in the standard question and answer pair in the first short sentence set.
[0072] In the method provided by an embodiment of the present invention, the process of acquiring the first feature set includes applying a preset sliding box to traverse each speech sentence in the first sentence set in turn, and during the traversal process, applying the feature extraction rules corresponding to the currently selected matching pattern to extract features from each speech sentence in the first sentence set to form the first feature set.
[0073] In the method provided by the embodiment of the present invention, in order to avoid repeated matching, after each matching pattern is completed, the voice sentences corresponding to all standard questions that have been marked in the current matching pattern can be removed from the first sentence set. When the next matching pattern is performed, the voice sentences corresponding to the standard questions that have been marked in the previous matching pattern in the first sentence set will not be compared repeatedly.
[0074] In the method provided in an embodiment of the present invention, similarly, determining whether there is a speech short sentence corresponding to the standard answer in the standard question and answer pair in the second short sentence set includes: determining a feature extraction rule corresponding to the currently selected matching mode; applying the feature extraction rule to extract a target matching feature corresponding to the standard answer in the standard question and answer pair; matching the target matching feature with each feature in the acquired second feature set corresponding to the second short sentence set; if there is a feature in the second feature set that matches the target matching feature, determining that there is a speech short sentence in the second short sentence set corresponding to the standard answer in the standard question and answer pair.
[0075] In the method provided in an embodiment of the present invention, similarly, the process of acquiring the second feature set includes applying a preset sliding box to traverse each speech short sentence in the second short sentence set in turn, and during the traversal process, applying the feature extraction rules corresponding to the currently selected matching pattern to extract features from each speech short sentence in the second short sentence set to form the second feature set.
[0076] In the method provided in an embodiment of the present invention, similarly, in order to avoid repeated matching, after each matching pattern is completed, the voice sentences corresponding to all standard answers that have been marked in the current matching pattern can be removed from the second sentence set. When the next matching pattern is performed, the voice sentences corresponding to the standard answers that have been marked in the previous matching pattern in the second sentence set will not be compared repeatedly.
[0077] S1043: If in the currently selected matching mode, all standard question-answer pairs in the target question-answer pair set are marked and pass the quality inspection of the speech to be inspected.
[0078] In the method provided by an embodiment of the present invention, after each selected matching mode is completed, it is detected whether all standard question and answer pairs in the target question and answer pair set of the currently selected matching mode are marked. If it indicates that the quality inspection of the speech to be inspected is qualified, there is no need to perform subsequent matching modes with a set order; if there are remaining unmarked target question and answer pairs in the target question and answer pair set of the currently selected matching mode, the next selected matching mode is performed.
[0079] S1044: If, in the last matching pattern in the matching set, all standard questions and answers in the target question and answer pair set corresponding to the last matching pattern are not all marked, it is determined that the quality inspection of the speech to be inspected fails.
[0080] In the method provided by an embodiment of the present invention, after all the set matching modes are completed, there are still unmarked standard question and answer pairs in the target question and answer pair set corresponding to the last matching mode, which means that there are still unmarked standard question and answer pairs in the standard question and answer pair set after passing through all the matching modes, and the questioner and the respondent have not completed the questions and answers of the specified multiple target questions as required, and therefore the speech quality inspection of the speech to be inspected is judged to be unqualified.
[0081] In order to more clearly illustrate the speech quality inspection process of this application, the embodiment of the present invention provides a specific example as follows:
[0082] First, perform speaker splitting to extract the questioner's voice data and the answerer's voice data from the voice data to be inspected. Device splitting can be used to record the questioner and the answerer separately using two microphones; software splitting can also be used to record the questioner and the answerer together using one microphone, and then extract the questioner's voice data and the answerer's voice data based on the questioner's voiceprint features and the answerer's voiceprint features. The voice data extracted based on the voiceprint features can also be denoised to improve the accuracy of voice quality inspection.
[0083] Then, short sentence segmentation is performed to segment the questioner's voice data into a plurality of questioner's voice short sentences to form the first short sentence set; and the answerer's voice data is segmented into a plurality of answerer's voice short sentences to form the second short sentence set.
[0084] Next, according to the voice business corresponding to the voice data to be inspected, the corresponding set of standard question-answer pairs is obtained. Assuming that the voice business corresponding to the voice data to be inspected is the sales business of financial management and agency products by the account manager of a financial institution, before the sales business of the financial management and agency products is carried out according to the requirements of laws and regulations, the account manager needs to ask the customer N relevant questions, and the customer also needs to make clear and compliant answers to each question. Therefore, a set of standard question-answer pairs is established based on the N relevant questions and corresponding answers, and the set of standard question-answer pairs includes N standard question-answer pairs.
[0085] Subsequently, a sliding window quality inspection method is used to perform speech quality inspection at multiple matching levels. The matching mode of each matching level is different. In this embodiment, three matching modes, namely FAQ (Frequently Asked Questions) matching mode, keyword matching mode and key sound matching mode are used as examples to perform speech quality inspection.
[0086] When performing FAQ matching mode, since it is the first selected matching mode, the standard question-answer pair set is determined as the target question-answer pair set of the FAQ matching mode, that is, the target question-answer pair set of the FAQ matching mode contains N standard question-answer pairs. Similarly, since it is the first selected matching mode, each questioner's voice short sentence in the first short sentence set is matched, and each answerer's voice short sentence in the second short sentence set is also matched.
[0087] During the FAQ matching process, a window sliding method can be used to perform polling matching on each questioner's voice short sentence in the first short sentence set according to all standard questions in the target question-answer pair set of the FAQ matching mode; at the same time, polling matching on each answerer's voice short sentence in the second short sentence set according to all standard answers in the target question-answer pair set of the FAQ matching mode. Figure 2 , two sliding windows simultaneously traverse each speech short sentence in the first short sentence set and the second short sentence set. The sliding window size can be set to 3 and the sliding step size can be set to 1. In practical applications, it can be adjusted according to specific circumstances. During polling matching, feature extraction is performed on each speech short sentence of the questioner in the first sliding window according to the feature extraction rules of the FAQ matching pattern, and compared with the target matching features corresponding to all standard questions in the target question and answer pair set of the FAQ matching pattern extracted according to the feature extraction rules of the FAQ matching pattern; and feature extraction is performed on each speech short sentence of the answerer in the second sliding window according to the feature extraction rules of the FAQ matching pattern, and compared with the target matching features corresponding to all standard answers in the target question and answer pair set of the FAQ matching pattern extracted according to the feature extraction rules of the FAQ matching pattern.
[0088] In some embodiments, it is considered that only when the standard question in a standard question-and-answer pair matches the corresponding voice sentence of the questioner in the first sliding window within a set time range, and the standard answer in the standard question-and-answer pair also matches the corresponding voice sentence of the answerer in the second sliding window, the standard question-and-answer pair is considered to be matched successfully and the standard question-and-answer pair is marked. The reason for setting a time range is that if the standard question and the corresponding standard answer of a standard question-and-answer pair are matched separately at a long time interval, it cannot be guaranteed that the standard answer corresponds to the standard question. And in this case, the voice sentences in the first sentence set and the second sentence set should be sorted according to the time when the sentences occurred.
[0089] The feature extraction rules of the FAQ matching mode can be: for the target question-answer pair set of the FAQ matching mode, obtain the text data of each standard question-answer pair in the target question-answer pair set of the FAQ matching mode, and convert the text data of the standard question of each standard question-answer pair into a vector, and also convert the text data of the standard answer of each standard question-answer pair into a vector. For the first short sentence set and the second short sentence set to be matched, convert each questioner's voice data in the first short sentence set into text and then into a vector, and convert each answerer's voice data in the second short sentence set into text and then into a vector. Each vector converted in the first short sentence set is respectively calculated with the vector converted from each standard question in the target question-answer pair set of the FAQ matching mode, and each vector converted in the second short sentence set is respectively calculated with the vector converted from each standard answer in the target question-answer pair set of the FAQ matching mode. The similarity calculation method can adopt cosine similarity calculation, and the calculation formula of cosine similarity is:
[0090]
[0091] Where A and B are two m-dimensional vectors, A represents the vector converted from the voice data of the questioner in the first short sentence set selected by the sliding window, and B represents the vector converted from the text data of the standard question in the target question-answer pair set of the FAQ matching mode; or A represents the vector converted from the voice data of the answerer in the second short sentence set selected by the sliding window, and B represents the vector converted from the text data of the standard answer in the target question-answer pair set of the FAQ matching mode. When cosθ is greater than the set threshold, it means that the voice short sentence corresponding to A at this time successfully matches the standard question or standard answer corresponding to B.
[0092] When a similarity calculation result is greater than a set threshold, it is considered that the two vectors corresponding to the similarity calculation result greater than the set threshold match, and the questioner's voice sentences in the first sentence set corresponding to the two vectors successfully match the standard questions in the target question and answer pair set of the FAQ matching pattern, or the respondent's voice sentences in the second sentence set corresponding to the two vectors successfully match the standard answers in the target question and answer pair set of the FAQ matching pattern.
[0093] When the standard question and the corresponding annotated answer of a standard question and answer pair in the target question and answer pair set of the FAQ matching mode are matched, it is considered that the question corresponding to the standard question and answer pair is asked and answered in the speech data to be quality inspected.
[0094] Assume that X1 standard question and answer pairs are marked in the FAQ matching mode, check whether X1 is equal to N. If so, end the voice quality inspection and output the quality inspection qualified result. If not, remove the marked X1 standard question and answer pairs from the target question and answer pair set of the FAQ matching mode and use them as the second matching mode, that is, the target question and answer pair set of the keyword matching mode. The target question and answer pair set of the keyword matching mode has N-X1 standard question and answer pairs.
[0095] During the keyword matching process, the window sliding method can also be used to remove the questioner's voice sentences and answerer's voice sentences corresponding to the X1 standard question and answer pairs marked in the FAQ matching mode from the first and second sentence sets, and then perform polling matching on all the remaining voice sentences. According to the feature extraction rules of the keyword matching mode, the keywords of the N-X1 standard question and answer pairs of the target question and answer pair set of the keyword matching mode are extracted, and compared with the keywords of each voice sentence in the sliding window extracted according to the feature extraction rules of the keyword matching mode. When the standard question and standard answer of a standard question and answer pair are matched, the standard question and answer pair is marked.
[0096] Assume that X2 standard question-answer pairs are marked in the keyword matching mode, check whether X1+X2 is equal to N. If so, end the speech quality inspection and output the qualified quality inspection result. If not, remove the marked X2 standard question-answer pairs from the target question-answer pair set of the keyword matching mode, and use it as the third matching mode, that is, the target question-answer pair set of the key sound matching mode. The target question-answer pair set of the key sound matching mode has N-X1-X2 standard question-answer pairs. Since there may be dialects in the speech data to be quality inspected, and dialects are difficult to accurately match at the semantic level, in view of this situation, this embodiment proposes to adopt the key sound matching mode in the third matching mode, and the key sound can be the key sound of the dialect or a foreign language.
[0097] During the key-sound matching mode, the window sliding method can also be used to remove the questioner's voice sentences and the answerer's voice sentences corresponding to the marked X1+X2 standard question-answer pairs in the FAQ matching mode and the keyword matching mode from the first and second sentence sets, and then perform polling matching on all the remaining voice sentences. According to the feature extraction rules of the key-sound matching mode, the key sounds of the N-X1-X2 standard question-answer pairs of the target question-answer pair set of the key-sound matching mode are extracted, and compared with the key sounds of each voice sentence in the sliding window extracted according to the feature extraction rules of the key-sound matching mode. When the standard question and standard answer of a standard question-answer pair are matched, the standard question-answer pair is marked.
[0098] Assume that X3 standard question-answer pairs are marked in the key sound matching mode, check whether X1+X2+X3 is equal to N. If so, end the speech quality inspection and output the result of qualified quality inspection. If not, end the speech quality inspection and output the result of unqualified quality inspection.
[0099] In the specific example provided by the embodiment of the present invention, the voice data to be inspected is quality inspected at three matching levels: FAQ matching, keywords, and key sounds. Natural language processing technology is used to achieve automated and rapid quality inspection of the voice data to be inspected, overcoming the shortcomings of manual quality inspection methods that consume manpower and material resources and cannot guarantee accuracy and stability. In addition, compared with single-level quality inspection, the accuracy of quality inspection results is improved and missed inspections are avoided.
[0100] The embodiment of the present invention also provides a speech quality inspection device corresponding to a speech quality inspection method, wherein the speech quality inspection device is used to implement the application of the speech quality inspection method in practice. The structure diagram of the speech quality inspection device can be referred to Figure 3 , the speech quality inspection device comprises:
[0101] A first determining unit, configured to determine a first short sentence set and a second short sentence set in the voice data to be quality checked, wherein the first short sentence set includes a plurality of voice short sentences corresponding to the questioner, and the second short sentence set includes a plurality of voice short sentences corresponding to the answerer;
[0102] A second determining unit is configured to obtain a set of standard question-answer pairs corresponding to a target voice service, wherein the set of standard question-answer pairs includes a plurality of standard question-answer pairs, each of which includes a standard question and a standard answer corresponding to the standard question; the target voice service is a voice service corresponding to the voice data to be quality inspected;
[0103] A third determining unit, configured to determine a matching set corresponding to the voice data to be quality-checked, wherein the matching set includes a plurality of matching patterns;
[0104] An execution unit is used to select matching modes in the matching set in sequence according to a preset selection order, and determine a target question-answer pair set corresponding to the standard question-answer pair set under the currently selected matching mode, wherein the target question-answer pair set is a subset of the standard question-answer pair set; under the currently selected matching mode, the execution unit determines whether to mark each standard question-answer pair in the target question-answer pair set, and if there is a speech short sentence corresponding to the standard question in the standard question-answer pair in the first short sentence set, and there is a speech short sentence corresponding to the standard answer in the standard question-answer pair in the second short sentence set, then the standard question-answer pair is marked;
[0105] A judgment unit is used to judge whether all standard question-answer pairs in the target question-answer pair set are marked under the currently selected matching mode, and if so, pass the quality inspection of the voice data to be inspected.
[0106] In the speech quality inspection device provided by an embodiment of the present invention, the judgment unit is also used to judge whether all standard questions and answers in the target question and answer pair set corresponding to the last matching mode in the matching set are all marked. If so, the quality inspection of the speech to be inspected is passed; otherwise, it is determined that the quality inspection of the speech to be inspected has failed.
[0107] The speech quality inspection device provided in an embodiment of the present invention includes a processor and a memory. The above-mentioned first determination unit, second determination unit, third determination unit, execution unit and judgment unit are all stored in the memory as program units, and the processor executes the above-mentioned program units stored in the memory to realize corresponding functions.
[0108] The processor contains a kernel, which calls the corresponding program unit from the memory. One or more kernels can be set, and the voice quality inspection process can be dynamically executed by adjusting the kernel parameters to improve the update efficiency.
[0109] An embodiment of the present invention provides a storage medium on which a program is stored. When the program is executed by a processor, the speech quality inspection method is implemented.
[0110] An embodiment of the present invention provides a processor, which is used to run a program, wherein the voice quality inspection method is executed when the program is running.
[0111] like Figure 4 As shown, an embodiment of the present invention provides an electronic device, the electronic device includes at least one processor, and at least one memory and a bus connected to the processor; wherein the processor and the memory communicate with each other through the bus; the processor is used to call the program instructions in the memory to execute the above-mentioned voice quality inspection method. The device in this article can be a server, a PC, a PAD, a mobile phone, etc.
[0112] The present application also provides a computer program product, which, when executed on a data processing device, is suitable for executing a program for initializing the following method steps, including:
[0113] A voice quality inspection method, comprising:
[0114] Determine a first short sentence set and a second short sentence set in the voice data to be quality checked, wherein the first short sentence set includes a plurality of voice short sentences corresponding to the questioner, and the second short sentence set includes a plurality of voice short sentences corresponding to the answerer;
[0115] Obtain a set of standard question-answer pairs corresponding to a target voice service, wherein the set of standard question-answer pairs includes a plurality of standard question-answer pairs, each of which includes a standard question and a standard answer corresponding to the standard question; the target voice service is a voice service corresponding to the voice data to be inspected;
[0116] Determine a matching set corresponding to the voice data to be quality-checked, wherein the matching set includes a plurality of matching patterns;
[0117] Select matching modes in the matching set in sequence according to a preset selection order, and determine a target question-answer pair set corresponding to the standard question-answer pair set under the currently selected matching mode, wherein the target question-answer pair set is a subset of the standard question-answer pair set;
[0118] In the currently selected matching mode, for each standard question-answer pair in the target question-answer pair set, if there is a speech short sentence corresponding to the standard question in the standard question-answer pair in the first short sentence set, and there is a speech short sentence corresponding to the standard answer in the standard question-answer pair in the second short sentence set, then the standard question-answer pair is marked;
[0119] If in the currently selected matching mode, all standard question-answer pairs in the target question-answer pair set are marked, and the quality inspection of the speech data to be inspected is passed.
[0120] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.
[0121] In a typical configuration, the device includes one or more processors (CPU), memory and bus. The device may also include input / output interface, network interface and the like.
[0122] The memory may include non-permanent memory in a computer-readable medium, random access memory (RAM) and / or non-volatile memory in the form of read-only memory (ROM) or flash RAM, and the memory includes at least one memory chip. The memory is an example of a computer-readable medium.
[0123] Computer readable media include permanent and non-permanent, removable and non-removable media that can be implemented by any method or technology to store information. Information can be computer readable instructions, data structures, program modules or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technology, compact disk read-only memory (CD-ROM), digital versatile disk (DVD) or other optical storage, magnetic cassettes, magnetic tape magnetic disk storage or other magnetic storage devices or any other non-transmission media that can be used to store information that can be accessed by a computing device. As defined herein, computer readable media does not include temporary computer readable media (transitory media), such as modulated data signals and carrier waves.
[0124] It should also be noted that the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, commodity or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, commodity or device. In the absence of more restrictions, the elements defined by the sentence "comprises a ..." do not exclude the existence of other identical elements in the process, method, commodity or device including the elements.
[0125] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment or an embodiment in combination with software and hardware. Moreover, the present application may adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that contain computer-usable program code.
[0126] The above are only embodiments of the present application and are not intended to limit the present application. For those skilled in the art, the present application may have various changes and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application should be included within the scope of the claims of the present application.
Claims
1. A voice quality inspection method, It is characterized in that include: Determine a first short sentence set and a second short sentence set in the voice data to be quality checked, wherein the first short sentence set includes a plurality of voice short sentences corresponding to the questioner, and the second short sentence set includes a plurality of voice short sentences corresponding to the answerer; Obtain a set of standard question-answer pairs corresponding to a target voice service, wherein the set of standard question-answer pairs includes a plurality of standard question-answer pairs, each of which includes a standard question and a standard answer corresponding to the standard question; the target voice service is a voice service corresponding to the voice data to be inspected; Determine a matching set corresponding to the voice data to be quality-checked, wherein the matching set includes a plurality of matching patterns; Select matching modes in the matching set in sequence according to a preset selection order, and determine a target question-answer pair set corresponding to the standard question-answer pair set under the currently selected matching mode, wherein the target question-answer pair set is a subset of the standard question-answer pair set; In the currently selected matching mode, for each standard question-answer pair in the target question-answer pair set, if there is a speech short sentence corresponding to the standard question in the standard question-answer pair in the first short sentence set, and there is a speech short sentence corresponding to the standard answer in the standard question-answer pair in the second short sentence set, then the standard question-answer pair is marked; If in the currently selected matching mode, all standard question-answer pairs in the target question-answer pair set are marked, and the quality inspection of the speech data to be inspected is passed.
2. The method according to claim 1, It is characterized in that The determining of the first set of short sentences and the second set of short sentences in the speech data to be quality checked comprises: Extracting the voice data of the questioner and the voice data of the answerer from the voice data to be inspected; Segmenting the questioner's voice data into a plurality of questioner's voice short sentences to form the first short sentence set; The respondent's voice data is segmented into a plurality of respondent's voice short sentences to form the second short sentence set.
3. The method according to claim 2, It is characterized in that The extracting the voice data of the questioner and the voice data of the answerer from the voice data to be inspected includes: Obtaining the preset voiceprint features of the questioner and the answerer; The questioner's voice data and the answerer's voice data are extracted from the voice data to be inspected based on the questioner's voiceprint features and the answerer's voiceprint features.
4. The method according to claim 1, It is characterized in that Determining the target question-answer pair set corresponding to the standard question-answer pair set under the currently selected matching mode includes: If the currently selected matching pattern is the first matching pattern selected in the matching set, determining the standard question-answer pair set as the target question-answer pair set; If the currently selected matching pattern is not the first matching pattern selected in the matching set, the unmarked standard question and answer pairs in the target question and answer set corresponding to the matching pattern previously selected before the currently selected matching pattern are combined into a new target question and answer pair set as the target question and answer pair set corresponding to the currently selected matching pattern.
5. The method according to claim 1, It is characterized in that Determining whether there is a speech sentence corresponding to the standard question in the standard question-answer pair in the first sentence set includes: Determine the feature extraction rule corresponding to the currently selected matching mode; Applying the feature extraction rule to extract the target matching feature corresponding to the standard question in the standard question-answer pair; Matching the target matching feature with each feature in the acquired first feature set corresponding to the first short sentence set; If there is a feature matching the target matching feature in the first feature set, it is determined that there is a speech sentence corresponding to the standard question in the standard question-answer pair in the first sentence set.
6. The method according to claim 5, It is characterized in that The process of acquiring the first feature set includes: A preset sliding box is applied to traverse each speech short sentence in the first short sentence set in turn, and during the traversal process, a feature extraction rule corresponding to the currently selected matching mode is applied to extract features from each speech short sentence in the first short sentence set to form the first feature set.
7. The method according to claim 1, It is characterized in that Also includes: If, in the last matching mode in the matching set, all standard questions and answers in the target question and answer pair set corresponding to the last matching mode are not all marked, it is determined that the quality inspection of the speech data to be inspected fails.
8. A voice quality inspection device, It is characterized in that include: A first determining unit, configured to determine a first short sentence set and a second short sentence set in the voice data to be quality checked, wherein the first short sentence set includes a plurality of voice short sentences corresponding to the questioner, and the second short sentence set includes a plurality of voice short sentences corresponding to the answerer; A second determining unit is configured to obtain a set of standard question-answer pairs corresponding to a target voice service, wherein the set of standard question-answer pairs includes a plurality of standard question-answer pairs, each of which includes a standard question and a standard answer corresponding to the standard question; the target voice service is a voice service corresponding to the voice data to be quality inspected; A third determining unit, configured to determine a matching set corresponding to the voice data to be quality-checked, wherein the matching set includes a plurality of matching patterns; An execution unit is used to select matching modes in the matching set in sequence according to a preset selection order, and determine a target question-answer pair set corresponding to the standard question-answer pair set under the currently selected matching mode, wherein the target question-answer pair set is a subset of the standard question-answer pair set; under the currently selected matching mode, the execution unit determines whether to mark each standard question-answer pair in the target question-answer pair set, and if there is a speech short sentence corresponding to the standard question in the standard question-answer pair in the first short sentence set, and there is a speech short sentence corresponding to the standard answer in the standard question-answer pair in the second short sentence set, then the standard question-answer pair is marked; A judgment unit is used to judge whether all standard question-answer pairs in the target question-answer pair set are marked under the currently selected matching mode, and if so, pass the quality inspection of the voice data to be inspected.
9. A storage medium, It is characterized in that The storage medium includes stored instructions, wherein when the instructions are executed, the device where the storage medium is located is controlled to execute the speech quality inspection method according to any one of claims 1 to 7.
10. An electronic device, It is characterized in that It includes at least one processor, and at least one memory and a bus connected to the processor; wherein the processor and the memory communicate with each other through the bus; the processor is used to call the program instructions in the memory to execute the speech quality inspection method as described in any one of claims 1 to 7.
Citation Information
Patent Citations
Voice quality inspection method and system
CN113270114A
AI voice quality inspection method and device based on voice information, equipment and storage medium
CN113808574A
