A method for enhancing the robustness of speech recognition based on fusion of multi-band speech signal features

Through the enhanced speech recognition robustness method of integrating multi-band voice signal characteristics, the inefficiency of information processing and information blocking of the power grid dispatch control center in severe abnormal situations is solved, automatic recording and efficient collaborative processing are realized, and the grid abnormality is improved. Efficiency and accuracy of disposal.

CN113744720BActive Publication Date: 2024-12-06STATE GRID HUBEI ELECTRIC POWER CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202111044761.3
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2021-09-07
Publication Date
2024-12-06
Estimated Expiration
2041-09-07

AI Technical Summary

Technical Problem

The power grid dispatching control center needs to spend a lot of time and energy on multi-level collaborative cooperation and information handling in severe abnormal situations. Telephone busyness may lead to information blockage, resulting in manual recording of the power grid abnormal handling process, which is inefficient.

Method used

Using a robust enhancement method of speech recognition based on the characteristics of fusion multi-band speech signal, the acoustic model training of speech signal feature extraction, fusion, speech recognition and vector space model construction are realized through speech signal feature extraction, fusion, speech recognition acoustic model training and vector space model construction, combined with TF-IDF and cosine similarity calculation, automatic recording and Information processing.

Benefits of technology

It reduces the workload and pressure of the dispatcher, avoids information blockage, and realizes automatic recording and efficient coordinated processing of the power grid abnormal handling process.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN113744720B_ABST
    Figure CN113744720B_ABST
Patent Text Reader

Abstract

The present invention discloses a method for enhancing the robustness of speech recognition based on the fusion of multi-band speech signal features, which specifically includes the following steps: S1, inputting information through a speech input module, performing feature selection through a feature selection module, and then performing feature extraction on speech signals PNCC of different frequency bands through a speech signal feature extraction unit, and further performing speech signal PNCC feature fusion after multiplying the speech signal PNCC feature of each frequency band by a weight coefficient through a speech multi-band feature fusion unit. The present invention relates to the field of power grid technology. The method for enhancing the robustness of speech recognition based on the fusion of multi-band speech signal features can make the power grid dispatching and control center only need to spend less time and energy of the dispatcher, and when serious power grid anomalies occur, there is no need for multi-level dispatching coordination and cooperation. In the scenario of commanding multiple units to uniformly handle anomalies, no serious information blockage will occur, and the power grid anomaly handling process record can be automatically recorded.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of power grid technology, and in particular to a method for enhancing the robustness of speech recognition based on the fusion of multi-band speech signal features. Background Art

[0002] With the rapid development of UHV AC / DC synchronous power grid, the large-scale centralized operation of new energy, and the in-depth promotion of power system reform, the coordinated handling of power grid anomalies is facing challenges. In order to further improve the level of emergency handling of power grid anomalies, it is necessary to take measures of integrated control and global resource allocation to deal with them. Especially in the case of power grid anomalies, joint early warning, coordinated pre-control, and coordinated handling must be achieved, which puts very high demands on the collaborative work level of institutions at all levels. The power grid control centers at all levels of the power system are the organization, command, guidance and coordination agencies for power grid operation. The dispatchers of the control centers, as the direct commanders of power grid operation, stick to the front line of power grid operation. With the continuous expansion of the national networking scale and the continuous improvement of voltage levels, the power grid has comprehensively surpassed the traditional UHV AC power grid in terms of the number of equipment, interconnection mode, coupling characteristics, and complexity. The problems faced by the power grid from natural, human, and internal factors are also more prominent, which has caused the workload, complexity and work pressure of dispatchers in dispatching the power grid to increase day by day. Therefore, the power sector uses artificial intelligence technology, and dispatchers can interact with machines verbally regarding dispatching and operating work. They can handle regular and repetitive work under the supervision and command of dispatchers, allowing machines to replace more human work, improve the intelligence level of dispatching and operating, improve work efficiency, eliminate human errors, and ensure the safe and stable operation of the power grid.

[0003] The existing power grid dispatching and control center requires dispatchers to spend a lot of time and energy. When serious power grid anomalies occur, multi-level dispatching is required to coordinate and cooperate to command multiple units to deal with the anomalies in a unified manner. If the phone is busy, serious information congestion will occur, and the power grid anomaly handling process needs to be recorded manually. Therefore, the present invention proposes a speech recognition robustness enhancement method based on the fusion of multi-band speech signal features to solve the above-mentioned problems. Summary of the invention

[0004] In view of the shortcomings of the prior art, the present invention provides a method for enhancing the robustness of speech recognition based on the fusion of multi-band speech signal features, which solves the problem that the existing power grid dispatching and control center requires dispatchers to spend a lot of time and energy, and when serious power grid anomalies occur, multi-level dispatching coordination and cooperation are required to command multiple units to deal with the anomalies in a unified manner. If the telephone is busy, serious information congestion will occur, and the power grid anomaly handling process needs to be recorded manually.

[0005] To achieve the above objectives, the present invention is implemented through the following technical solutions: a method for enhancing the robustness of speech recognition based on the fusion of multi-band speech signal features, specifically comprising the following steps:

[0006] S1, input information from the voice input module, perform feature selection through the feature selection module, then perform feature extraction on the PNCC of voice signals of different frequency bands through the voice signal feature extraction unit, and further perform voice signal PNCC feature fusion after multiplying the PNCC feature of the voice signal of each frequency band by a weight coefficient through the voice multi-band feature fusion unit;

[0007] S2, after the speech signal PNCC feature fusion is performed by the speech multi-band feature fusion unit, the model is generated and trained by the speech recognition acoustic model training unit;

[0008] S3, improving the robustness of the speech recognition model through the speech signal z-score normalization unit;

[0009] S4. Construct a vector space model (VSM): Use the power grid fault dictionary to segment the power grid fault cases and count the word frequencies, and combine the power grid fault entity weight table to obtain the power grid fault case feature vector;

[0010] S5, calculate the weight TF-IDF of each term in the vector;

[0011] S6, cosine similarity calculation;

[0012] S7. Perform similarity calculation based on the calculated value. The larger the value, the higher the information similarity of the feature vectors of the two power grid fault cases, and the more likely it is that duplicate names or multiple names will occur.

[0013] Preferably, the speech recognition acoustic model training unit in step S2 includes a training sample acquisition module, an original acoustic model acquisition module, an acoustic feature determination module, a state description model generation module and an acoustic model generation module. After acquiring the original acoustic model through the original acoustic model acquisition module, the acoustic state in the original acoustic model corresponding to each training text is determined.

[0014] Preferably, the acoustic feature determination module then determines the acoustic feature corresponding to each acoustic state based on the acoustic state and acoustic features corresponding to each training text, and then the state description model generation module uses the acoustic features corresponding to each acoustic state to retrain and obtain the state description model of the acoustic state.

[0015] Preferably, the acoustic model generation module finally uses the state model obtained by retraining to update the original state description model in the original acoustic model to obtain the updated acoustic model.

[0016] Preferably, the speech signal z-score standardization unit in step S3 includes a low-frequency signal enhancement module, a high-frequency signal suppression module and a frequency band speech signal weight coefficient adjustment module. The low-frequency signal enhancement module is used to enhance the low-frequency signal (human voice signal) characteristics, and the high-frequency signal suppression module is used to suppress the high-frequency signal (non-human voice signal) characteristics. The weight coefficients of speech signals in different frequency bands are adjusted to thereby improve the robustness of the speech recognition model.

[0017] Preferably, the power grid fault case feature vector in step S4 is composed of n weights, each word has a weight, and different words affect the importance of document relevance according to their weights in the document.

[0018] Preferably, in step S2, TF-IDF represents the product of TF (term frequency) and IDF (inverse document frequency): wherein: TF-IDF=term frequency (TF)×inverse document frequency (IDF).

[0019] Beneficial Effects

[0020] The present invention provides a method for enhancing the robustness of speech recognition based on the fusion of multi-band speech signal features. Compared with the prior art, the method has the following beneficial effects:

[0021] The method for enhancing the robustness of speech recognition based on the fusion of multi-band speech signal features specifically includes the following steps: S1, inputting information through a speech input module, performing feature selection through a feature selection module, then performing feature extraction on PNCC of speech signals of different frequency bands through a speech signal feature extraction unit, and further performing speech signal PNCC feature fusion after multiplying the speech signal PNCC feature of each frequency band by a weight coefficient through a speech multi-band feature fusion unit; S2, performing model generation and training through a speech recognition acoustic model training unit after performing speech signal PNCC feature fusion through the speech multi-band feature fusion unit; S3, improving the speech recognition model through a speech signal z-score normalization unit Robustness; S4. Construct a vector space model (VSM): Use the power grid fault dictionary to segment the power grid fault cases and count the word frequencies, and combine the power grid fault entity weight table to obtain the power grid fault case feature vector; S5. Calculate the weight TF-IDF of each term in the vector; S6. Cosine similarity calculation; S7. Perform similarity calculation based on the calculated value. The larger the value, the higher the information similarity of the two power grid fault case feature vectors, and the more likely it is that duplicate names or multiple names will occur. This allows the power grid dispatching and control center to spend less time and energy on dispatchers. When serious power grid anomalies occur, there is no need for multi-level dispatching to coordinate and cooperate. In the scenario of commanding multiple units to uniformly handle anomalies, if the phone is busy, there will be no serious information congestion, and the power grid anomaly handling process record can be automatically recorded. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 It is a schematic diagram of the overall structure of the present invention;

[0023] Figure 2 Schematic diagram of a sound recognition acoustic model training unit of the present invention;

[0024] Figure 3 Schematic diagram of a speech signal z-score normalization unit according to the present invention. DETAILED DESCRIPTION

[0025] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0026] See also Figure 1-3 The present invention provides a technical solution: a method for enhancing the robustness of speech recognition based on the fusion of multi-band speech signal features, which specifically includes the following steps:

[0027] S1, input information from the voice input module, perform feature selection through the feature selection module, then perform feature extraction on the PNCC of voice signals of different frequency bands through the voice signal feature extraction unit, and further perform voice signal PNCC feature fusion after multiplying the PNCC feature of the voice signal of each frequency band by a weight coefficient through the voice multi-band feature fusion unit;

[0028] S2, after the speech signal PNCC feature fusion is performed by the speech multi-band feature fusion unit, the model is generated and trained by the speech recognition acoustic model training unit;

[0029] S3, improving the robustness of the speech recognition model through the speech signal z-score normalization unit;

[0030] S4. Construct a vector space model (VSM): Use the power grid fault dictionary to segment the power grid fault cases and count the word frequencies, and combine the power grid fault entity weight table to obtain the power grid fault case feature vector;

[0031] S5, calculate the weight TF-IDF of each term in the vector;

[0032] S6, cosine similarity calculation;

[0033] S7. Perform similarity calculation based on the calculated value. The larger the value, the higher the information similarity of the feature vectors of the two power grid fault cases, and the more likely it is that duplicate names or multiple names will occur.

[0034] In an embodiment of the present invention, the speech recognition acoustic model training unit in step S2 includes a training sample acquisition module, an original acoustic model acquisition module, an acoustic feature determination module, a state description model generation module and an acoustic model generation module. After the original acoustic model is acquired through the original acoustic model acquisition module, the acoustic state in the original acoustic model corresponding to each training text is determined.

[0035] In an embodiment of the present invention, the acoustic feature determination module determines the acoustic features corresponding to each acoustic state based on the acoustic state and acoustic features corresponding to each training text, and then the state description model generation module uses the acoustic features corresponding to each acoustic state to retrain and obtain the state description model of the acoustic state.

[0036] In the embodiment of the present invention, the acoustic model generation module finally uses the state model obtained by retraining to update the original state description model in the original acoustic model to obtain the updated acoustic model.

[0037] In an embodiment of the present invention, the speech signal z-score standardization unit in step S3 includes a low-frequency signal enhancement module, a high-frequency signal suppression module and a frequency band speech signal weight coefficient adjustment module. The low-frequency signal enhancement module is used to enhance the low-frequency signal (human voice signal) characteristics, and the high-frequency signal suppression module is used to suppress the high-frequency signal (non-human voice signal) characteristics. The weight coefficients of speech signals in different frequency bands are adjusted to thereby improve the robustness of the speech recognition model.

[0038] In the embodiment of the present invention, the power grid fault case feature vector in step S4 is composed of n weights, each word has a weight, and different words affect the importance of document relevance according to their weights in the document.

[0039] In the embodiment of the present invention, the TF-IDF in step S2 represents the product of TF (term frequency) and IDF (inverse document frequency): wherein: TF-IDF=term frequency (TF)×inverse document frequency (IDF).

[0040] Meanwhile, the contents not described in detail in this specification belong to the prior art known to those skilled in the art.

[0041] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article 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, article or device.

[0042] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for enhancing the robustness of speech recognition based on the fusion of multi-band speech signal features, characterized in that: The specific steps include: S1, input information from the voice input module, perform feature selection through the feature selection module, then perform feature extraction on the PNCC of voice signals of different frequency bands through the voice signal feature extraction unit, and further perform voice signal PNCC feature fusion after multiplying the PNCC feature of the voice signal of each frequency band by a weight coefficient through the voice multi-band feature fusion unit; S2, after the speech signal PNCC feature fusion is performed by the speech multi-band feature fusion unit, the model is generated and trained by the speech recognition acoustic model training unit; S3. Improving the robustness of the speech recognition model through a speech signal z-score standardization unit, wherein the speech signal z-score standardization unit includes a low-frequency signal enhancement module, a high-frequency signal suppression module, and a frequency band speech signal weight coefficient adjustment module. The low-frequency signal enhancement module is used to enhance the low-frequency signal (human voice signal) feature, and the high-frequency signal suppression module is used to suppress the high-frequency signal (non-human voice signal) feature. The weight coefficients of speech signals in different frequency bands are adjusted to thereby improve the robustness of the speech recognition model. S4. Construct a vector space model (VSM): Use the power grid fault dictionary to segment the power grid fault cases and count the word frequencies, and combine the power grid fault entity weight table to obtain the power grid fault case feature vector; S5, calculate the weight TF-IDF of each term in the vector; S6, cosine similarity calculation; S7. Perform similarity calculation based on the calculated value. The larger the value, the higher the information similarity of the feature vectors of the two power grid fault cases, and the more likely it is that duplicate names or multiple names will occur.

2. The method for enhancing the robustness of speech recognition based on fusion of multi-band speech signal features according to claim 1, characterized in that: The speech recognition acoustic model training unit in step S2 includes a training sample acquisition module, an original acoustic model acquisition module, an acoustic feature determination module, a state description model generation module and an acoustic model generation module. After the original acoustic model is acquired through the original acoustic model acquisition module, the acoustic state in the original acoustic model corresponding to each training text is determined.

3. The method for enhancing the robustness of speech recognition based on fusion of multi-band speech signal features according to claim 2, characterized in that: Then, the acoustic feature determination module determines the acoustic features corresponding to each acoustic state based on the acoustic state and acoustic features corresponding to each training text. Then, the state description model generation module uses the acoustic features corresponding to each acoustic state to retrain and obtain the state description model of the acoustic state.

4. The method for enhancing the robustness of speech recognition based on fusion of multi-band speech signal features according to claim 2, characterized in that: Finally, the acoustic model generation module uses the retrained state model to update the original state description model in the original acoustic model to obtain the updated acoustic model.

5. The method for enhancing the robustness of speech recognition based on fusion of multi-band speech signal features according to claim 1, characterized in that: The power grid fault case feature vector in step S4 is composed of n weights, each word has a weight, and different words affect the importance of document relevance according to their weights in the document.

6. The method for enhancing the robustness of speech recognition based on fusion of multi-band speech signal features according to claim 1, characterized in that: In step S2, TF-IDF represents the product of TF (term frequency) and IDF (inverse document frequency): wherein: TF-IDF=term frequency (TF)×inverse document frequency (IDF).