Speech recognition intelligent rescue carriage and speech recognition method

Abstract

The invention relates to the technical field of voice recognition, and discloses a voice recognition intelligent rescue vehicle and a voice recognition method, and the method comprises the steps: calculating a prior probability and a conditional probability through historical data, and dynamically calculating an adjustment factor in combination with a distance coefficient between adjacent frames in a real-time feature vector to optimize a feature vector sequence, therefore, the accuracy and robustness of the speech recognition system are remarkably improved. Specifically, the prior probability provides global label distribution information to help the model to make reasonable prediction in the absence of clear feature support; the conditional probability reflects local association between each feature vector and a specific label, and the distinguishing ability of the model to different labels is enhanced. Meanwhile, context information is introduced by calculating the distance coefficient of the adjacent frames in real time, so that the current frame not only depends on the characteristic value of the current frame, but also considers the correlation of the front and rear frames, and the effectiveness of characteristic representation is further improved.

Description

Technical Field

[0001] The present invention relates to the technical field of speech recognition, and in particular to a speech recognition intelligent rescue vehicle and a speech recognition method. Background Art

[0002] Speech recognition is a technology that converts the content of human speech into computer-readable input. In simple terms, it enables machines to "understand" what people say and convert it into text or other forms of data. A common speech recognition method is an automatic speech recognition (ASR) system based on deep learning. The human voice signal is captured through a microphone and other devices, and the captured sound signal is processed by noise reduction and standardization to improve the accuracy of subsequent steps. Features are extracted from the speech signal using technologies such as Mel-frequency cepstral coefficients (MFCC). These features can effectively represent important information about the speech. Using a large amount of speech data and corresponding text labels, the model is trained through deep learning algorithms (such as recurrent neural networks (RNNs), long short-term memory networks (LSTMs), or convolutional neural networks (CNNs), so that the model can learn how to map speech features to corresponding text outputs. Speech signals are usually concentrated in the low frequency band (between 20Hz and 8kHz), and many common noise sources also occupy this frequency band. The presence of noise may cause the value of the feature vector to shift, making the originally clear speech features unclear, resulting in the inability of the model to obtain accurate text output when converting text. Summary of the invention

[0003] In order to solve the above technical problems, the present invention provides a speech recognition method, comprising: Step 1, capturing real-time voice signals through a microphone; Step 2, preprocessing the real-time speech signal to obtain a preprocessed real-time speech signal; Step 3, extracting features from the preprocessed real-time speech signal to obtain a real-time sequence of feature vectors; Step 4, constructing an adjustment factor for the real-time sequence of the feature vector, and adjusting the real-time sequence of the feature vector based on the adjustment factor to obtain an adjusted sequence of the feature vector; Step 5: Input the feature vector adjustment sequence into the pre-trained deep learning model and output the recognition result.

[0004] Furthermore, the feature vector real-time sequence includes one or several real-time feature vectors, and the real-time feature vector includes several frame real-time features.

[0005] Furthermore, an adjustment factor is constructed for the real-time sequence of the feature vectors, and the real-time sequence of the feature vectors is adjusted based on the adjustment factor to obtain an adjusted sequence of the feature vectors, specifically including: Step 41, presetting a number of estimated labels; Step 42, calculating the conditional probability of each frame of the historical feature of the historical feature vector sequence under each estimated label according to the historical speech data; Step 43, calculating the prior probability of each estimated label based on the historical speech recognition data; Step 44, calculating the distance coefficient between the real-time features of adjacent frames in each real-time feature vector; Step 45, calculating the adjustment factor of each real-time feature vector according to the conditional probability of each frame historical feature under each estimated label, the prior probability of each estimated label, and the distance coefficient between the real-time features of adjacent frames in each real-time feature vector; Step 46, obtaining an adjusted feature vector according to the adjustment factor and the real-time feature vector, and obtaining a feature vector adjustment sequence according to the adjusted feature vector.

[0006] Furthermore, the estimated tags include words, phrases, phonemes, characters, phoneme states, and subword units.

[0007] Furthermore, the conditional probability of each frame of historical features in the feature vector sequence under each estimated label is calculated based on the historical speech recognition data, including the following steps: Step 421, the historical speech recognition data includes a historical speech signal that is subjected to feature extraction to obtain a historical feature vector sequence, and the historical feature vector sequence is identified by a deep learning model to obtain an estimated label; Step 422, calibrating an estimated label for each historical feature in the historical feature vector sequence, and calculating the conditional probability of each frame of historical feature under each estimated label; The prior probability of each estimated label is calculated based on the historical speech recognition data, including: counting the number of all estimated labels, counting the number of historical features labeled as the current estimated label, and calculating the prior probability of the current estimated label based on the number of all estimated labels and the number of historical features labeled as the current estimated label.

[0008] Further, calculating the distance coefficient between the real-time features of adjacent frames in each real-time feature vector includes the following steps: Step 441, obtaining the real-time feature of the current frame from each real-time feature vector, obtaining the real-time feature of the previous frame according to the real-time feature of the current frame, and calculating the first adjacent similarity according to the real-time feature of the current frame and the real-time feature of the previous frame; Step 442, obtaining the real-time feature of the next frame according to the real-time feature of the current frame, and calculating the second adjacent similarity according to the real-time feature of the current frame and the real-time feature of the next frame; Step 443, calculating an adjustment coefficient according to the first adjacent similarity and the second adjacent similarity; Step 444: Calculate the distance coefficient of each feature vector according to the first adjacent similarity, the second adjacent similarity, and the adjustment coefficient.

[0009] Furthermore, the calculation formula of the adjustment factor is: ; In the formula, represents the adjustment factor of the ith real-time feature vector, represents the conditional probability of the lth frame real-time feature in the i-th real-time feature vector under the j-th estimated label, represents the prior probability that the l-th frame real-time feature in the i-th real-time feature vector is labeled as the j-th estimated label, represents the distance coefficient between the real-time feature of the lth frame and the real-time feature of the adjacent frame in the ith real-time feature vector, n represents the total number of real-time feature vectors, m represents the total number of estimated labels, q represents the total number of feature frames, Represents the adjustment factor.

[0010] Furthermore, the distance coefficient is calculated as follows: ; In the formula, represents the first adjacent similarity, represents the second adjacent similarity, The feature value representing the real-time feature of the lth frame, represents the feature value of the real-time feature of the previous frame l-1, Represents the feature value of the real-time feature of the next frame l+1 after the real-time feature of the lth frame.

[0011] Furthermore, the calculation formula of the feature vector adjustment sequence is: ; In the formula, represents the feature vector adjustment sequence, represents the adjustment factor of the ith real-time feature vector, Represents a real-time sequence of feature vectors.

[0012] A voice recognition intelligent rescue vehicle, comprising a rescue vehicle body and a voice control device installed on the rescue vehicle body, for executing a voice recognition method as described in any one of claims 1 to 9 above, characterized in that: an installation cavity is provided on the upper surface of the rescue vehicle body, a connection terminal is fixedly connected to the side wall of the installation cavity, the connection terminal is matched and aligned with the voice control device, a limiting opening is provided on the bottom surface of the installation cavity away from one end of the connection terminal, a spring cavity is provided on the bottom surface of the limiting opening, a lifting cavity is provided through one side of the spring cavity, a limiting component is connected inside the spring cavity, the upper end of the limiting component passes through the limiting opening to the inside of the installation cavity and abuts against the side wall of the voice control device, a pressing component is connected to the side end of the limiting component, the other end of the pressing component passes through the lifting cavity to the outside of the rescue vehicle body, a shielding component is connected to the inner wall of the installation cavity, and the other end of the shielding component abuts against the side of the voice control device away from the connection terminal.

[0013] The embodiments of the present invention have the following technical effects: The present invention calculates the prior probability and conditional probability through historical data, and combines the distance coefficient between adjacent frames in the real-time feature vector to dynamically calculate the adjustment factor to optimize the feature vector sequence, thereby significantly improving the accuracy and robustness of the speech recognition system. Specifically, the prior probability provides global label distribution information, helping the model to make reasonable predictions in the absence of clear feature support; the conditional probability reflects the local association between each feature vector and a specific label, enhancing the model's ability to distinguish different labels. At the same time, the real-time calculation of the distance coefficient of adjacent frames introduces contextual information, so that the current frame not only depends on its own eigenvalue, but also considers the correlation between the previous and next frames, further improving the effectiveness of feature representation. Combining these statistical information with contextual information, the generated adjustment factor can weight the original feature vector, highlight important features and suppress the influence of noise, thereby generating a more accurate feature vector adjustment sequence. Finally, the optimized feature vector adjustment sequence is input into the pre-trained deep learning model, which significantly improves the accuracy and reliability of the recognition results. This comprehensive method not only effectively improves the performance of the speech recognition system in complex environments, but also enhances its ability to cope with challenges such as noise and voice change, providing strong support for high-precision speech recognition in practical application scenarios. In addition, this method has good scalability and flexibility, and can be customized and optimized according to different application scenarios and requirements, further improving the adaptability and practicality of the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] In order to more clearly illustrate the specific implementation methods of the present invention or the technical solutions in the prior art, the drawings required for use in the specific implementation methods or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some implementation methods of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0015] Figure 1 is a flow chart of a speech recognition method provided by an embodiment of the present invention; Figure 2 A schematic diagram of the overall appearance of a voice recognition intelligent rescue vehicle provided by an embodiment of the present invention; Figure 3 A partial cross-sectional enlarged schematic diagram of a rescue vehicle body of a voice recognition intelligent rescue vehicle provided by an embodiment of the present invention; Figure 4 A detailed connection diagram of components such as a voice control device, a shielding component, and a limiting component of a voice recognition intelligent rescue vehicle provided by an embodiment of the present invention; Figure 5 For the present invention Figure 4 Exploded diagram of the various components.

[0016] In the figure: 1. rescue vehicle body; 2. voice control device; 3. connecting plate; 4. accordion plate; 5. connecting terminal; 6. clamping block; 7. sliding block; 8. pushing spring; 9. limiting plate; 10. limiting spring; 11. pressure plate; 12. shielding plate; 13. abutment plate; 101. installation cavity; 102. limiting opening; 103. spring cavity; 104. lifting cavity; 105. slide groove; 201. clamping mouth. DETAILED DESCRIPTION

[0017] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be described clearly and completely below. 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 belong to the scope of protection of the present invention.

[0018] Figure 1 is a flow chart of a speech recognition method provided by an embodiment of the present invention. Figure 1 , including: Step 1: Capture real-time speech signals through a microphone.

[0019] Step 2: preprocess the real-time speech signal to obtain a preprocessed real-time speech signal.

[0020] Preprocessing includes: Framing: Divide the continuous real-time speech signal into several short time frames (for example, 25 milliseconds per frame) with some overlap between frames (for example, 10 milliseconds) to analyze the local characteristics of the speech signal.

[0021] Windowing: Apply a Hanning window or other types of window functions to each frame of real-time speech signal to reduce spectral leakage, smooth frame boundaries, and make the signal more stable within the frame.

[0022] Pre-emphasis: Pre-emphasis is performed on each frame of the signal, usually through a high-pass filter to enhance the high-frequency components, thereby improving the quality of the voice signal.

[0023] Noise removal: If there is background noise, noise reduction algorithms (such as spectral subtraction) can be used to reduce noise interference and improve the clarity of the speech signal.

[0024] Step 3: extract features from the preprocessed real-time speech signal to obtain a real-time sequence of feature vectors.

[0025] Calculating the Mel-frequency cepstral coefficients (MFCC) is one of the most commonly used frequency domain feature extraction methods. The specific steps include: calculating the fast Fourier transform (FFT) of each frame to obtain the spectrum. Mapping the spectrum to the Mel scale to obtain the Mel spectrum. Taking the logarithm of the Mel spectrum and performing a discrete cosine transform (DCT) to obtain the MFCC coefficients. Directly using the energy of the Mel filter bank as a feature. The above is the prior art and will not be described in detail in this embodiment.

[0026] The feature vector real-time sequence includes one or several real-time feature vectors, and the real-time feature vector includes several frame real-time features.

[0027] Step 4: construct an adjustment factor for the real-time sequence of feature vectors, and adjust the real-time sequence of feature vectors based on the adjustment factor to obtain an adjusted sequence of feature vectors.

[0028] Step 41, presetting a number of estimated labels; Step 42, calculating the conditional probability of each frame of the historical feature of the historical feature vector sequence under each estimated label based on the historical speech data.

[0029] Step 421, the historical speech recognition data includes a historical feature vector sequence obtained after feature extraction of the historical speech signal, and the historical feature vector sequence is recognized by a deep learning model to obtain an estimated label; the estimated label includes words, phrases, phonemes, characters, phoneme states, and sub-word units.

[0030] Step 422, calibrate an estimated label for each historical feature in the historical feature vector sequence, and calculate the conditional probability of each frame of historical feature under each estimated label.

[0031] Step 43, calculating the prior probability of each estimated label based on the historical speech recognition data.

[0032] Count the number of all estimated labels, count the number of historical features that are labeled as the current estimated label, and calculate the prior probability of the current estimated label based on the number of all estimated labels and the number of historical features labeled as the current estimated label.

[0033] The process of calculating the prior probability and conditional probability in the above steps 42-43 belongs to the prior art and will not be described in detail in this embodiment.

[0034] Step 44, calculating the distance coefficient between the real-time features of adjacent frames in each real-time feature vector.

[0035] Step 441, obtaining the real-time feature of the current frame from each real-time feature vector, obtaining the real-time feature of the previous frame according to the real-time feature of the current frame, and calculating the first adjacent similarity according to the real-time feature of the current frame and the real-time feature of the previous frame.

[0036] Step 442, obtaining the real-time feature of the next frame according to the real-time feature of the current frame, and calculating the second adjacent similarity according to the real-time feature of the current frame and the real-time feature of the next frame.

[0037] The first adjacent similarity and the second similarity are calculated by cosine similarity.

[0038] Step 443, calculating an adjustment coefficient according to the first adjacent similarity and the second adjacent similarity; Step 444: Calculate the distance coefficient of each feature vector according to the first adjacent similarity, the second adjacent similarity, and the adjustment coefficient.

[0039] Step 45, calculating the adjustment factor of each real-time feature vector according to the conditional probability of each frame historical feature under each estimated label, the prior probability of each estimated label, and the distance coefficient between the real-time features of adjacent frames in each real-time feature vector; ; In the formula, represents the adjustment factor of the ith real-time feature vector, represents the conditional probability of the lth frame real-time feature in the i-th real-time feature vector under the j-th estimated label, represents the prior probability that the lth frame real-time feature in the i-th real-time feature vector is labeled as the j-th estimated label, represents the distance coefficient between the real-time feature of the lth frame and the real-time feature of the adjacent frame in the ith real-time feature vector, n represents the total number of real-time feature vectors, m represents the total number of estimated labels, q represents the total number of feature frames, Represents the adjustment factor.

[0040] ; In the formula, represents the first adjacent similarity, represents the second adjacent similarity, The feature value representing the real-time feature of the lth frame, represents the feature value of the real-time feature of the previous frame l-1, Represents the feature value of the real-time feature of the next frame l+1 after the real-time feature of the lth frame.

[0041] Step 46, obtaining an adjusted feature vector according to the adjustment factor and the real-time feature vector, and obtaining a feature vector adjustment sequence according to the adjusted feature vector.

[0042] ; In the formula, represents the feature vector adjustment sequence, represents the adjustment factor of the ith real-time feature vector, Represents a real-time sequence of feature vectors.

[0043] Step 5: Input the feature vector adjustment sequence into the pre-trained deep learning model and output the recognition result.

[0044] The feature vector adjustment sequence is fed into the model as input and forward propagated through the network layer. The model contains multiple levels of transformations and nonlinear activation functions, which can automatically extract complex patterns and structural information from the input data. Decoding output: The output of the model is usually a probability distribution of an estimated label. For speech recognition tasks, this means finding the text sequence that is most likely to correspond to the input speech content. Text transcription: Based on the probability distribution given by the model, the most likely character or word sequence is selected as the final recognition result.

[0045] See also Figure 2-Figure 5The present embodiment also discloses a voice recognition intelligent rescue vehicle, comprising a rescue vehicle body 1 and a voice control device 2 installed on the rescue vehicle body 1. An installation cavity 101 is provided on the upper surface of the rescue vehicle body 1. A connection terminal 5 is fixedly connected to the side wall of the installation cavity 101. The connection terminal 5 is matched and aligned with the voice control device 2. A limiting opening 102 is provided on the bottom surface of the installation cavity 101 away from the end of the connection terminal 5. A spring cavity 103 is provided on the bottom surface of the limiting opening 102. A lifting cavity 104 is provided on one side of the spring cavity 103. A limiting component is connected inside the spring cavity 103. The upper end of the limiting component passes through the limiting opening 102 to the inside of the installation cavity 101 and abuts against the side wall of the voice control device 2. A pressing component is connected to the side end of the limiting component. The other end of the pressing component passes through the lifting cavity 104 to the outside of the rescue vehicle body 1. A shielding component is connected to the inner wall of the installation cavity 101. The other end of the shielding component abuts against the side of the voice control device 2 away from the connection terminal 5.

[0046] The voice recognition intelligent rescue vehicle in the present invention is similar in structure to the existing voice recognition intelligent rescue vehicle, such as the voice recognition intelligent rescue vehicle disclosed in the patent with publication number CN215651533U. Figures 2 to 5 As shown, when the voice control device 2 of the voice recognition intelligent rescue vehicle in the present invention needs to be removed from the rescue vehicle body 1, it is only necessary to press the pressing component toward the bottom of the rescue vehicle body 1, and then the pressing component will descend with the limiting component, and then the limiting component will descend to cancel the limiting of the voice control device 2 from the side of the voice control device 2 away from the connecting terminal 5. At this time, the voice control device 2 is pushed toward the side away from the connecting terminal 5, and during the pushing process of the voice control device 2, the voice control device 2 will compress and tighten the shielding component again for a distance. When the connection between the voice control device 2 and the connecting terminal 5 is disconnected, the voice control device 2 can be unplugged from the top of the installation cavity 101, thereby separating the voice control device 2 and the rescue vehicle body 1, and as the voice control device 2 is pulled out, the shielding component lacks the extrusion limit, and the upper opening of the entire installation cavity 101 will be blocked, thereby reducing the probability of foreign objects falling here (if the installation cavity 101 needs to be cleaned and disinfected, the shielding component can be opened again), which is convenient and easy to operate.

[0047] After the voice control device 2 is cleaned and disinfected (and before the voice control device 2 is installed in the installation cavity 101 again, the installation cavity 101 can also be cleaned and disinfected), the shielding component is pushed toward the side away from the connecting terminal 5, and then the bottom surface of the voice control device 2 is pressed above the limiting component. Then, as the voice control device 2 descends, the limiting component will be pressed back into the spring cavity 103. After the voice control device 2 is inserted into the installation cavity 101, the voice control device 2 is pushed toward the direction of the connecting terminal 5 to form an effective electrical connection between the voice control device 2 and the rescue vehicle body 1. Then, the shielding component is released, and the shielding component will shield the cavity between the voice control device 2 and the rescue vehicle body 1 again, thereby reducing the probability of external contamination entering the installation cavity 101.

[0048] It should be noted that the connection terminal 5 and the voice control device 2 are electrically connected via wires, and the electrical connection is a mature technology, so it will not be described in detail here.

[0049] like Figure 2-Figure 4 As shown, the limit assembly includes a limit plate 9 and a plurality of limit springs 10, the lower ends of the plurality of limit springs 10 are fixedly connected to the bottom surface of the spring cavity 103, the upper ends of the plurality of limit springs 10 are fixedly connected to the bottom surface of the limit plate 9, the upper end of the limit plate 9 passes through the limit opening 102 to the installation cavity 101 and abuts against the side wall of the voice control device 2, the outer wall of the limit plate 9 is slidably connected to the inner wall of the limit opening 102, and the side wall of the limit plate 9 is connected to one end of the pressing assembly located in the spring cavity 103.

[0050] More specifically, in the limiting stage, the limiting plate 9 is pressed against the top of the spring cavity 103 under the action of a number of limiting springs 10, and the upper end of the limiting plate 9 is located in the installation cavity 101, and the side of the voice control device 2 away from the connecting terminal 5 is pressed against by one side of the limiting plate 9, thereby avoiding the loosening and separation of the voice control device 2 and the connecting terminal 5 due to the vibration of the rescue vehicle body 1.

[0051] When the limit needs to be released, the pressing component only needs to be pressed down, and the pressing component will descend with the limit plate 9, thereby removing the obstruction to the voice control device 2.

[0052] like Figure 1-Figure 4 As shown, the pressing assembly includes a pressing plate 11 and a shielding plate 12. One end of the pressing plate 11 is fixedly connected to the side wall of the limiting plate 9, and the other end of the pressing plate 11 passes through the lifting cavity 104 to the outside of the rescue vehicle body 1 and is fixedly connected to the upper end of the shielding plate 12. The width and thickness of the shielding plate 12 are greater than the width and thickness of the lifting cavity 104.

[0053] More specifically, when it is necessary to loosen the limit plate 9 on the voice control device 2, it is only necessary to press the pressure plate 11. The pressure plate 11 will descend with the shielding plate 12, and the pressure plate 11 will also descend with the limit plate 9, thereby removing the shielding of the limit plate 9 behind the voice control device 2, so as to facilitate the voice control device 2 to be slid out of the installation cavity 101.

[0054] It should be noted here that: no matter the voice control device 2 is installed in the installation cavity 101 or taken out for cleaning and disinfection, the limit plate 9 and the pressure plate 11 will be located at the upper end of the spring cavity 103 and the lifting cavity 104 (only at the moment of disassembly, the limit plate 9 and the pressure plate 11 are located at the lower end of the spring cavity 103 and the lifting cavity 104). Therefore, by setting the baffle plate 12, it is possible to prevent debris in the external environment from falling into the lifting cavity 104 and the spring cavity 103 and causing blockage.

[0055] like Figure 1-Figure 5 As shown, the shielding assembly includes a connecting plate 3 and an accordion plate 4, one end of the accordion plate 4 is fixedly connected to the inner wall of the installation cavity 101 away from the connecting terminal 5, the other end of the accordion plate 4 is fixedly connected to one side of the connecting plate 3, the other side of the connecting plate 3 is abutted against the side wall of the voice control device 2, and the side walls at both ends of the connecting plate 3 are slidably connected to the two sides of the inner wall of the installation cavity 101 respectively.

[0056] More specifically, when the voice control device 2 is normally located in the installation cavity 101, the connecting plate 3 and the accordion board 4 are located on the side away from the connecting terminal 5. When the voice control device 2 is taken out of the installation cavity 101, the accordion board 4 is pulled and unfolded by the connecting plate 3, so that the opening above the installation cavity 101 can be protected and shielded by the accordion board 4.

[0057] like Figure 2-Figure 5 As shown, slide grooves 105 are provided on both sides of the installation cavity 101, and sliders 7 are fixedly connected to both ends of the connecting plate 3. The outer walls of the two sliders 7 are slidably connected to the inner walls of the two slide grooves 105 respectively, and the two sliders 7 are fixedly connected to the side close to the accordion plate 4 with push springs 8, and the other ends of the two push springs 8 are fixedly connected to the inner walls of the two slide grooves 105 respectively.

[0058] More specifically, by providing the slide groove 105 and the slider 7, not only can the connecting plate 3 and the accordion plate 4 slide normally, but the connecting plate 3 and the accordion plate 4 will not fall off from the installation cavity 101. By providing the push spring 8, after the voice control device 2 is taken out of the installation cavity 101, the connecting plate 3 can be automatically pushed toward the connection terminal 5, thereby saving the steps and time of manual operation.

[0059] like Figure 2-Figure 5As shown, a butt plate 13 is fixedly connected to the bottom surface of the connecting plate 3 , and a side of the butt plate 13 away from the accordion plate 4 butts against a side wall of the limiting plate 9 .

[0060] More specifically, by providing the abutment plate 13, when the voice control device 2 is located in the installation cavity 101, the abutment plate 13 and the limiting plate 9 abut against each other to prevent the connecting plate 3 from continuously and permanently abutting against the outside of the voice control device 2, which may easily cause damage to the voice control device 2.

[0061] like Figure 2-Figure 5 As shown, blocks 6 are fixedly connected on both sides of the installation cavity 101 near the connecting terminal 5, and bayonet holes 201 are opened on both sides of the voice control device 2 near the connecting terminal 5. The outer walls of the two blocks 6 are slidably connected to the inner walls of the two bayonet holes 201 respectively.

[0062] More specifically, by providing the clamping block 6 and the clamping port 201 , the connection stability between the voice control device 2 and the installation cavity 101 can be further increased, thereby preventing the voice control device 2 from “bending” the connection terminal 5 due to up and down shaking or other reasons.

Claims

1. A speech recognition method, characterized in that: include: Step 1, capturing real-time voice signals through a microphone; Step 2, preprocessing the real-time speech signal to obtain a preprocessed real-time speech signal; Step 3, extracting features from the preprocessed real-time speech signal to obtain a real-time sequence of feature vectors; Step 4, constructing an adjustment factor for the real-time sequence of the feature vector, and adjusting the real-time sequence of the feature vector based on the adjustment factor to obtain an adjusted sequence of the feature vector; Step 5: Input the feature vector adjustment sequence into the pre-trained deep learning model and output the recognition result.

2. A speech recognition method according to claim 1, characterized in that: The feature vector real-time sequence includes one or several real-time feature vectors, and the real-time feature vector includes several frame real-time features.

3. A speech recognition method according to claim 1, characterized in that: An adjustment factor is constructed for the real-time sequence of the feature vector, and the real-time sequence of the feature vector is adjusted based on the adjustment factor to obtain an adjusted sequence of the feature vector, specifically including: Step 41, presetting a number of estimated labels; Step 42, calculating the conditional probability of each frame of the historical feature of the historical feature vector sequence under each estimated label according to the historical speech data; Step 43, calculating the prior probability of each estimated label based on the historical speech recognition data; Step 44, calculating the distance coefficient between the real-time features of adjacent frames in each real-time feature vector; Step 45, calculating the adjustment factor of each real-time feature vector according to the conditional probability of each frame historical feature under each estimated label, the prior probability of each estimated label, and the distance coefficient between the real-time features of adjacent frames in each real-time feature vector; Step 46, obtaining an adjusted feature vector according to the adjustment factor and the real-time feature vector, and obtaining a feature vector adjustment sequence according to the adjusted feature vector.

4. A speech recognition method according to claim 3, characterized in that: The estimated labels include words, phrases, phones, characters, phoneme states, and subword units.

5. A speech recognition method according to claim 3, characterized in that: Calculating the conditional probability of each frame of historical features in the feature vector sequence under each estimated label based on historical speech recognition data includes the following steps: Step 421, the historical speech recognition data includes a historical speech signal that is subjected to feature extraction to obtain a historical feature vector sequence, and the historical feature vector sequence is identified by a deep learning model to obtain an estimated label; Step 422, calibrating an estimated label for each historical feature in the historical feature vector sequence, and calculating the conditional probability of each frame of historical feature under each estimated label; The prior probability of each estimated label is calculated based on the historical speech recognition data, including: counting the number of all estimated labels, counting the number of historical features labeled as the current estimated label, and calculating the prior probability of the current estimated label based on the number of all estimated labels and the number of historical features labeled as the current estimated label.

6. A speech recognition method according to claim 5, characterized in that: Calculating the distance coefficient between the real-time features of adjacent frames in each real-time feature vector includes the following steps: Step 441, obtaining the real-time feature of the current frame from each real-time feature vector, obtaining the real-time feature of the previous frame according to the real-time feature of the current frame, and calculating the first adjacent similarity according to the real-time feature of the current frame and the real-time feature of the previous frame; Step 442, obtaining the real-time feature of the next frame according to the real-time feature of the current frame, and calculating the second adjacent similarity according to the real-time feature of the current frame and the real-time feature of the next frame; Step 443, calculating an adjustment coefficient according to the first adjacent similarity and the second adjacent similarity; Step 444: Calculate the distance coefficient of each feature vector according to the first adjacent similarity, the second adjacent similarity, and the adjustment coefficient.

7. A speech recognition method according to claim 3, characterized in that: The adjustment factor is calculated as: ； In the formula, represents the adjustment factor of the ith real-time feature vector, represents the conditional probability of the lth frame real-time feature in the i-th real-time feature vector under the j-th estimated label, represents the prior probability that the lth frame real-time feature in the i-th real-time feature vector is labeled as the j-th estimated label, represents the distance coefficient between the real-time feature of the lth frame and the real-time feature of the adjacent frame in the ith real-time feature vector, n represents the total number of real-time feature vectors, m represents the total number of estimated labels, q represents the total number of feature frames, Represents the adjustment factor.

8. A speech recognition method according to claim 7, characterized in that: The distance coefficient is calculated as: ； In the formula, represents the first adjacent similarity, represents the second adjacent similarity, The feature value representing the real-time feature of the lth frame, represents the feature value of the real-time feature of the previous frame l-1, Represents the feature value of the real-time feature of the next frame l+1 after the real-time feature of the lth frame.

9. A speech recognition method according to claim 1, characterized in that: The calculation formula of the feature vector adjustment sequence is: ； In the formula, represents the feature vector adjustment sequence, represents the adjustment factor of the ith real-time feature vector, Represents a real-time sequence of feature vectors.

10. A voice recognition intelligent rescue vehicle, comprising a rescue vehicle body and a voice control device installed on the rescue vehicle body, used to execute a voice recognition method according to any one of claims 1 to 9, characterized in that: An installation cavity is provided on the upper surface of the rescue vehicle body, and a connecting terminal is fixedly connected to the side wall of the installation cavity, and the connecting terminal is matched and aligned with the voice control device. A limiting opening is provided on the bottom surface of the installation cavity away from the end of the connecting terminal, and a spring cavity is provided on the bottom surface of the limiting opening. A lifting cavity is penetrated through one side of the spring cavity, and a limiting component is connected to the inside of the spring cavity. The upper end of the limiting component penetrates the limiting opening to the inside of the installation cavity and abuts against the side wall of the voice control device. A pressing component is connected to the side end of the limiting component, and the other end of the pressing component penetrates the lifting cavity to the outside of the rescue vehicle body. A shielding component is connected to the inner wall of the installation cavity, and the other end of the shielding component abuts against the side of the voice control device away from the connecting terminal.

Images