DFCNN-based speech recognition method for elevator emergency rescue

Abstract

The embodiment of the invention relates to the technical field of voice recognition, in particular to a DFCNN-based voice recognition method for elevator emergency rescue, which comprises the following steps: acquiring voice data of trapped passengers in the elevator emergency rescue as sample data, and taking corresponding text data as labels for labeling to obtain a training sample set; constructing a DFCNN framework-based speech recognition model composed of a speech spectrogram conversion module, a DFCNN framework and an output module; based on a layer-by-layer greedy training algorithm, performing iterative training on the constructed speech recognition model by using the training sample set to obtain a trained speech recognition model; and deploying the trained voice recognition model in an artificial intelligence alarm receiving platform, performing voice recognition on the voice data of the trapped passengers received by the artificial intelligence alarm receiving platform, and outputting corresponding text data. According to the method, the speed and the stability of voice recognition are improved, and a technical basis with strong performance is provided for elevator rescue.

Description

Technical Field

[0001] The embodiments of the present application relate to the field of speech recognition technology, and in particular to a DFCNN-based speech recognition method for elevator emergency rescue. Background Art

[0002] Speech recognition technology, also known as Automatic Speech Recognition (ASR), is a computer technology that uses machine learning, deep learning, and other algorithms to automatically identify, understand, and process human speech signals and convert them into text or instructions. It is widely used in applications such as human-computer interaction and smart devices. The emergence of speech recognition technology enables computers and smart devices to understand human voice commands, questions, or conversations, and perform corresponding operations or provide corresponding information accordingly.

[0003] Speech recognition technology began in the 1950s. The development of its mainstream algorithmic models has gone through four stages: cognition, inception, development and application, and integrated application. In the cognition stage, speech recognition technology was primarily implemented using template matching. During this stage, systems (speech recognition systems) could only recognize isolated English numbers and a limited number of words, resulting in low accuracy in non-specific contexts. In the inception stage, probabilistic statistical models such as hidden Markov models and Gaussian mixture models gradually replaced template matching, allowing systems to advance from recognizing isolated words and numbers to recognizing connected words and continuous speech with a large vocabulary. Entering the development and application stage, the maturity and widespread adoption of deep neural networks has led to a qualitative breakthrough in speech recognition technology, significantly improving the accuracy of continuous speech recognition for large vocabularies and enabling the emergence of a wide range of speech functions and products. Currently, speech recognition technology is rapidly entering a new stage of integrated application. This phase, focused on meeting the core needs of diverse sectors, integrates multi-faceted pattern recognition, big data, the Internet of Things, cloud services, and other technologies to create numerous customized industry solutions, thereby facilitating human productivity and daily life.

[0004] As an indispensable means of transportation in modern society, elevators are closely integrated into human life and production, becoming a unique and crucial piece of equipment. When an elevator accident occurs, trapping passengers, quickly and effectively resolving the fault and ensuring safe rescue are crucial to their lives. Therefore, elevator cars are often equipped with emergency call systems, allowing trapped passengers to call for help. However, in the event of force majeure, such as high-energy loads causing widespread power outages, or torrential rain and lightning damaging high-voltage power lines, resulting in a large number of trapped people in a short period of time, the volume of incoming calls can increase dramatically, making it difficult for emergency personnel to promptly handle all the calls from trapped passengers. Furthermore, the elevator rescue system's control center could potentially be paralyzed by force majeure, preventing a large number of trapped passengers from receiving timely assistance.

[0005] To address this situation, a research team has applied speech recognition technology to elevator rescue scenarios, establishing an AI-powered alarm response platform to handle trapped passengers' calls for help. While current speech recognition technology can convert the vocabulary of human speech into computer-readable binary codes or character sequences, it still faces various difficulties in supporting deeper logic, understanding, judgment, decision-making, thinking, and analysis. Furthermore, maintaining a high recognition rate in the face of various uncertainties in continuous speech recognition involving large vocabularies is a challenge currently faced by speech recognition technology. Summary of the Invention

[0006] In order to solve the above technical problems, the embodiments of the present application propose a DFCNN-based speech recognition method for elevator emergency rescue, construct and train a DFCNN-based speech recognition framework to perform speech recognition tasks, greatly improve the speed and stability of speech recognition, and provide a powerful technical foundation for the artificial intelligence alarm receiving platform in elevator rescue, thereby better protecting the lives of trapped passengers.

[0007] On the one hand, an embodiment of the present application proposes a DFCNN-based speech recognition method for elevator emergency rescue, which is suitable for an artificial intelligence alarm receiving platform of an elevator emergency rescue system. The method includes the following steps: obtaining the voice data of trapped passengers in the elevator emergency rescue as sample data, and using the corresponding text data as labels for annotation, thereby obtaining a training sample set; constructing a speech recognition model based on the DFCNN framework; wherein, the speech recognition model based on the DFCNN framework is composed of a spectrogram conversion module, a DFCNN framework and an output module, and the DFCNN framework is composed of a number of convolutional pooling units connected in series. The spectrogram conversion module is used to convert the input voice data into a spectrogram, and the several convolutional pooling units are used to model their own input to express the long-term correlation of the voice data, and the output module is used to output text data corresponding to the input voice data; based on a layer-by-layer greedy training algorithm, the speech recognition model based on the DFCNN framework is iteratively trained using the training sample set to obtain a trained speech recognition model; the trained speech recognition model is deployed to the artificial intelligence alarm receiving platform, and the voice data of the trapped passengers received by the artificial intelligence alarm receiving platform are subjected to speech recognition, and the corresponding text data is output.

[0008] Optionally, voice data of trapped passengers in elevator emergency rescue is obtained as sample data, and the corresponding text data is used as labels for annotation, thereby obtaining a training sample set, including: crawling the voice data of trapped passengers from the historical trapped accident rescue records of the elevator emergency rescue system and dividing it into sentences, and incorporating it into the basic data pool; recording the voice data of volunteers simulating trapped passengers and dividing it into sentences, and incorporating it into the basic data pool; using a speech generation model, based on an artificial intelligence speech knowledge base for elevator emergency rescue, to simulate and generate voice data of trapped passengers and dividing it into sentences, and incorporating it into the basic data pool; performing data cleaning on the voice data in the basic data pool, using the cleaned voice data as sample data, and using the corresponding text data as labels for annotation, thereby obtaining a training sample set.

[0009] Optionally, the spectrogram conversion module consists of a pre-emphasis unit, a frame windowing unit, an FFT transformation unit and a multi-frame splicing unit; the pre-emphasis unit is used to filter the input voice data through a high-pass filter to improve the signal-to-noise ratio of the high-frequency part of the input voice data to obtain pre-emphasized voice data; the frame windowing unit is used to perform frame windowing processing on the pre-emphasized voice data according to a preset frame windowing size to obtain several frames of pre-emphasized voice data, and there is overlap between two adjacent frames of pre-emphasized voice data; the FFT transformation unit is used to perform FFT transformation on each frame of pre-emphasized voice data to obtain the spectrum of each frame of pre-emphasized voice data, and perform coordinate axis transformation using time and frequency as two dimensions to obtain a spectrogram of each frame of pre-emphasized voice data; the multi-frame splicing unit is used to splice the spectrogram of each frame of pre-emphasized voice data in the time dimension to finally obtain a spectrogram corresponding to the input voice data.

[0010] Optionally, each convolutional pooling unit in the DFCNN framework consists of several convolutional layers and a pooling layer connected in series. Each convolutional layer uses a small convolution kernel of size 3×3. The DFCNN framework accumulates a large number of convolutional pooling units to model the spectrogram corresponding to the input speech data to express the long-term correlation of the input speech data and obtain the spectrogram features of the input speech data.

[0011] Optionally, the output unit decodes based on the spectrogram features of the input speech data, and outputs text data that matches the syllables or Chinese characters and corresponds to the input speech data.

[0012] Optionally, when iteratively training the speech recognition model based on the DFCNN framework using the training sample set based on the layer-by-layer greedy training algorithm, only one layer of the speech recognition model based on the DFCNN framework is trained in each step. In the k-th step of training, the model parameters of the first k-1 layers that have been trained are fixed, and only the model parameters of the k-th layer are trained.

[0013] Optionally, while obtaining the training sample set, some labeled sample data are selected to generate a test sample set. Based on the layer-by-layer greedy training algorithm, the training sample set is used to iteratively train the speech recognition model based on the DFCNN framework to obtain the trained speech recognition model. The method also includes: based on the test sample set, performing a performance test on the trained speech recognition model, and the performance test items include word recognition accuracy, word recognition accuracy, sentence recognition accuracy, short sentence recognition average response time, long sentence recognition average response time and recognition throughput; when the test results of all items meet the preset standards, the trained speech recognition model is used as a mature speech recognition model and deployed to the artificial intelligence alarm receiving platform; when the test results of at least one item do not meet the preset standards, the model parameters of the trained speech recognition model are fine-tuned until the test results of all items meet the preset standards.

[0014] This application proposes a DFCNN-based speech recognition method for elevator emergency rescue, which constructs and trains a speech recognition model based on the DFCNN framework to perform speech recognition tasks. The speech recognition model based on the DFCNN framework directly converts the input speech data into a spectrogram for processing through a spectrogram conversion unit, with time and frequency as two dimensions, which effectively avoids the information loss caused by the filter group in the frequency domain and the information loss caused by the large frame shift in the time domain. After the spectrogram corresponding to the input speech data enters the DFCNN framework, several convolutional pooling units in series in the DFCNN framework will model the spectrogram to express the long-term correlation of the speech data. Such a DFCNN architecture can see very long historical and future information, thereby effectively improving the robustness of speech recognition. The output module inherits the DFCNN architecture and can directly output text data corresponding to the input speech data that matches the syllables or Chinese characters. When training the speech recognition model based on the DFCNN framework, a layer-by-layer greedy training algorithm is used to achieve end-to-end training of the entire model, and the training speed and model stability are greatly improved. Deploying such a DFCNN framework-based speech recognition model to the artificial intelligence alarm receiving platform can provide a powerful technical foundation for the artificial intelligence alarm receiving platform, thereby better protecting the lives of trapped passengers.

[0015] On the second aspect, the embodiment of the present application proposes a DFCNN-based speech recognition device for elevator emergency rescue, which is suitable for an artificial intelligence alarm receiving platform of an elevator emergency rescue system. The system includes: a sample data collection module for obtaining the voice data of trapped passengers in elevator emergency rescue as sample data, and using the corresponding text data as labels for annotation, thereby obtaining a training sample set; a model construction module for constructing a speech recognition model based on the DFCNN framework. The speech recognition model based on the DFCNN framework consists of a spectrogram conversion module, a DFCNN framework and an output module. The DFCNN framework consists of several convolutional pooling units connected in series. The spectrogram conversion module is used to convert the input speech data into a spectrogram. Several convolutional pooling units are used to model their own input to express the long-term correlation of the speech data. The output module is used to output text data corresponding to the input speech data; the model layer-by-layer training module is used to iteratively train the speech recognition model based on the DFCNN framework using the training sample set based on the layer-by-layer greedy training algorithm to obtain a trained speech recognition model; the model deployment and use module is used to deploy the trained speech recognition model to the artificial intelligence alarm platform, perform speech recognition on the voice data of trapped passengers received by the artificial intelligence alarm platform, and output the corresponding text data.

[0016] In the third aspect, an embodiment of the present application proposes an electronic device comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute a DFCNN-based speech recognition method for elevator emergency rescue as described in the first aspect above.

[0017] In the fourth aspect, an embodiment of the present application proposes a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, it can implement a DFCNN-based speech recognition method for elevator emergency rescue as described in the first aspect above.

[0018] It can be understood that the beneficial effects of the second to fourth aspects mentioned above can be found in the relevant description of the first aspect mentioned above, and will not be repeated here. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the related technologies, the following is a brief introduction to the drawings required for use in the embodiments of the present application or the description of the related technologies. Obviously, the following drawings are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work. The drawings described here are only used to explain the present application and are not used to limit the present application.

[0020] Figure 1 This is a flowchart of a DFCNN-based speech recognition method for elevator emergency rescue provided in one embodiment of the present application;

[0021] Figure 2 It is a structural diagram of a speech recognition model based on the DFCNN framework provided in one embodiment of the present application;

[0022] Figure 3 is a structural diagram of a spectrogram conversion module provided in one embodiment of the present application;

[0023] Figure 4 is a spectrogram provided in one embodiment of the present application;

[0024] Figure 5 It is a schematic diagram of the structure of the DFCNN framework provided in one embodiment of the present application;

[0025] Figure 6 1 is a structural diagram of a DFCNN-based speech recognition device for elevator emergency rescue provided in another embodiment of the present application;

[0026] Figure 7 It is a structural diagram of an electronic device provided in another embodiment of the present application. DETAILED DESCRIPTION

[0027] In order to make the purpose, technical solutions and advantages of the embodiments of the present application clearer, each embodiment of the present application will be described in detail below with reference to the accompanying drawings. Those skilled in the art will appreciate that in each embodiment of the present application, many technical details are provided to enable the reader to better understand the present application. However, even without these technical details and various changes and modifications based on the following embodiments, the technical solutions claimed in the present application can be implemented. The division of the following embodiments is for convenience of description and should not constitute any limitation on the specific implementation of the present application. The various embodiments can be combined with each other and referenced to each other under the premise of no contradiction.

[0028] An embodiment of the present application proposes a DFCNN-based speech recognition method for elevator emergency rescue, which is suitable for an artificial intelligence alarm receiving platform of an elevator emergency rescue system and is applied to electronic devices, wherein the electronic device can be a terminal or a server. This embodiment and the following embodiments are all described using a server as an example. The following is a specific description of the implementation details of the DFCNN-based speech recognition method for elevator emergency rescue proposed in this embodiment. The following content is only the implementation details provided for the convenience of understanding and is not necessary for the implementation of this solution.

[0029] The specific process of the DFCNN-based speech recognition method for elevator emergency rescue proposed in this embodiment can be as follows: Figure 1 Shown, including:

[0030] Step 101: Acquire voice data of trapped passengers in elevator emergency rescue as sample data, and use corresponding text data as labels for annotation, thereby obtaining a training sample set.

[0031] In practice, the server first needs to perform the basic work of model training. The first step is to build a training sample set. The server can obtain voice data from trapped passengers during elevator emergency rescue as sample data and annotate the corresponding text data as labels to obtain the training sample set.

[0032] In one example, while constructing the training sample set, the server also needs to select some labeled sample data to generate a test sample set in preparation for subsequent performance testing of the trained model.

[0033] In one example, the server obtains the voice data of trapped passengers in an elevator emergency rescue as sample data, and uses the corresponding text data as labels for annotation, and then divides the annotated sample data according to a preset ratio (such as 7:3) to obtain a training sample set and a test sample set.

[0034] In one example, given that elevator entrapment incidents are rare, simply acquiring the voice data of trapped passengers during elevator emergency rescue operations is insufficient as sample data. Therefore, the server can expand the sample data set in three ways. The first involves conventional real-world data. The server crawls the voice data of trapped passengers from historical rescue records of trapped passengers in the elevator emergency rescue system, segments it into sentences, and incorporates it into the base data pool. The second involves real-life simulations. The server records the voice data of volunteers simulating trapped passengers and segments it into sentences, which is then incorporated into the base data pool. The third involves model generation. Using a speech generation model, the server simulates and generates voice data of trapped passengers based on an AI speech knowledge base for elevator emergency rescue. This data is segmented into sentences and incorporated into the base data pool. The data in the base data pool varies in length and may contain noise. Therefore, the server performs data cleansing (including but not limited to filtering, alignment, normalization, and standardization) on the voice data in the base data pool. The cleaned voice data is used as sample data, and the corresponding text data is used as labels for annotation, thereby generating a training sample set.

[0035] Step 102, construct a speech recognition model based on the DFCNN framework. The speech recognition model based on the DFCNN framework consists of a spectrogram conversion module, a DFCNN framework and an output module. The DFCNN framework consists of several convolutional pooling units connected in series.

[0036] In the specific implementation, while building the training sample set, the server also needs to build a speech recognition model based on the DFCNN framework. The structure of the speech recognition model based on the DFCNN framework can be as follows: Figure 2 As shown in the figure, it specifically consists of a spectrogram conversion module, a DFCNN framework, and an output module. The DFCNN framework consists of several convolutional pooling units connected in series. The spectrogram conversion module is used to convert the input speech data into a spectrogram. Several convolutional pooling units are used to model their own input to express the long-term correlation of the speech data. The output module is used to output text data corresponding to the input speech data.

[0037] In one example, the structure of the spectrogram conversion module is as follows Figure 3 As shown, it is specifically composed of a pre-emphasis unit, a frame windowing unit, an FFT transformation unit and a multi-frame splicing unit.

[0038] The pre-emphasis unit filters the input voice data through a high-pass filter, improving the signal-to-noise ratio of the high-frequency portion of the input voice data, resulting in pre-emphasized voice data. The power spectrum of a voice signal decreases as frequency increases, resulting in most of the voice data's energy being concentrated in the low-frequency portion. This results in a very low signal-to-noise ratio in the high-frequency portion. Without pre-emphasis, useful information in the high-frequency portion is lost.

[0039] The frame windowing unit is configured to perform frame windowing on the pre-emphasized voice data according to a preset frame windowing size, thereby obtaining a plurality of frames of pre-emphasized voice data, wherein adjacent frames of pre-emphasized voice data overlap. After frame windowing, the voice data is converted from one-dimensional to two-dimensional.

[0040] The FFT transformation unit is used to perform FFT transformation on each frame of pre-emphasized speech data to obtain the spectrum of each frame of pre-emphasized speech data. In the spectrum, the horizontal axis represents frequency and the vertical axis represents amplitude. At this time, the coordinate axis needs to be rotated 90 degrees so that the horizontal axis represents amplitude and the vertical axis represents frequency. The amplitude is mapped by quantization, with 0 representing white and 255 representing black. The larger the amplitude, the darker the corresponding area, thereby removing the amplitude dimension and leaving an extra dimension to express time. That is, time and frequency are used as two dimensions for coordinate axis transformation, and finally a spectrogram of each frame of pre-emphasized speech data is obtained.

[0041] The multi-frame splicing unit is used to splice the spectrogram of each frame of pre-emphasized speech data in the time dimension, and finally obtain the spectrogram corresponding to the input speech data. Figure 4 .

[0042] In one example, Figure 2 、 Figure 5 As shown in the figure, each convolutional pooling unit in the DFCNN framework consists of several convolutional layers and a pooling layer in series (in Figure 5 In the

[15] , each convolutional pooling unit is composed of the first convolutional layer, the second convolutional layer and the pooling layer connected in series), and each convolutional layer uses a small convolution kernel of size 3×3. The DFCNN framework accumulates a large number of convolutional pooling units to model the spectrogram corresponding to the input speech data to express the long-term correlation of the input speech data and obtain the spectrogram features of the input speech data.

[0043] In one example, the output unit decodes the input speech data based on the spectrogram features and outputs text data that matches the syllables or Chinese characters and corresponds to the input speech data.

[0044] Step 103, based on the layer-by-layer greedy training algorithm, use the training sample set to iteratively train the speech recognition model based on the DFCNN framework to obtain a trained speech recognition model.

[0045] In the specific implementation, after completing the construction of the speech recognition model based on the DFCNN framework, the server will iteratively train the speech recognition model based on the DFCNN framework using the training sample set based on the layer-by-layer greedy training algorithm to obtain a trained speech recognition model.

[0046] When iteratively training the speech recognition model based on the DFCNN framework using the training sample set based on the layer-by-layer greedy training algorithm, each step of training will only train one layer of the speech recognition model based on the DFCNN framework. Therefore, in the k-th step of training, it is necessary to fix the model parameters of the first k-1 layers that have been trained, and only train the model parameters of the k-th layer.

[0047] In one example, after completing the model training, it is also necessary to perform a performance test on the trained speech recognition model based on the test sample set. The performance test items include but are not limited to word recognition accuracy, word recognition accuracy, sentence recognition accuracy, short sentence recognition average response time, long sentence recognition average response time and recognition throughput, etc. When the test results of all items meet the preset standards, the server can use the trained speech recognition model as a mature speech recognition model and deploy it to the artificial intelligence alarm receiving platform. When the test results of at least one item do not meet the preset standards, the server needs to fine-tune the model parameters of the trained speech recognition model until the test results of all items meet the preset standards.

[0048] The word recognition accuracy rate (Corr) is the number of correctly recognized words divided by the total number of labeled words. The word recognition accuracy rate (Acc) is (the number of correctly recognized words - the number of incorrectly recognized words) divided by the total number of labeled words. The sentence recognition accuracy rate (Sent) is the number of correctly recognized sentences divided by the total number of sentences.

[0049] The average response time for short sentence recognition is the average of all short sentence recognition response times. The short sentence recognition response time is the time the short sentence recognition result is returned minus the time the short sentence is received. Generally speaking, the average response time for short sentence recognition should be around 200ms. The average response time for long sentence recognition is the average of all long sentence recognition response times. The long sentence recognition response time is the time the long sentence recognition result is returned minus the time the long sentence is received. Generally speaking, the average response time for long sentence recognition should be around 700ms.

[0050] Throughput indicates how many seconds of speech data the speech recognition model can process per second. For example, a throughput of 100 means the speech recognition model can process 100 seconds of speech data in 1 second.

[0051] Step 104: deploy the trained speech recognition model to the artificial intelligence alarm receiving platform, perform speech recognition on the voice data of the trapped passengers received by the artificial intelligence alarm receiving platform, and output corresponding text data.

[0052] This embodiment proposes a DFCNN-based speech recognition method for elevator emergency rescue, which constructs and trains a speech recognition model based on the DFCNN framework to perform speech recognition tasks. The speech recognition model based on the DFCNN framework directly converts the input speech data into a spectrogram for processing using time and frequency as two dimensions through a spectrogram conversion unit, which effectively avoids the information loss caused by the filter group in the frequency domain and the information loss caused by large frame shift in the time domain. After the spectrogram corresponding to the input speech data enters the DFCNN framework, several convolutional pooling units in series in the DFCNN framework will model the spectrogram to express the long-term correlation of the speech data. Such a DFCNN architecture can see very long historical and future information, thereby effectively improving the robustness of speech recognition. The output module inherits the DFCNN architecture and can directly output text data corresponding to the input speech data that matches the syllables or Chinese characters. When training the speech recognition model based on the DFCNN framework, a layer-by-layer greedy training algorithm is used to achieve end-to-end training of the entire model, and the training speed and model stability are greatly improved. Deploying such a DFCNN framework-based speech recognition model to the artificial intelligence alarm receiving platform can provide a powerful technical foundation for the artificial intelligence alarm receiving platform, thereby better protecting the lives of trapped passengers.

[0053] The steps of the various methods described above are divided for clarity of description only. During implementation, they can be combined into a single step, or some steps can be broken down into multiple steps. As long as the same logical relationships are present, they are all within the scope of protection of this application. Adding minor modifications to the algorithm or process, or introducing minor design changes that do not change the core design of the algorithm or process, are also within the scope of protection of this application.

[0054] Another embodiment of the present application proposes a DFCNN-based speech recognition device for elevator emergency rescue, which is suitable for an artificial intelligence alarm receiving platform for an elevator emergency rescue system. The following is a detailed description of the details of the DFCNN-based speech recognition device for elevator emergency rescue proposed in this embodiment. The following content is only for the convenience of understanding the implementation details, and is not necessary for the implementation of this example. Figure 6 This is a structural diagram of a DFCNN-based speech recognition device for elevator emergency rescue proposed in this embodiment, including: a sample data collection module 201, a model construction module 202, a model layer-by-layer training module 203 and a model deployment and use module 204.

[0055] The sample data collection module 201 is used to obtain the voice data of trapped passengers in elevator emergency rescue as sample data, and use the corresponding text data as labels for annotation, so as to obtain a training sample set.

[0056] The model building module 202 is used to build a speech recognition model based on the DFCNN framework. The speech recognition model based on the DFCNN framework consists of a spectrogram conversion module, a DFCNN framework and an output module. The DFCNN framework consists of several convolutional pooling units connected in series. The spectrogram conversion module is used to convert the input speech data into a spectrogram. Several convolutional pooling units are used to model their own input to express the long-term correlation of the speech data. The output module is used to output text data corresponding to the input speech data.

[0057] The model layer-by-layer training module 203 is used to iteratively train the speech recognition model based on the DFCNN framework using the training sample set based on the layer-by-layer greedy training algorithm to obtain a trained speech recognition model.

[0058] The model deployment and use module 204 is used to deploy the trained speech recognition model to the artificial intelligence alarm receiving platform, perform speech recognition on the voice data of trapped passengers received by the artificial intelligence alarm receiving platform, and output corresponding text data.

[0059] It is not difficult to find that this embodiment is an apparatus embodiment corresponding to the above-mentioned method embodiment, and this embodiment can be implemented in conjunction with the above-mentioned method embodiment. The relevant technical details and technical effects mentioned in the above-mentioned method embodiment are still valid in this embodiment, and to reduce repetition, they are not repeated here. Accordingly, the relevant technical details mentioned in this embodiment can also be applied to the above-mentioned method embodiment.

[0060] It is worth mentioning that all modules and units involved in this embodiment are logical modules. In actual applications, a logical unit can be a physical unit, a part of a physical unit, or a combination of multiple physical units. In addition, to highlight the innovation of this application, this embodiment does not include units that are not closely related to solving the technical problem proposed by this application. However, this does not mean that other units do not exist in this embodiment.

[0061] Another embodiment of the present application provides an electronic device, the specific structure of which is as follows: Figure 7 As shown, the electronic device includes: at least one processor 301; and a memory 302 communicatively connected to the at least one processor 301; wherein the memory 302 stores instructions that can be executed by the at least one processor 301, and the instructions are executed by the at least one processor 301 so that the at least one processor 301 can execute a DFCNN-based speech recognition method for elevator emergency rescue as described in the above-mentioned method embodiments.

[0062] The memory and processor are connected using a bus, which includes any number of interconnected buses and bridges. The bus connects various circuits of one or more processors and memories. The bus can also connect various other circuits such as peripheral devices, voltage regulators, and power management circuits. These are all well known in the art and therefore will not be described further in this article. The bus interface provides an interface between the bus and the transceiver. The transceiver can be a single component or multiple components, such as multiple receivers and transmitters, providing a unit for communicating with various other devices on a transmission medium. Data processed by the processor is transmitted on a wireless medium via an antenna. Furthermore, the antenna also receives data and transmits it to the processor.

[0063] The processor is responsible for managing the bus and general processing, and can also provide various functions, including timing, peripheral interfaces, voltage regulation, power management, and other control functions. Memory can be used to store data used by the processor when performing operations.

[0064] Another embodiment of the present application proposes a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, it can implement a DFCNN-based speech recognition method for elevator emergency rescue as described in the above-mentioned method embodiments.

[0065] That is, those skilled in the art will understand that all or part of the steps in the above method embodiments can be implemented by instructing the relevant hardware through a program, which is stored in a storage medium and includes a number of instructions for causing a device (such as a single-chip microcomputer, chip, etc.) or a processor to execute all or part of the steps of the method described in each embodiment of the present application. Storage media include: U disk, mobile hard disk, read-only memory, random access memory, magnetic disk or optical disk, etc., various media that can store program code.

[0066] Those skilled in the art will appreciate that the above embodiments are specific embodiments for implementing the present application. In actual applications, various modifications may be made to the embodiments in form and detail without departing from the spirit and scope of the present application. Those skilled in the art may also make improvements and modifications without departing from the principles of the present application, and such improvements and modifications are also considered to be within the scope of protection of the present application.

Claims

1. A DFCNN-based speech recognition method for elevator emergency rescue, suitable for an artificial intelligence alarm receiving platform for an elevator emergency rescue system, characterized in that: The method comprises: Acquire the voice data of trapped passengers in elevator emergency rescue as sample data, and use the corresponding text data as labels for annotation to obtain a training sample set; Construct a speech recognition model based on the DFCNN framework; wherein, the speech recognition model based on the DFCNN framework consists of a spectrogram conversion module, a DFCNN framework and an output module; the DFCNN framework consists of several convolutional pooling units connected in series; the spectrogram conversion module is used to convert the input speech data into a spectrogram; several convolutional pooling units are used to model their own input to express the long-term correlation of the speech data; the output module is used to output text data corresponding to the input speech data; Based on the layer-by-layer greedy training algorithm, the speech recognition model based on the DFCNN framework is iteratively trained using the training sample set to obtain a trained speech recognition model; The trained speech recognition model is deployed to the artificial intelligence alarm receiving platform, and the voice data of trapped passengers received by the artificial intelligence alarm receiving platform are subjected to speech recognition to output the corresponding text data.

2. A DFCNN-based speech recognition method for elevator emergency rescue according to claim 1, characterized in that: The voice data of trapped passengers in elevator emergency rescue is obtained as sample data, and the corresponding text data is used as labels for annotation to obtain a training sample set, including: From the historical rescue records of trapped passengers in the elevator emergency rescue system, we crawled and segmented the voice data of trapped passengers into sentences and incorporated them into the basic data pool. Record voice data from volunteers simulating trapped passengers, segment them into sentences, and incorporate them into the basic data pool; Using a speech generation model and an AI speech knowledge base for elevator emergency rescue, we simulated and generated speech data of trapped passengers, segmented it by sentence, and incorporated it into the basic data pool. The speech data in the basic data pool is cleaned, the cleaned speech data is used as sample data, and the corresponding text data is used as labels for annotation, so as to obtain a training sample set.

3. A DFCNN-based speech recognition method for elevator emergency rescue according to claim 1, characterized in that: The spectrogram conversion module consists of a pre-emphasis unit, a frame windowing unit, an FFT transformation unit, and a multi-frame splicing unit; The pre-emphasis unit is used to filter the input voice data through a high-pass filter to improve the signal-to-noise ratio of the high-frequency part of the input voice data to obtain pre-emphasized voice data; The frame windowing unit is used to perform frame windowing processing on the pre-emphasized voice data according to a preset frame windowing size to obtain a plurality of frames of pre-emphasized voice data, and there is overlap between two adjacent frames of pre-emphasized voice data; The FFT transform unit is used to perform FFT transform on each frame of pre-emphasized speech data to obtain the spectrum of each frame of pre-emphasized speech data, and perform coordinate axis transformation using time and frequency as two dimensions to obtain a spectrogram of each frame of pre-emphasized speech data; The multi-frame splicing unit is used to splice the spectrogram of each frame of pre-emphasized speech data in the time dimension, and finally obtain the spectrogram corresponding to the input speech data.

4. A DFCNN-based speech recognition method for elevator emergency rescue according to claim 3, characterized in that: Each convolutional pooling unit in the DFCNN framework consists of several convolutional layers and a pooling layer connected in series. Each convolutional layer uses a small convolution kernel of size 3×3. The DFCNN framework accumulates a large number of convolutional pooling units to model the spectrogram corresponding to the input speech data to express the long-term correlation of the input speech data and obtain the spectrogram features of the input speech data.

5. A DFCNN-based speech recognition method for elevator emergency rescue according to claim 4, characterized in that: The output unit decodes the input speech data based on the spectrogram features and outputs text data that matches the syllables or Chinese characters and corresponds to the input speech data.

6. A DFCNN-based speech recognition method for elevator emergency rescue according to claim 1, characterized in that: When the speech recognition model based on the DFCNN framework is iteratively trained using the training sample set based on the layer-by-layer greedy training algorithm, only one layer of the speech recognition model based on the DFCNN framework is trained in each step. In the k-th step of training, the model parameters of the first k-1 layers that have been trained are fixed, and only the model parameters of the k-th layer are trained.

7. The DFCNN-based speech recognition method for elevator emergency rescue according to claim 6 is characterized in that: While obtaining the training sample set, a portion of the labeled sample data is selected to generate a test sample set. Based on a layer-by-layer greedy training algorithm, the training sample set is used to iteratively train a speech recognition model based on the DFCNN framework. After obtaining the trained speech recognition model, the method further includes: Based on the test sample set, the trained speech recognition model is tested for performance. The performance test items include word recognition accuracy, word recognition accuracy, sentence recognition accuracy, average response time for short sentence recognition, average response time for long sentence recognition, and recognition throughput. When the test results of all projects meet the preset standards, the trained speech recognition model will be regarded as a mature speech recognition model and deployed to the artificial intelligence alarm receiving platform; When the test result of at least one item does not meet the preset standard, the model parameters of the trained speech recognition model are fine-tuned until the test results of all items meet the preset standard.

8. A DFCNN-based speech recognition device for elevator emergency rescue, suitable for an artificial intelligence alarm receiving platform for an elevator emergency rescue system, characterized in that: The device comprises: The sample data collection module is used to obtain the voice data of trapped passengers in the elevator emergency rescue as sample data, and use the corresponding text data as labels for annotation, so as to obtain a training sample set; The model building module is used to build a speech recognition model based on the DFCNN framework. The speech recognition model based on the DFCNN framework consists of a spectrogram conversion module, a DFCNN framework, and an output module. The DFCNN framework consists of several convolutional pooling units connected in series. The spectrogram conversion module is used to convert the input speech data into a spectrogram. Several convolutional pooling units are used to model their own input to express the long-term correlation of the speech data. The output module is used to output text data corresponding to the input speech data. The model layer-by-layer training module is used to iteratively train the speech recognition model based on the DFCNN framework using the training sample set based on the layer-by-layer greedy training algorithm to obtain the trained speech recognition model; The model deployment and usage module is used to deploy the trained speech recognition model to the artificial intelligence alarm receiving platform, perform speech recognition on the voice data of trapped passengers received by the artificial intelligence alarm receiving platform, and output the corresponding text data.

9. An electronic device, characterized in that: include: at least one processor; as well as, a memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute a DFCNN-based speech recognition method for elevator emergency rescue as described in any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, it can implement a DFCNN-based speech recognition method for elevator emergency rescue as described in any one of claims 1 to 7.

Images