An intelligent system and method for speech recognition based on communication scheduling instructions

Abstract

The present invention belongs to speech recognition technology in the field of computers, and in particular to an intelligent system and method for speech recognition based on communication scheduling instructions. The present invention manages the scheduling instruction data throughout its life cycle. The deep residual shrinkage network DRSN‑CS is used to reduce noise or process redundant information in speech data. In the process of model training, an open source dictionary is used in combination with a custom professional dictionary to improve the accuracy.

Description

Technical Field

[0001] The invention belongs to the field of computer speech recognition technology, in particular to an intelligent system and method for speech recognition based on communication scheduling instructions. Background Art

[0002] The work of intelligent communication dispatching is characterized by strong professionalism, complex operation and high accuracy requirements. At present, dispatching work mainly involves operators manually identifying the identities of the personnel who issue instructions for communication dispatching, and then the operators operate according to the instructions. The existing intelligent communication dispatching work process mainly uses the dispatching telephone system, and the commander and operator complete the entire process of issuing orders, recording, repeating, confirming, operating and returning orders. In this process, the content of the instructions is manually checked, and the efficiency, accuracy and safety need to be improved.

[0003] The present invention uses voiceprint recognition and speech recognition in the communication intelligent dispatching system to remotely collect and process the voice signals of dispatchers, analyze the voiceprint information to confirm the dispatcher and operator, and convert the dispatching voice instructions into text for storage. The present invention recognizes the dispatching instructions, effectively improves the accuracy and work efficiency of speech recognition, eliminates manual operation errors, and can better meet the needs of speech recognition in the process of communication intelligent dispatching. Summary of the invention

[0004] The present invention solves the deficiencies of the prior art and provides an intelligent system and method for speech recognition based on communication scheduling instructions, which is characterized by comprising the following steps:

[0005] S01 Speech Training Corpus

[0006] The training corpus consists of two parts: open source speech corpus and communication dispatch instruction speech corpus;

[0007] Open source speech corpus; the communication dispatch corpus collects speech data of daily dispatch instructions of communication dispatch staff. The speech data contains professional terminology in the communication industry, or speech data with local characteristics. The communication dispatch instruction speech corpus is derived from the daily language of communication dispatch commands, and contains richer content on professional names than the open source corpus.

[0008] S02 Feature Extraction

[0009] Step S021: Increase the amplitude of the high-frequency component of the speech signal compared to the low-frequency component. The speech energy is mainly concentrated in the low frequency. Improving the high frequency helps to improve the signal-to-noise ratio and remove the influence of glottal excitation, oral and nasal radiation, and greater high-frequency attenuation during propagation. Suppose the speech sampling value at time n is x(n), and the result y(n) after pre-emphasis processing is:

[0010]

[0011] Step S022: divide the audio into frames, and divide the sound into many small segments. There is a certain overlap between frames. The length of each frame is 25ms, the frame shift is 10ms, and there is an overlap of 25-10=15ms between two frames.

[0012] Step S023: After the framing, it is usually necessary to perform windowing on the signal of each frame, and the window function is:

[0013]

[0014] Here 0 ≤ n ≤ N-1, N is the width of the window;

[0015] Step S024: For each frame of the windowed signal, perform an N-point FFT transform, also known as short-time Fourier transform (STFT), where N is usually 256 or 512.

[0016] Step S025: After the above steps, the Mel filter bank is applied to the energy spectrum to extract the FBank feature;

[0017] Step S026: removing the mean value to reduce the mismatch between the training set and the test set, equalize the spectrum, and improve the signal-to-noise ratio;

[0018] S03 Acoustic Model Training

[0019] 1) Constructing the model Constructing the acoustic model, which includes: deep neural network, hidden Markov model; in this model, HMM is used to describe the dynamic changes of speech signals, and each output node of DNN is used to estimate the posterior probability of a certain state of the continuous density HMM. The overall input of the model is the fbank feature, and then DNN models the likelihood of all clustered states to obtain the posterior probability. Then it is decoded in combination with HMM;

[0020] 2) Training model The main training steps of the DNN-HMM model are as follows:

[0021] ① First, train a state-sharing triphone GMM-HMM Chinese recognition system, use a decision tree to decide how to share the state, and assume that the trained system is gmm-hmm;

[0022] ② Use the gmm-hmm obtained in step ① to initialize a new hidden Markov model and generate a DNN-HMM model, assuming that the model is dnn-hmm1;

[0023] ③Pre-train the deep neural network in the dnn-hmm1 system, and the obtained deep neural network is ptdnn;

[0024] ④ Use the gmm-hmm system to arrange the speech training data, and the obtained data is set as align-raw;

[0025] ⑤Use the data obtained in step ④ to fine-tune the parameters of ptdnn, and set the obtained deep neural network as dnn;

[0026] ⑥ Use dnn and dnn-hmm1 and the maximum likelihood algorithm to re-estimate the parameters of the hidden Markov, and let the newly obtained system be dnn-hmm2;

[0027] ⑦ If the accuracy of step ⑥ is no longer improved, exit the algorithm, otherwise use dnn and dnn-hmm2 to generate new permutation data of speech training data, and then return to step ⑤;

[0028] ⑧ Estimate the probability value using training data;

[0029] S04 Acoustic Model

[0030] Through step S03: after the acoustic model is trained, an acoustic model suitable for the communication scheduling instruction voice is obtained;

[0031] S05 Dictionary

[0032] The open source dictionary DaCiDian adjusts the open source dictionary according to the direction of communication scheduling, covers as many words related to communication scheduling as possible, and discards unnecessary words to improve retrieval efficiency and recognition performance;

[0033] S06 Text training corpus

[0034] Organize the plain text data of sentences that conform to daily language grammar and sentences of communication call instructions into text training corpus; S07 language model training

[0035] 1) Model building A language model is built using the chain rule. The probability of each word can be calculated from the corpus. Assuming that sentence W is composed of a word sequence w1, w2, w3…wn, it can be expressed by the conditional probability formula:

[0036]

[0037] The N-gram model assumes that the probability distribution of each word depends only on the first few words that appear. In the N-gram model, the probability distribution of each word depends only on the first n-1 words.

[0038] 2) Training model The main training steps of the N-gram model are as follows:

[0039] ① Obtain the original text training corpus;

[0040] ② Segment the original corpus obtained in the previous step using the open source Chinese word segmentation library jieba, and customize the professional terms of communication scheduling in the word library to help improve the accuracy;

[0041] ③ Perform word frequency statistics on the word segmentation results obtained in the previous step, specifically using Spark RDD programming to perform word frequency statistics on the original corpus;

[0042] ④ Use Laplace smoothing to smooth the N-gram model to obtain the final N-gram model. Laplace Smoothing algorithm assumes that each word in a sentence is represented by Wi, where V represents the number of words in the sentence:

[0043]

[0044] S08 Language Model

[0045] Through step S08: after the language model is trained, an N-gram model is obtained;

[0046] S09 Voice to be analyzed

[0047] Input a communication dispatch instruction voice data;

[0048] S10 Speech Preprocessing

[0049] The deep residual shrinkage network with shared thresholds between channels performs noise reduction or redundant information processing on speech data. DRSN-CS is an improvement on the Deep Residual Network (ResNet), which introduces soft thresholding and automatically sets the threshold in the sub-network.

[0050] S11 Feature Extraction

[0051] The feature extraction in step S11 is the same as the feature extraction in step S02;

[0052] S12 Decode Search

[0053] Step S121: The input parameter is the FBank feature of the speech extracted in step S11, and a feature vector sequence is obtained;

[0054] Step S122: Given the relevant feature vector, the probability is output by the deep neural network DNN as the emission probability, and the phoneme sequence is marked by combining the hidden Markov model HMM and the viterbi algorithm;

[0055] Step S123: Generate a character string sequence according to the dictionary;

[0056] Step S124: Based on the string sequence obtained in the previous step, the N-gram model is used in combination with the dictionary to realize the phonetic-character conversion. For multiple conversion results, the Viterbi decoding algorithm is used to determine the path using the principle of dynamic programming to give the character sequence with the highest probability.

[0057] The protection of the entire dispatch instruction intelligent recognition process in the present invention. 1) The dispatch instruction data of the present invention is managed throughout the life cycle. 2) Voice preprocessing: Use Deep Residual Shrinkage Networks with Channel-shared Thresholds (DRSN-CS) to reduce noise or process redundant information in voice data. 3) Language model: Use open source dictionaries combined with custom professional dictionaries to improve accuracy during model training. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 The overall flow chart of speech recognition for communication scheduling instructions;

[0059] Figure 2 Schematic diagram of FBank features for extracting speech;

[0060] Figure 3 It is a model hierarchy diagram;

[0061] Figure 4 This is a schematic diagram of an open source dictionary;

[0062] Figure 5 The schematic diagram of the overall network structure of DRSN-CS;

[0063] Figure 6 Schematic diagram of the improved residual module;

[0064] Figure 7 Search schematic diagram for decoding;

[0065] Figure 8 The overall logic for output analysis is as follows;

[0066] Fig. 9 Flowchart for speech recognition of communication scheduling instructions. DETAILED DESCRIPTION

[0067] a) Overview of the overall process

[0068] The invention comprises the following parts: collecting communication dispatch instruction voice data to form corpus; preprocessing voice data; the voice data contains professional noun voices in the communication industry; extracting voice features respectively; training the features; and obtaining a voice recognition model suitable for communication intelligent dispatch.

[0069] The overall flow chart of speech recognition of communication scheduling instructions is as follows:

[0070] Detailed description of the nodes in flowchart 1:

[0071] i. Speech training corpus S01

[0072] The training corpus consists of two parts: the open source speech corpus and the communication dispatch instruction speech corpus. The open source speech corpus (aidatatang_1505zh); the communication dispatch corpus collects the speech data of daily dispatch instructions of communication dispatch staff, which contains professional terminology in the communication industry, or speech data with local characteristics (dialects). The communication dispatch instruction speech corpus comes from the daily language of communication dispatch commands (for example: close the switch, open the knife switch), which contains richer content of professional names than the open source corpus.

[0073] ii. Feature Extraction S02

[0074] Extract the FBank features of speech. The specific process is as follows:

[0075] Step S021: Increase the amplitude of the high-frequency component of the speech signal compared to the low-frequency component. The speech energy is mainly concentrated in the low frequency. Improving the high frequency helps to improve the signal-to-noise ratio and remove the effects of glottal excitation, oral and nasal radiation, and greater high-frequency attenuation during propagation.

[0076] Assume that the speech sampling value at time n is x(n), and the result y(n) after pre-emphasis processing is:

[0077]

[0078] Step S022: divide the audio into frames, and divide the sound into many small segments. There is a certain overlap between frames. As shown in FIG3 , the length of each frame is 25 ms, the frame shift is 10 ms, and there is an overlap of 25-10=15 ms between two frames.

[0079] Step S023: After the framing, it is usually necessary to perform windowing on the signal of each frame. The purpose is to make the two ends of the frame decay smoothly, so as to reduce the strength of the side lobes after the subsequent Fourier transform and obtain a higher quality spectrum. Here, the Hamming window is used, and its window function is:

[0080]

[0081] Here 0 ≤ n ≤ N-1, N is the width of the window.

[0082] Step S024: For each frame of the windowed signal, perform an N-point FFT transform, also known as short-time Fourier transform (STFT), where N is usually 256 or 512.

[0083] Step S025: After the above steps, the Mel filter bank is applied to the energy spectrum to extract the FBank features.

[0084] Step S026: removing the mean value reduces the mismatch between the training set and the test set, balances the spectrum, and improves the signal-to-noise ratio.

[0085] Acoustic model training S03

[0086] 1) Build the model

[0087] Construct an acoustic model (DNN-HMM model), which includes: Deep Neural Networks (DNN) and Hidden Markov Model (HMM).

[0088] In this model, HMM is used to describe the dynamic changes of speech signals, and each output node of DNN is used to estimate the posterior probability of a certain state of continuous density HMM. The overall input of this model is fbank features, and then DNN models the likelihood of all clustered states (such as triphones after state binding) to obtain the posterior probability. It is then decoded in combination with HMM.

[0089] Model hierarchy description:

[0090] 2) Training the model

[0091] The main training steps of the DNN-HMM model are as follows:

[0092] ① First, train a state-sharing triphone GMM-HMM Chinese recognition system and use a decision tree to decide how to share states. Let the trained system be gmm-hmm.

[0093] ② Use the gmm-hmm obtained in step ① to initialize a new hidden Markov model (including transition probability, observation probability, and state of the hidden Markov model), and generate a DNN-HMM model, assuming that the model is dnn-hmm1.

[0094] ③Pre-train the deep neural network in the dnn-hmm1 system, and the obtained deep neural network is ptdnn.

[0095] ④ Use the gmm-hmm system to arrange the speech training data (that is, find out which state in the hidden Markov model the training data corresponds to), and set the obtained data as align-raw.

[0096] ⑤Use the data obtained in step ④ to fine-tune the parameters of ptdnn (you can use the stochastic gradient descent algorithm). Let the obtained deep neural network be dnn.

[0097] ⑥ Use dnn, dnn-hmm1 and maximum likelihood algorithm to re-estimate the parameters (transition probability, observation probability) in the hidden Markov, and let the newly obtained system be dnn-hmm2.

[0098] ⑦ If the accuracy of step ⑥ is no longer improved, exit the algorithm, otherwise use dnn and dnn-hmm2 to generate new permutation data of speech training data, and then return to step ⑤.

[0099] ⑧Use training data to estimate the probability value.

[0100] iii. Acoustic Model (AM) S04

[0101] Through step S03: after acoustic model training, an acoustic model (DNN-HMM model) suitable for communication scheduling instruction speech is obtained.

[0102] iv. Dictionary (Lexicon) S05

[0103] Open source dictionary DaCiDian. The communication scheduling dictionary counts the words used in communication scheduling instructions, and annotates the pinyin to form a one-to-one corresponding data set. In Chinese, it is the correspondence between pinyin and Chinese characters. Its purpose is to find the corresponding Chinese characters (words) or words based on the phonemes recognized by the acoustic model, which is used to build a bridge between the acoustic model and the language model to connect the two. The function of the pronunciation dictionary is to specify the pronunciation rules of words and get words from phonemes. According to the adjustment of the open source dictionary in the direction of communication scheduling, as many words related to communication scheduling as possible are covered, and unnecessary words are discarded to improve retrieval efficiency and recognition performance.

[0104] The dictionary example is as follows:

[0105] v. Text training corpus S06

[0106] The sentence (plain text) data that conforms to the grammar of daily language and the sentences of communication call instructions are organized into text training corpus.

[0107] vi. Language model training S07

[0108] 1) Build the model

[0109] Construct a language model (N-gram model). The language model in speech recognition is also used to process text sequences. It combines the output of the acoustic model. The probability of the entire sentence is equal to the product of the probabilities of each word. The text sequence with the highest probability is used as the speech recognition result. Since the language model represents the probability of a certain text sequence, it is generally expressed using the chain rule. The probability of each word can be obtained by statistical calculation in the corpus. Assuming that the sentence W is composed of the word sequence w1, w2, w3…wn, it can be expressed by the conditional probability related formula as follows:

[0110]

[0111] The N-gram model assumes that the probability distribution of each word depends only on the first few words that appear. In the N-gram model, the probability distribution of each word depends only on the first n-1 words.

[0112] 2) Training model The main training steps of the N-gram model are as follows:

[0113] ① Obtain the original text training corpus;

[0114] ② Segment the original corpus obtained in the previous step using the open source Chinese word segmentation library Jieba, and customize the professional terms of communication scheduling in the word library to help improve the accuracy;

[0115] ③ Perform word frequency statistics on the word segmentation results obtained in the previous step, specifically using Spark RDD programming to perform word frequency statistics on the original corpus;

[0116] ④ Use Laplace Smoothing to smooth the N-gram model to obtain the final N-gram model. Laplace Smoothing algorithm, assuming that each word in a sentence is represented by Wi, where V represents the number of words in the sentence:

[0117]

[0118] vii. Language Model (LM) S08

[0119] Through step S08: after the language model is trained, an N-gram model is obtained.

[0120] viii. Voice to be analyzed S09

[0121] Input a section of communication dispatch instruction voice data.

[0122] ix. Speech preprocessing S10

[0123] In a speech data, any information irrelevant to the current pattern recognition task can be considered as noise or redundant information. These noises or redundant information are likely to have an adverse effect on the current pattern recognition task. For any two speech data, their noise or redundant content is often different. To handle the above situation, the present invention uses: Deep Residual Shrinkage Networks with Channel-shared Thresholds (DRSN-CS for short) to reduce noise or process redundant information on speech data.

[0124] DRSN-CS is an improvement of Deep Residual Network (ResNet), which introduces soft thresholding and automatically sets the threshold in the sub-network.

[0125] Figure 5 below shows the overall network structure of DRSN-CS. C represents the number of channels of the feature map, W represents the width of the feature map, and 1 represents that the height of the feature map is always 1. In the brackets after Conv, K represents the number of convolution kernels in the convolution layer. When K=C, the number of channels of the output feature map is C. When K=2C, the number of channels of the output feature map is 2C. / 2 means that the step length of each movement of the convolution kernel is 2, so that the width of the output feature map is halved.

[0126] Figure 6 below shows an improved residual module named "Residual Shrinkage Building Unit with Channel-shared thresholds (RSBU-CS)". M represents the number of neurons in the fully connected layer. M=C means that the number of neurons is equal to the number of channels in the previous feature map, and M=1 means that the number of neurons is 1. Absolute means taking the absolute value of each element in the feature map, and Average means finding the average value of all elements in the feature map.

[0127] x. Feature extraction S11

[0128] This part of feature extraction is the same as step: feature extraction S02 in Figure 1.

[0129] xi. Decode Search S12

[0130] This part searches for decoding and searching algorithms (obtaining the best text output based on acoustic models, dictionaries, and language models). The specific steps are as follows:

[0131] Step S121: The input parameter is the FBank feature of the speech extracted in step S11 in FIG. 1, and a feature vector sequence is obtained.

[0132] Step S122: Given the relevant feature vector, the probability is output through the deep neural network (DNN) as the emission probability, and the phoneme sequence is marked by combining the hidden Markov model (HMM) and the Viterbi algorithm.

[0133] Step S123: Generate a character string sequence according to the dictionary. For example: kun ming gong dian ju tongxin jifang

[0134] Step S124: Based on the string sequence obtained in the previous step, the language model (N-gram model) is used in combination with the dictionary to realize the phonetic-character conversion. For multiple conversion results, the Viterbi decoding algorithm is used to quickly determine the most appropriate path using the principle of dynamic programming to give the text sequence with the highest probability. Viterbi algorithm pruning optimization steps:

[0135] Step 1: Initialize parameters

[0136]

[0137] Step 2: Recursively, for t=2, 3, ..., T; record the maximum probability as δ t(i); record the previous state j of the path with maximum probability as ψ t(i).

[0138]

[0139] Step 3: Calculate the terminal state, calculate the maximum δ T(i) at time T as P*, which is the probability of the most likely hidden state sequence. Calculate the maximum ψ T(i) at time T as i*T, which is the most likely hidden state at time T.

[0140]

[0141] Step 4: i*t is the backtracking of the optimal path, for t=T-1, T-2, ..., 1; I* is the optimal path.

[0142]

[0143] Step S125: Output the analyzed optimal text.

[0144] Text Output S13

[0145] Output the converted text data.

[0146] The beneficial effects of the present invention are that the solution greatly improves the actual on-site operation and maintenance efficiency, improves the safety of on-site communication scheduling work; reduces human intervention in the scheduling process to improve efficiency; and utilizes new technologies to improve the accuracy of speech recognition.

[0147] b) The communication dispatch instruction voice recognition process is as follows:

[0148] Step A1: Communication dispatch staff A issues dispatch voice instructions through the communication telephone system (for example: close the switch, open the knife switch, change the protection to tripping, and disable the protection).

[0149] Step A2: Communication dispatch staff B receives the voice instruction.

[0150] Step A3: The voice recognition system obtains the voice command, and finally outputs text data through the above process steps S9->S10->S11->S04->S08->S12->S13.

[0151] Step A4: Analyze and obtain the dispatch text instruction corresponding to the voice, and store it in the instruction system.

[0152] Step A5: After receiving the voice command, communication dispatch staff B confirms it with the dispatch text command of the command system and then performs specific operations.

Claims

1. An intelligent system and method for speech recognition based on communication scheduling instructions, characterized in that The steps include: S01 speech training corpus The training corpus consists of two parts: open source speech corpus and communication dispatch instruction speech corpus; Open source speech corpus; the communication dispatch instruction speech corpus collects speech data of daily dispatch instructions of communication dispatch staff. The speech data contains professional terminology in the communication industry, or speech data with local characteristics. The communication dispatch instruction speech corpus is derived from the daily language of communication dispatch commands, and contains richer content on professional names than the open source corpus. S02 Feature Extraction Step S021: Increase the amplitude of the high-frequency component of the speech signal compared to the low-frequency component. The speech energy is mainly concentrated in the low frequency. Improving the high frequency helps to improve the signal-to-noise ratio and remove the influence of glottal excitation, oral and nasal radiation, and greater high-frequency attenuation during propagation. Suppose the speech sampling value at time n is x(n), and the result y(n) after pre-emphasis processing is: Step S022: Frame the audio and divide the audio into many small segments. There is a certain overlap between frames. The length of each frame is 25ms, the frame shift is 10ms, and there is an overlap of 25-10=15ms between two frames. Step S023: After the framing, it is usually necessary to perform windowing on the signal of each frame, and the window function is: Here 0≤n≤N-1, N is the width of the window; Step S024: For each frame of the windowed signal, perform an N-point FFT transform, also known as short-time Fourier transform (STFT), where N is usually 256 or 512; Step S025: After the above steps, the Mel filter bank is applied to the energy spectrum to extract the FBank feature; Step S026: removing the mean value to reduce the mismatch between the training set and the test set, equalize the spectrum, and improve the signal-to-noise ratio; S03 Acoustic Model Training 1) Build the model Construct an acoustic model, which includes: deep neural network and hidden Markov model; in this model, HMM is used to describe the dynamic changes of speech signals, and each output node of DNN is used to estimate the posterior probability of a certain state of continuous density HMM. The overall input of this model is fbank features, and then DNN models the likelihood of all clustered states to obtain the posterior probability, which is then decoded in combination with HMM; 2) Training the model The main training steps of the DNN-HMM model are as follows: ① First, train a state-sharing triphone GMM-HMM Chinese recognition system, use a decision tree to decide how to share the state, and assume that the trained system is gmm-hmm; ② Use the gmm-hmm obtained in step ① to initialize a new hidden Markov model and generate a DNN-HMM model, assuming that the model is dnn-hmm1; ③Pre-train the deep neural network in the dnn-hmm1 system, and the obtained deep neural network is ptdnn; ④ Use the gmm-hmm system to arrange the speech training data, and the obtained data is set as align-raw; ⑤Use the data obtained in step ④ to fine-tune the parameters of ptdnn, and set the obtained deep neural network as dnn; ⑥ Use dnn and dnn-hmm1 and the maximum likelihood algorithm to re-estimate the parameters of the hidden Markov, and let the newly obtained system be dnn-hmm2; ⑦ If the accuracy of step ⑥ is no longer improved, exit the algorithm, otherwise use dnn and dnn-hmm2 to generate new permutation data of speech training data, and then return to step ⑤; ⑧ Estimate the probability value using training data; S04 Acoustic Model Through step S03: after the acoustic model is trained, an acoustic model suitable for the communication scheduling instruction voice is obtained; S05 Dictionary The open source dictionary DaCiDian adjusts the open source dictionary according to the direction of communication scheduling, covers as many words related to communication scheduling as possible, and discards unnecessary words to improve retrieval efficiency and recognition performance; S06 text training corpus Organize the plain text data of sentences that conform to the grammar of daily language and the sentences of communication call instructions into text training corpus; S07 Language Model Training 1) Build the model The language model is constructed using the chain rule. The probability of each word can be obtained by statistical calculation in the corpus. Assuming that the sentence W is composed of the word sequence w1, w2, w3…wn, it can be expressed by the conditional probability related formula as follows: P(W)=P(w1)*p(w2)*p(w3)***p(wn) =p(w1)*p(w2|w1)*p(w3|w1w2)***p(wn|w1w2w3…wn) The N-gram model assumes that the probability distribution of each word depends only on the first few words that appear. In the N-gram model, the probability distribution of each word depends only on the first n-1 words. 2) Training the model The main training steps of the N-gram model are as follows: ① Obtain the original text training corpus; ② Segment the original corpus obtained in the previous step using the open source Chinese word segmentation library jieba, and customize the professional terms of communication scheduling in the word library to help improve the accuracy; ③ Perform word frequency statistics on the word segmentation results obtained in the previous step, specifically using Spark RDD programming to perform word frequency statistics on the original corpus; ④ Use Laplace smoothing to smooth the N-gram model to obtain the final N-gram model. LaplaceSmoothing algorithm assumes that each word in a sentence is represented by Wi, where V represents the number of words in the sentence: S08 Language Model Through step S08: after the language model is trained, an N-gram model is obtained; S09 Voice to be analyzed Input a communication dispatch instruction voice data; S10 Voice Preprocessing The deep residual shrinkage network with shared thresholds between channels performs noise reduction or redundant information processing on speech data; DRSN-CS is an improvement of Deep Residual Network (ResNet), which introduces soft thresholding and automatically sets the threshold in the sub-network. S11 feature extraction Step S11 feature extraction is the same as feature extraction S02; S12 Decode Search Step S121: The input parameter is the FBank feature of the speech extracted in step S11, and a feature vector sequence is obtained; Step S122: Given the relevant feature vector, the probability is output by the deep neural network DNN as the emission probability, and the phoneme sequence is marked by combining the hidden Markov model HMM and the viterbi algorithm; Step S123: Generate a character string sequence according to the dictionary; Step S124: Based on the string sequence obtained in the previous step, the N-gram model is used in combination with a dictionary to realize the phonetic-character conversion. For multiple conversion results, the Viterbi decoding algorithm is used, and the path is determined by the principle of dynamic programming to give the text sequence with the highest probability.

Images