A speech recognition optimization method and system based on adaptive dynamic programming

Abstract

This invention provides a speech recognition optimization method and system based on adaptive dynamic programming, belonging to the field of speech recognition technology. The method first acquires speech data and preprocesses it to construct an acoustic feature sequence. Then, it constructs an acoustic state transition model, modeling the speech recognition process as a finite-time optimal control problem, determining the optimization objective and policy distribution. Next, it trains a Conformer speech recognition model in two stages to learn and optimize the policy distribution. The first stage uses a connection-time classification (CTC) framework for pre-training, and the second stage introduces an adaptive dynamic programming (ADP) algorithm and a self-critical sequence training method for training. Finally, the trained Conformer speech recognition model is used for speech recognition, outputting the recognition result. This invention's method is simple to implement, highly compatible with existing ASR architectures, and significantly reduces the word error rate while ensuring stability.

Description

Technical Field

[0001] This invention relates to the field of speech recognition technology, and in particular to a speech recognition optimization method and system based on adaptive dynamic programming. Background Technology

[0002] Automatic speech recognition is used to transcribe speech signals into text, and it is widely used in human-computer interaction, voice input, and intelligent assistants. In engineering, recognition accuracy, such as word error rate (WER) or character error rate (CER), is usually used as an evaluation metric. In speech recognition methods within the deep learning framework, Transformer and Conformer models based on self-attention have become the backbone.

[0003] The Transformer and Conformer speech recognition methods based on self-attention mechanisms are currently the most representative end-to-end models in automatic speech recognition systems. Their basic process is as follows: First, continuous speech signals are sampled and framed, and acoustic feature sequences such as log-Mel spectrum are extracted as input to the Transformer model. Next, the feature sequences are deeply represented by an encoder. The Transformer model's encoder consists of multiple layers of self-attention and feedforward networks, capable of capturing long-distance dependencies across the entire speech segment, thus better understanding the global semantic structure of the speech. In the decoding stage, the model generates text progressively using an autoregressive approach. The decoder ensures temporal consistency of the generated sequence by masking self-attention and guides the output by reading the encoder's representation through cross-attention. To enhance training stability, a common practice is to add a Connected Temporal Classification (CTC) auxiliary branch in addition to the backbone cross-entropy loss, achieving multi-task joint optimization to obtain better alignment and convergence speed.

[0004] The Conformer model further introduces a convolutional module, adding depthwise separable convolutions between the self-attention and feedforward networks to capture local temporal patterns in speech. This gives the Conformer model both the sensitivity of convolution to local structures and the modeling ability of self-attention to global dependencies, resulting in more robust performance under conditions of long speech, noisy environments, and diverse accents. During the inference phase, both types of models typically generate multiple candidate transcriptions through beam search, and then sort and filter them using an external language model or regularization strategy, ultimately outputting the optimal text.

[0005] Existing self-attention-based end-to-end speech recognition models have two limitations. First, the training objective and evaluation metric are inconsistent. Training typically optimizes cross-entropy or connection-temporal classification loss, while actual evaluation relies on word error rate or character error rate calculated from the whole sentence edit distance. The difference between these two means that a decrease in training error rate does not guarantee a synchronous decrease in recognition error rate. Especially in long sentences, small errors in the early stages can cause overall alignment shifts, and existing objective functions cannot effectively characterize this global amplification effect.

[0006] Secondly, there is a difference between teacher-mandated training and free inference. During training, the model always relies on true prefixes for prediction, while inference must rely on its own historical output. Once an error occurs during inference, the model enters an "error context" not covered in training, which can easily lead to error accumulation and a sharp decline in performance. This exposure bias is a common stability problem faced by end-to-end speech recognition. Summary of the Invention

[0007] The technical problem to be solved by the present invention is to address the shortcomings of the prior art by providing a speech recognition optimization method and system based on adaptive dynamic programming to achieve speech recognition.

[0008] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows:

[0009] On one hand, the present invention provides a speech recognition optimization method based on adaptive dynamic programming, comprising:

[0010] Acquire voice data and construct validation and test datasets;

[0011] The acquired speech data is preprocessed to construct an acoustic feature sequence;

[0012] An acoustic state transition model is constructed to model the speech recognition process as a finite-time optimal control problem, and the optimization objective and policy distribution are determined.

[0013] The Conformer speech recognition model is trained in two stages to learn and optimize the policy distribution. In the first stage, the Connectionist Temporal Classification (CTC) framework is used to pre-train the Conformer speech recognition model. In the second stage, the Adaptive Dynamic Programming (ADP) algorithm and the self-critic sequence training method are introduced to train the Conformer speech recognition model in the second stage.

[0014] The trained Conformer speech recognition model is used for speech recognition, and the recognition results are output.

[0015] Furthermore, the method extracts the Mel frequency cepstral coefficients of the speech data as acoustic feature parameters of the speech data to construct an acoustic feature sequence.

[0016] Furthermore, the specific method for constructing the acoustic state transition model is as follows:

[0017] Viewing the speech recognition process as a discrete-time, time-varying dynamic system, the state of the system is defined as... ,in This represents a complete sequence of acoustic features. This represents the historical speech symbols generated up to time t;

[0018] The action is defined as selecting the next symbol. , For the letters or subwords identified at time t+1, the system transitions to a new state after the action is performed. ;

[0019] Define sequence-level reward function That is, the sequence-level reward function is defined as the negative value of the edit distance between the predicted sequence and the reference text, where, Y* represents the predicted sequence of characters recognized by the speech recognition model, Y* represents the reference text consisting of correctly labeled characters in the dataset, and ED represents the edit distance, which indicates the minimum number of operation steps required to convert the predicted sequence into the reference text.

[0020] The optimization objective is to maximize the initial state. Expected reward:

[0021] ;

[0022] in, Represents a given sequence of acoustic features At that time, generate The probability of; , Represents the policy distribution, that is, in the state Next, select the probability of generating the next symbol.

[0023] Furthermore, the Conformer speech recognition model is pre-trained using the Connectionist Temporal Classification (CTC) framework, specifically as follows:

[0024] The predicted sequence is obtained by folding the alignment path. By introducing the whitespace symbol method, the alignment problem in speech recognition is transformed into the sum of probabilities of all possible paths. In the recognition, "selecting one letter in each frame from the first frame to the last frame" is regarded as an alignment path Π. Repeated characters and whitespace characters in the path are removed to obtain the final recognized text.

[0025] The training objective is to minimize the negative log-likelihood, where log-likelihood is the total probability of all paths generating correct text.

[0026] ;

[0027] in, The training objective is to use the Connectionist Temporal Classification (CTC) framework to pre-train the Conformer speech recognition model. This indicates that all text that has been folded is for reference. The set of paths It is the conditional probability of path Π given an input acoustic feature sequence X.

[0028] Furthermore, the specific method for the second-stage training of the Conformer speech recognition model using the Adaptive Dynamic Programming (ADP) algorithm and the self-criticism sequence training method is as follows:

[0029] During this training phase, the Adaptive Dynamic Programming (ADP) algorithm is based on the Bellman equations in the finite-time domain:

[0030] ;

[0031] in, It is the optimal state value function, representing the maximum expected reward from the current moment to the end; Let P(u) represent the expectation function. t |s t ) indicates that in state s t Take action u t The probability distribution, i.e., the policy distribution; t represents the time step. Indicates an immediate reward;

[0032] Solving for the optimal state value function This leads to the optimal strategy distribution of the acoustic state transition model;

[0033] Policy gradient optimization is performed using self-critical sequence training. Specifically, beam search is used to generate N candidate sequences from the Conformer speech recognition model. The expected gradient is estimated using Monte Carlo policy gradients on these candidate sequences, and a self-generated baseline reward is used as a baseline to reduce variance. Based on this, the following gradient approximation is obtained:

[0034] ;

[0035] in, To optimize the target The gradient approximation, The baseline reward is used to select the reward for finding the optimal path in the beam search, which is used to reduce variance. This represents the nth candidate prediction sequence; N is the number of candidate text sequences generated in one beam search. The weights of the Conformer speech recognition model, for The gradient.

[0036] Furthermore, the training of the conformer speech recognition model employs a joint loss function:

[0037] , ;

[0038] in, For the joint loss function, This represents the training loss based on reward optimization. The training loss is used when employing the CTC framework. These are the weighting coefficients.

[0039] Furthermore, the method evaluates the recognition performance of the conformer speech recognition model using word error rate (WER) or edit distance (ED):

[0040] Word error rate ;

[0041] Edit distance ;

[0042] in, These represent the number of errors where one character needs to be replaced, deleted, or inserted between the recognized text and the correct text, respectively, and N is the number of words transcribed from the reference text.

[0043] On the other hand, the present invention also provides a speech recognition optimization system based on adaptive dynamic programming, including a speech acquisition module, a preprocessing module, a process modeling module, a training module and a recognition module;

[0044] The voice acquisition module is used to acquire voice data and construct verification datasets and test datasets;

[0045] The preprocessing module is used to preprocess the acquired speech data and construct an acoustic feature sequence;

[0046] The process modeling module is used to construct an acoustic state transition model, which models the speech recognition process as a finite-time optimal control problem and determines the optimization objective and policy distribution.

[0047] The training module trains the Conformer speech recognition model in two stages to learn and optimize the policy distribution. In the first stage, the Conformer speech recognition model is pre-trained using the Connectionist Temporal Classification (CTC) framework. In the second stage, the Conformer speech recognition model is trained using the Adaptive Dynamic Programming (ADP) algorithm and the self-critic sequence training method.

[0048] The recognition module uses the trained Conformer speech recognition model to perform speech recognition and outputs the recognition results.

[0049] Thirdly, this application proposes a computer-readable storage medium storing executable instructions that, when executed, cause a processor to perform the aforementioned adaptive dynamic programming-based speech recognition optimization method.

[0050] Fourthly, this application proposes a computer program product, including a computer program or instructions that, when executed by a processor, implement the aforementioned speech recognition optimization method based on adaptive dynamic programming.

[0051] The beneficial effects of adopting the above technical solution are as follows: The present invention provides a speech recognition optimization method and system based on adaptive dynamic programming, which constructs a time-varying acoustic state transition model containing acoustic features and historical speech symbols, and implements policy iteration specifically optimized for speech recognition (automatic speech recognition, i.e., ASR) tasks within the ADP framework; it proposes a novel ADP algorithm based on the Monte Carlo policy gradient method (REINFORCE), which introduces a sequence-level edit distance reward function, takes edit distance as the sole optimization objective, effectively bridges the gap between training and inference, and avoids the mismatch between training objectives and evaluation metrics and strong teacher dependence.

[0052] The method of this invention is simple to implement and highly compatible with existing ASR architectures. It significantly reduces word error rate while ensuring stability, thus meeting the dual requirements of recognition accuracy and engineering usability. Attached Figure Description

[0053] Figure 1 A flowchart of a speech recognition optimization method based on adaptive dynamic programming provided in Embodiment 1 of the present invention;

[0054] Figure 2 This is a comparison chart of word error rates for the Conformer model provided in Embodiment 1 of the present invention, whether or not the ADP algorithm is used. Detailed Implementation

[0055] The specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples. The following examples are for illustrative purposes only and are not intended to limit the scope of the invention.

[0056] In this embodiment, a speech recognition optimization method based on adaptive dynamic programming is described, such as... Figure 1 As shown, it includes the following steps:

[0057] Step 1: Acquire voice data;

[0058] This embodiment selects a 100-hour clean speech subset from the LibriSpeech dataset for verification, a dev-clean dataset for testing, and a test-clean subset for testing.

[0059] Step 2: Preprocess the acquired speech data, extract the Mel frequency cepstral coefficients of the speech data as the acoustic feature parameters of the speech data, and construct the acoustic feature sequence;

[0060] In this embodiment, the acoustic feature sequence of the speech data after feature extraction is represented as follows: ,in, The acoustic features of frame T are given.

[0061] Step 3: Construct an acoustic state transition model to model the speech recognition process as a finite-time optimal control problem;

[0062] The speech recognition process (inputting a speech feature sequence X and progressively generating a symbol sequence) can be viewed as a discrete-time, time-varying dynamic system, with the system's state defined as follows: ,in This represents a complete sequence of acoustic features. Represents the historical speech symbols (letters or subwords) generated up to time t.

[0063] The action is defined as selecting the next symbol. , For the letters or subwords identified at time t+1, the system transitions to a new state after the action is performed. ;

[0064] To measure recognition quality, a sequence-level reward function is defined. That is, the sequence-level reward function is defined as the negative value of the edit distance between the predicted sequence and the reference text, where, Y* represents the predicted sequence of characters recognized by the speech recognition algorithm, Y* represents the reference text consisting of correctly labeled characters in the dataset, and ED represents the edit distance, which indicates the minimum number of operation steps required to convert the predicted sequence into the reference text (two texts) so that they are identical. The operation includes replacing a character, deleting a character, and inserting a character. The larger the reward, the closer the predicted sequence is to the transcribed reference text.

[0065] Therefore, the optimization objective is to maximize the initial state. Expected reward: ,in, Represents a given sequence of acoustic features At that time, generate The probability of; , Represents the policy distribution, that is, in the state Next, we choose the probability of generating the next symbol. Therefore, maximizing... The essence is to optimize the policy distribution. To make the recognition result the same as Predicted sequences that are closer and have a smaller edit distance.

[0066] Step 4: Train the Conformer speech recognition model in two stages. In the first stage, the Conformer speech recognition model is pre-trained using the Connectionist Temporal Classification (CTC) framework. In the second stage, the Conformer speech recognition model is trained using the Adaptive Dynamic Programming (ADP) algorithm and the self-criticism sequence training method.

[0067] Directly using sequence edit distance as a reward (with feedback only at the endpoint) leads to problems such as signal sparsity, large variance, and difficulty in credit allocation. Connectionist Temporal Classification (CTC) can significantly reduce this instability. Therefore, this invention learns and optimizes the policy distribution by performing two-stage training on the conformer speech recognition model. This improves performance in the initial state. Expected reward .

[0068] First, the Conformer speech recognition model is pre-trained using the Connectionist Temporal Classification (CTC) framework. The predicted sequence can be obtained by folding the alignment path (merging repeated characters and removing whitespace). Therefore, by introducing whitespace, the alignment problem in speech recognition is transformed into summing the probabilities of all possible paths. In speech recognition, the alignment problem refers to treating "selecting one letter in each frame from the first to the last" as an alignment path Π. Repeated characters and whitespace characters are removed from the path to obtain the final recognized text. Therefore, the same recognized text often corresponds to many possible paths Π. The training objective is to minimize the negative log-likelihood, which is the total probability of all paths generating the correct text.

[0069] ;

[0070] in, This indicates that all text that has been folded is for reference. The set of paths Given input X, path Π is the conditional probability. Training the conformer speech recognition model with CTC framework loss ensures that the model can learn the basic correspondence between speech and text, but its goal is not entirely consistent with the final evaluation metric WER.

[0071] To address this, based on the pre-training of the Conformer speech recognition model within the CTC framework, an Adaptive Dynamic Programming (ADP) algorithm and a self-critical sequence training method are introduced to perform a second-stage training of the Conformer speech recognition model, thereby directly optimizing the model's sequence-level performance.

[0072] During this training phase, the Adaptive Dynamic Programming (ADP) algorithm is based on the Bellman equations in the finite-time domain:

[0073] ;

[0074] in, It is the optimal state value function, representing the maximum expected reward from the current moment to the end; P(u) represents the expected function; t |s t ) indicates that in state s t Take action u t The probability distribution, i.e., the policy distribution); t represents the time step, Indicates an immediate reward;

[0075] Solving for the optimal state value function This allows the optimal strategy to be induced. In the ADP algorithm, the best recognition result is obtained by... The derived unknowns are not directly solvable using the Bellman equation, but rather optimized using approximations (policy gradient + beam sampling), making the implicit... This is better, thus allowing for a smaller edit distance in the predicted sequence.

[0076] To concentrate the probability quality towards recognition results with smaller edit distances (higher rewards), thereby directly increasing the optimization objective J in step 3, this invention employs self-critical sequence training for policy gradient optimization. Specifically, it uses beam search to generate N candidate sequences from the Conformer speech recognition model, estimates the expected gradient on these candidate sequences using Monte Carlo policy gradients, and uses a self-generated baseline reward as a baseline to reduce variance. Based on this, the following gradient approximation is obtained:

[0077] ;

[0078] in, To optimize the target The gradient approximation, The baseline reward is typically taken as the reward for searching the optimal path in the beam search, in order to reduce variance. This represents the nth candidate prediction sequence; N is the number of candidate text sequences generated in one beam search. The weights of the Conformer speech recognition model, for The gradient;

[0079] Finally, to balance training stability and optimal performance, the Conformer speech recognition model employs a joint loss function during training:

[0080] , ;

[0081] in, For the joint loss function, This represents the loss based on reward optimization. For CTC loss, it acts as a regularization term. Weighting coefficients (usually a small value, such as 0.001) are used to prevent training instability of the model during sequence optimization.

[0082] Step 5: Use the trained Conformer speech recognition model to perform speech recognition and output the recognition results; use the trained and optimized Conformer speech recognition model to recognize the speech data in the validation set and directly output the text transcription as the final result.

[0083] In this embodiment, the recognition performance of the Conformer speech recognition model is evaluated using the word error rate (WER) or edit distance (ED):

[0084] Word error rate ;

[0085] Edit distance ;

[0086] in, These represent the number of errors where one character needs to be replaced, deleted, or inserted between the recognized text and the correct text, respectively, and N is the number of words transcribed from the reference text.

[0087] In this embodiment, based on the LibriSpeech dataset, the word error rate of the Conformer speech recognition model is compared to whether it uses the ADP algorithm for speech recognition. Figure 2As shown in the figure, the blue curve represents the speech recognition result of the Conformer speech recognition model without using the ADP algorithm, with a word error rate of 24.4%. The red curve represents the speech recognition result of the Conformer speech recognition model after using the ADP algorithm, with a word error rate of 18.9%. It can be seen that using the ADP algorithm can significantly improve the recognition performance of the Conformer speech recognition model.

[0088] Example 2:

[0089] This embodiment provides a speech recognition optimization system based on adaptive dynamic programming, including a speech acquisition module, a preprocessing module, a process modeling module, a training module, and a recognition module;

[0090] The voice acquisition module is used to acquire voice data and construct verification datasets and test datasets;

[0091] The preprocessing module is used to preprocess the acquired speech data and construct an acoustic feature sequence;

[0092] The process modeling module is used to construct an acoustic state transition model, which models the speech recognition process as a finite-time optimal control problem and determines the optimization objective and policy distribution.

[0093] The training module trains the Conformer speech recognition model in two stages to learn and optimize the policy distribution. In the first stage, the Conformer speech recognition model is pre-trained using the Connectionist Temporal Classification (CTC) framework. In the second stage, the Conformer speech recognition model is trained using the Adaptive Dynamic Programming (ADP) algorithm and the self-critic sequence training method.

[0094] The recognition module uses the trained Conformer speech recognition model to perform speech recognition and outputs the recognition results.

[0095] Example 3:

[0096] This embodiment proposes a computer-readable storage medium that stores executable instructions. When these instructions are executed, if they are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium.

[0097] The computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the adaptive dynamic programming-based speech recognition optimization method described in the various embodiments of this application.

[0098] The aforementioned storage media include: flash memory, hard disks, multimedia cards, card-type memory (e.g., SD (Secure Digital Memory Card) or DX (Memory Data Register, MDR) memory), random access memory (RAM), static random-access memory (SRAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), programmable read-only memory (PROM), magnetic storage, disks, optical discs, servers, APP (Application) application stores, and other media capable of storing program verification codes. These media store computer programs, which, when executed by a processor, can implement the various steps of the aforementioned adaptive dynamic programming-based speech recognition optimization method.

[0099] Example 4:

[0100] This embodiment proposes a computer program product, including a computer program or instructions, which, when executed by a processor, implements the aforementioned speech recognition optimization method based on adaptive dynamic programming.

[0101] Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or part of the technical solution, can be embodied in the form of a computer program product.

[0102] The various embodiments in this application are described in a progressive manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.

[0103] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope defined by the present invention.

Claims

1. A speech recognition optimization method based on adaptive dynamic programming, characterized in that, include: Acquire voice data and construct validation and test datasets; The acquired speech data is preprocessed to construct an acoustic feature sequence; An acoustic state transition model is constructed to model the speech recognition process as a finite-time optimal control problem, and the optimization objective and policy distribution are determined. The Conformer speech recognition model is trained in two stages to learn and optimize the policy distribution. In the first stage, the Connectionist Temporal Classification (CTC) framework is used to pre-train the Conformer speech recognition model. In the second stage, the Adaptive Dynamic Programming (ADP) algorithm and the self-critic sequence training method are introduced to train the Conformer speech recognition model in the second stage. The trained Conformer speech recognition model is used for speech recognition, and the recognition results are output.

2. The speech recognition optimization method based on adaptive dynamic programming according to claim 1, characterized in that, The method extracts the Mel frequency cepstral coefficients of the speech data as acoustic feature parameters of the speech data, and constructs an acoustic feature sequence.

3. The speech recognition optimization method based on adaptive dynamic programming according to claim 1, characterized in that, The specific method for constructing the acoustic state transition model is as follows: Viewing the speech recognition process as a discrete-time, time-varying dynamic system, the state of the system is defined as... ,in This represents a complete sequence of acoustic features. This represents the historical speech symbols generated up to time t; The action is defined as selecting the next symbol. , For the letters or subwords identified at time t+1, the system transitions to a new state after the action is performed. ; Define sequence-level reward function That is, the sequence-level reward function is defined as the negative value of the edit distance between the predicted sequence and the reference text, where, Y* represents the predicted sequence of characters recognized by the speech recognition model, Y* represents the reference text consisting of correctly labeled characters in the dataset, and ED represents the edit distance, which indicates the minimum number of operation steps required to convert the predicted sequence into the reference text. The optimization objective is to maximize the initial state. Expected reward: ； in, Represents a given sequence of acoustic features At that time, generate The probability of; , Represents the policy distribution, that is, in the state Next, select the probability of generating the next symbol.

4. The speech recognition optimization method based on adaptive dynamic programming according to claim 3, characterized in that, The Conformer speech recognition model is pre-trained using the Connectionist Temporal Classification (CTC) framework, specifically as follows: The predicted sequence is obtained by folding the alignment path. By introducing the whitespace symbol method, the alignment problem in speech recognition is transformed into the sum of probabilities of all possible paths. In recognition, "selecting one letter in each frame from the first frame to the last frame" is regarded as an alignment path Π. Repeated characters and whitespace characters in the path are removed to obtain the final recognized text. The training objective is to minimize the negative log-likelihood, where log-likelihood is the total probability of all paths generating correct text. ； in, The training objective is to use the Connectionist Temporal Classification (CTC) framework to pre-train the Conformer speech recognition model. This indicates that all text that has been folded is for reference. The set of paths It is the conditional probability of path Π given an input acoustic feature sequence X.

5. The speech recognition optimization method based on adaptive dynamic programming according to claim 4, characterized in that, The specific method for the second-stage training of the Conformer speech recognition model by introducing the Adaptive Dynamic Programming (ADP) algorithm and the self-critic sequence training method is as follows: During this training phase, the Adaptive Dynamic Programming (ADP) algorithm is based on the Bellman equations in the finite-time domain: ； in, It is the optimal state value function, representing the maximum expected reward from the current moment to the end; Let P(u) represent the expectation function. t |s t ) indicates that in state s t Take action u t The probability distribution, i.e., the policy distribution; t represents the time step. Indicates an immediate reward; Solving for the optimal state value function This leads to the optimal strategy distribution of the acoustic state transition model; Policy gradient optimization is performed using self-critical sequence training. Specifically, beam search is used to generate N candidate sequences from the Conformer speech recognition model. The expected gradient is estimated using Monte Carlo policy gradients on these candidate sequences, and a self-generated baseline reward is used as a baseline to reduce variance. Based on this, the following gradient approximation is obtained: ； in, To optimize the target The gradient approximation, The baseline reward is used to select the reward for finding the optimal path in the beam search, which is used to reduce variance. This represents the nth candidate prediction sequence; N is the number of candidate text sequences generated in one beam search. The weights of the Conformer speech recognition model, for The gradient.

6. The speech recognition optimization method based on adaptive dynamic programming according to claim 5, characterized in that, The Conformer speech recognition model is trained using a joint loss function: ， ； in, For the joint loss function, This represents the training loss based on reward optimization. The training loss is used when employing the CTC framework. These are the weighting coefficients.

7. The speech recognition optimization method based on adaptive dynamic programming according to claim 6, characterized in that, The method evaluates the recognition performance of the Conformer speech recognition model using word error rate (WER) or edit distance (ED): Word error rate ; Edit distance ; in, These represent the number of errors where one character needs to be replaced, deleted, or inserted between the recognized text and the correct text, respectively, and N is the number of words transcribed from the reference text.

8. A speech recognition optimization system based on adaptive dynamic programming, implemented based on the speech recognition optimization method based on adaptive dynamic programming as described in claim 1, characterized in that, It includes a speech acquisition module, a preprocessing module, a process modeling module, a training module, and a recognition module; The voice acquisition module is used to acquire voice data and construct verification datasets and test datasets; The preprocessing module is used to preprocess the acquired speech data and construct an acoustic feature sequence; The process modeling module is used to construct an acoustic state transition model, which models the speech recognition process as a finite-time optimal control problem and determines the optimization objective and policy distribution. The training module trains the Conformer speech recognition model in two stages to learn and optimize the policy distribution. In the first stage, the Conformer speech recognition model is pre-trained using the Connectionist Temporal Classification (CTC) framework. In the second stage, the Conformer speech recognition model is trained using the Adaptive Dynamic Programming (ADP) algorithm and the self-critic sequence training method. The recognition module uses the trained Conformer speech recognition model to perform speech recognition and outputs the recognition results.

9. A computer-readable storage medium storing executable instructions for performing the speech recognition optimization method based on adaptive dynamic programming as described in any one of claims 1-7, characterized in that, When the instruction is executed, it causes the processor to perform the adaptive dynamic programming-based speech recognition optimization method.

10. A computer program product for executing the speech recognition optimization method based on adaptive dynamic programming as described in any one of claims 1-7, characterized in that, This includes a computer program or instructions that, when executed by a processor, implement the aforementioned adaptive dynamic programming-based speech recognition optimization method.

Images