Speech recognition system and method based on double-weight directed graph

Abstract

The invention discloses a speech recognition system and method based on a double-weight directed graph, and the method comprises the steps: 1), dividing continuous speech into a plurality of audio frames through a speech stream module, and transmitting the obtained audio frames to an acoustic module; 2) the acoustic module converts the audio frame into an acoustic model; (3) the language module generates a directed and weighted decoded graph according to the text knowledge base and sends the weighted and directed decoded graph to a double-weight wide search module; 4) the double-weight wide search module processes the acoustic model and the weighted and directed decoded graph to preliminarily generate a recognition result; and 5) the recognition output module receives the recognition result and the total path loss value sent by the double-weight wide search module, further processes the recognition result and the total path loss value to obtain a language recognition result, and outputs the language recognition result. According to the method and the device, logic language output and non-logic voice recognition can be considered, and the recognition rate of non-logic languages such as telephone numbers is improved while a language model is kept to be used for removing the influence influencing pronunciation approaching.

Description

Technical Field

[0001] The present invention relates to the field of speech recognition, and in particular to a speech recognition system and method based on a dual-weighted directed graph. Background Art

[0002] A weighted finite state transducer based on a directed graph is used to integrate the acoustic model and the text model, and a balanced weight is used to control the speech recognition tendency towards the acoustic model or the text model. When the weight of the text model is increased, the output is more consistent with the language logic, but it affects the illogical recognition such as telephone number recognition. When the weight of the acoustic model is increased, the recognition rate of illogical language is improved, but it affects the recognition rate of similar pronunciations. It cannot guarantee the improvement of both effects at the same time. Therefore, it is very necessary to propose a speech recognition system and method based on a dual-weighted directed graph. Summary of the invention

[0003] The purpose of the present invention is to provide a speech recognition system and method based on a dual-weighted directed graph, so that in the speech recognition process, a language model is used to remove the influence of pronunciation proximity while improving the recognition rate of illogical language, so as to solve the existing technical defects and unattainable technical requirements.

[0004] To achieve the above object, the present invention provides the following technical solution: a speech recognition system based on a dual-weighted directed graph, comprising:

[0005] A language stream module, which converts speech into video frames and then sends the video frames to the acoustic module;

[0006] An acoustic module, which receives the video frames sent by the language stream module, converts the audio into an acoustic model, and then sends the acoustic model to the dual-weighted wideband search module;

[0007] A language module, which converts the text knowledge base into a directed and weighted decoding graph, and sends the directed and weighted decoding graph to the dual-weighted wide search module;

[0008] A dual-weighted wide search module, which receives the acoustic model sent by the acoustic module and the directed weighted decoding graph sent by the language module, sets two balancing weights, uses the two balancing weights to perform a width search on the directed weighted decoding graph, finds all possible results, and sends all possible results to the recognition output module;

[0009] The recognition output module is used to receive all possible results sent by the dual-weighted wide search module, recognize all possible results, and output language recognition results.

[0010] A method for speech recognition based on a dual-weight directed graph comprises the following steps:

[0011] 1) The language stream module divides the continuous speech into multiple audio frames and sends the obtained multiple audio frames to the acoustic module;

[0012] 2) The acoustic module converts the audio frame into an acoustic model

[0013] 2.1) The acoustic module receives the audio frames sent by the language stream module;

[0014] 2.2) The acoustic module multiplies the audio frame by the Hamming window function;

[0015] 2.3) The audio frames processed in step 2.2) of the acoustic module are converted into frequency domain data by FFT;

[0016] 2.4) The acoustic module processes the frequency domain data and generates an acoustic model;

[0017] 2.5) The acoustic module sends the acoustic model to the dual-weighted wide search module;

[0018] 3) The language module generates a directed and weighted decoding graph based on the text knowledge base, and sends this weighted and directed decoding graph to the dual-weighted wide search module;

[0019] 4) The dual-weighted wide search module processes the acoustic model and the weighted directed decoding graph to generate preliminary recognition results

[0020] 4.1) The dual-weighted wide search module receives the acoustic model sent by the acoustic module and the weighted directed decoding graph sent by the language module;

[0021] 4.2) The dual-weighted wide search module sets two balancing weights;

[0022] 4.3) The dual-weighted wide search module starts the wide search from the starting node of the directed weighted decoding graph, and adds the starting node and the total path loss value to the next round of wide search queue;

[0023] 4.4) Every time the dual-weight wide search module receives the acoustic model of an audio frame, it traverses the wide search queue to take out the current processing node and the total path loss value, and processes them in sequence to obtain a new node;

[0024] 4.5) The dual-weighted wide search module uses the new node to obtain the loss value of the vocabulary, and further processes the loss value of the vocabulary to obtain the loss value of the two paths of the vocabulary;

[0025] 4.6) The dual-weighted width search module further processes the total path loss value in step 4.3 to obtain two total path loss values;

[0026] 4.7) The dual-weighted width search module compares the total loss values ​​of the two paths with the truncation loss values, filters out the paths that need to be discarded, and further processes the paths that are not discarded.

[0027] 4.8) When the audio of a speech ends, the dual-weighted wide search module generates the recognition result of the path and the total path loss value;

[0028] 4.9) The dual-weighted width search module sends the recognition result and the total path loss value to the recognition output module;

[0029] 5) The recognition output module receives the recognition result and the total path loss value sent by the dual-weighted width search module, further processes the two, obtains the language recognition result, and outputs the language recognition result.

[0030] It should be noted that there is a sound length screening process between step 4.2) and step 4.3): the current processing node represents the phoneme of a word in the acoustic model output query probability value. If the probability value is greater than the same word threshold, it is considered that the pronunciation of the word has not ended, and the current processing node and the total path loss value are added to the next round of wide search queue. If the probability value is less than the same word threshold, it means that the current processing node has no upper or lower node in the directed authorized decoding graph, then it is considered that some partial recognition results have been found, and the path loss value is accumulated to the total path loss value. Then perform the processing of step 4.3), that is, the initial node and the total path loss value of the starting node of the directed authorized decoding graph are initially zero, and the total path loss value and the initial node are added to the next round of wide search queue to identify the remaining speech. The remaining speech here is the remaining speech after the above steps. It can be understood that the content of 4.3 processing is the speech after the sound length screening in the above steps.

[0031] Preferably, in step 1), the specific requirements for dividing the continuous speech into multiple audio frames are:

[0032] There is a 1 / 2 overlap area between two adjacent frames.

[0033] Preferably, the specific contents of step 2.4) include:

[0034] 2.4.1) The acoustic module uses a Mel filter bank to filter the power spectrum in the frequency domain data to simulate the hearing characteristics of the human ear;

[0035] Among them, the Mel filter bank consists of a series of triangular filters;

[0036] 2.4.2) The acoustic module takes the logarithm of the filtered signal to obtain the Fbank feature;

[0037] 2.4.3) The acoustic module uses the Fbank feature input into the Transformer neural network encoder and outputs the result as the acoustic model.

[0038] Preferably, the specific contents processed in step 4.4) are:

[0039] 4.4.1) The dual-weighted wide search module extracts the weights of all next nodes and the edges between two nodes from the directed and weighted decoding graph according to the current processing node, and processes all next nodes as new nodes one by one;

[0040] 4.4.2) The dual-weighted width search module uses one minus the weights of the two node edges to obtain the loss value of the two node edges.

[0041] Preferably, the specific contents of step 4.5) are:

[0042] 4.5.1) The dual-weighted width search module uses a new node to represent the probability value of a phoneme of a word that can be queried in the acoustic model output. This probability value is the possibility of the word. The dual-weighted width search module uses one minus the probability value to obtain the loss value of the word;

[0043] 4.5.2) The dual-weighted wide search module uses the two balancing weights in step 4.2) to multiply the loss value of the vocabulary in step 4.5.1) to obtain the loss values ​​of the two paths of the vocabulary, namely the loss value with a high balancing weight and the loss value with a low balancing weight.

[0044] Preferably, the specific content of step 4.6) is: the dual-weighted width search module adds the total path loss value in step 4.3) to the loss value of the two node edges in step 4.4.2), and then adds the loss values ​​of the two paths of the vocabulary of the new node in step 4.5.2) to obtain two total path loss values.

[0045] 8. A method for speech recognition based on a dual-weight directed graph according to claim 2, characterized in that the specific contents of step 4.7) are:

[0046] 4.7.1), the dual-weighted width search module compares the total loss values ​​of the two paths in step 4.6) with the truncation loss values;

[0047] 4.7.2) Set the traverse loss value. If the total path loss value is greater than the truncation loss value, the path is discarded.

[0048] 4.7.3) If the total path loss value is less than the truncation loss value, the dual-weight wide search module adds the new node and the new path total loss value to the next round of wide search queue. A new node can be added to the next round at most twice.

[0049] Preferably, the specific contents of the identification result of the path and the total path loss value generated by the dual-weighted width search module in step 4.8) are:

[0050] 4.8.1) The dual-weight wide search module traverses the wide search queue and retrieves the total path loss value;

[0051] 4.8.2) Take the upper words of the path and combine them into the preliminary recognition result of the path.

[0052] 10. A method for speech recognition based on a dual-weight directed graph according to claim 2, characterized in that the specific contents of further processing the recognition result and the total path loss value in step 5) include:

[0053] 5.1) Set the reduction coefficient and character number threshold. The recognition output module performs special vocabulary statistics on the preliminary recognition results in step 4.8.2). If it exceeds the set character number threshold, the total path loss value is divided by the reduction coefficient; the character number threshold means the set value for the number of characters.

[0054] 5.2) The identification output module sorts the adjusted total path loss values ​​from small to large;

[0055] 5.3) The recognition output module outputs the recognition result with the smallest total path loss value as the final language recognition result.

[0056] Compared with the prior art, the present invention has the following beneficial effects:

[0057] 1. This application can take into account both logical language output and non-logical speech recognition, retain the use of language models to remove the influence of pronunciation proximity, while improving the recognition rate of non-logical languages ​​such as telephone numbers. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 It is a logical relationship diagram between the modules in the present invention;

[0059] In the figure: language stream module 1; acoustic model 2; language module 3; dual-weighted wide search module 4; recognition output module 5. DETAILED DESCRIPTION

[0060] The following will be combined with the attached embodiment of the present invention Figure 1 , the technical solutions in the embodiments of the present invention are clearly and completely described. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0061] See also Figure 1 , embodiments of the present invention:

[0062] Example:

[0063] like Figure 1 As shown: A speech recognition system based on a dual-weight directed graph, comprising:

[0064] A language stream module 1, which converts speech into video frames and then sends the video frames to an acoustic module 2;

[0065] Acoustic module 2, which receives the video frame sent by the language stream module 1, converts the audio into an acoustic model, and then sends the acoustic model to the dual-weighted wide search module 4;

[0066] A language module 3 converts the text knowledge base into a directed and weighted decoding graph, and sends the directed and weighted decoding graph to a dual-weighted wide search module 4;

[0067] A dual-weighted wide search module 4 receives the acoustic model sent by the acoustic module 2 and the directed weighted decoding graph sent by the language module 3, sets two balancing weights, uses the two balancing weights to perform a width search on the directed weighted decoding graph, finds all possible results, and sends all possible results to the recognition output module 5;

[0068] The recognition output module 5 is used to receive all possible results sent by the dual-weighted wide search module 4, recognize all possible results, and output language recognition results.

[0069] A method for speech recognition based on a dual-weight directed graph comprises the following steps:

[0070] 1) The language stream module 1 divides the continuous speech into multiple audio frames, and sends the obtained multiple audio frames to the acoustic module 2;

[0071] Specifically, the language stream module 1 divides the continuous speech into multiple audio frames. A frame is a collection of N sampling points into an observation unit. There is a 1 / 2 overlapping area between two adjacent frames to avoid excessive changes between two adjacent frames.

[0072] 2) Acoustic module 2 converts audio frames into acoustic models

[0073] 2.1) Acoustic module 2 receives the audio frames sent by language stream module 1;

[0074] 2.2) Acoustic module 2 multiplies the audio frame by the Hamming window function to increase the continuity of the left and right ends of the frame and reduce spectrum leakage;

[0075] 2.3) The audio frames processed in step 2.2) of acoustic module 2 are converted into frequency domain data by FFT;

[0076] 2.4) Acoustic module 2 processes the frequency domain data to generate an acoustic model;

[0077] The specific contents of the step 2.4) include:

[0078] 2.4.1) Acoustic module 2 applies a Mel filter bank to filter the power spectrum in the frequency domain data to simulate the auditory characteristics of the human ear. The Mel filter bank consists of a series of triangular filters.

[0079] 2.4.2), acoustic module 2 takes the logarithm of the filtered signal to obtain the Fbank feature;

[0080] 2.4.3) Acoustic module 2 uses Fbank features to input into the Transformer neural network encoder and outputs the result as the acoustic model.

[0081] Specifically, it can be understood that the acoustic module 2 uses the Fbank feature as the input of the Transformer neural network encoder, and finally obtains the acoustic model output. The acoustic model output is expressed as a probability value of a phoneme series, and each audio frame corresponds to a phoneme series probability value.

[0082] 2.5) The acoustic module 2 sends the acoustic model to the dual-weighted wide search module 4;

[0083] 3) The language module 3 generates a directed and weighted decoding graph according to the text knowledge base, and sends the weighted and directed decoding graph to the dual-weighted wide search module 4;

[0084] Specifically, based on the occurrence of a large number of sentences and words in the text knowledge base, the language module 3 uses a statistical language model to automatically learn syntactic rules through text data, and finally generates a directed and authorized decoding graph, in which the numbers of years, etc. have no logic. Before using the statistical language model, the language module 3 performs preprocessing to remove number-related sentences.

[0085] 4) Dual-weighted wide search module 4 processes the acoustic model and the weighted directed decoding graph to generate a preliminary recognition result

[0086] 4.1) The dual-weighted wide search module 4 receives the acoustic model sent by the acoustic module 2 and the weighted directed decoding graph sent by the language module 3;

[0087] 4.2) The dual-weighted wide search module 4 sets two balancing weights, taking into account the output of the acoustic model and the high-weight output of the directed and weighted decoding graph, taking into account the recognition of the logical language output and the non-logical speech;

[0088] 4.3) The dual-weighted wide search module 4 starts the wide search from the starting node of the directed right decoding graph, and adds the starting node and the total path loss value to the next round of wide search queue, and the starting node and the total path loss value are zero;

[0089] 4.4) Every time the dual-weight wide search module 4 receives the acoustic model of an audio frame, it traverses the wide search queue to take out the current processing node and the total path loss value, and processes them in sequence to obtain a new node;

[0090] The specific contents processed in step 4.4) are as follows:

[0091] 4.4.1) The dual-weighted wide search module 4 extracts the weights of all next nodes and the edges between two nodes from the directed and weighted decoding graph according to the current processing node, and processes all next nodes as new nodes one by one;

[0092] The nodes of the directed weighted decoding graph represent a word, and the weight of the edge between two nodes represents the possibility of one word and another word;

[0093] 4.4.2) The dual-weighted wide search module 4 uses one minus the weights of the two node edges to obtain the loss values ​​of the two node edges.

[0094] The loss value of the edge between two nodes represents the cost of going from one word to another, that is, the cost represents one minus the likelihood of two nodes.

[0095] 4.5) The dual-weight wide search module 4 uses the new node to obtain the loss value of the vocabulary, and further processes the loss value of the vocabulary to obtain the two-way loss value of the vocabulary, wherein the loss value with a high balance weight will increase the weight of the acoustic model, which will increase the loss value of the vocabulary and shorten the path length found by the wide search in the directed weighted decoding graph, thereby being more conducive to the recognition of content without language logic; wherein the loss value with a low balance weight will reduce the loss value of the vocabulary, and the wide search can find a longer path, making the recognized content more language-logical.

[0096] The specific contents of step 4.5) are:

[0097] 4.5.1) The double-weighted width search module 4 uses a new node to represent the probability value of a phoneme of a word that can be queried in the acoustic model output. This probability value is the possibility of the word. The double-weighted width search module 4 uses one minus the probability value to obtain the loss value of the word;

[0098] 4.5.2) The dual-weighted wide search module 4 uses the two balancing weights in step 4.2) to multiply the loss value of the vocabulary in step 4.5.1) to obtain the loss values ​​of the two paths of the vocabulary.

[0099] 4.6) The dual-weighted width search module 4 further processes the total path loss value in step 4.3 to obtain two total path loss values;

[0100] The specific content of step 4.6) is: the dual-weighted wide search module 4 adds the total path loss value in step 4.3) to the loss value of the two node edges in step 4.4.2), and then adds the loss values ​​of the two paths of the vocabulary of the new node in step 4.5.2) to obtain two total path loss values.

[0101] 4.7) The dual-weighted width search module 4 compares the total loss values ​​of the two paths with the truncation loss values, selects the paths that need to be discarded, and further processes the paths that are not discarded.

[0102] The specific contents of step 4.7) are:

[0103] 4.7.1), the dual-weighted width search module 4 compares the total loss values ​​of the two paths in step 4.6) with the truncation loss values;

[0104] 4.7.2) Set the traverse loss value. If the total path loss value is greater than the truncation loss value, the path is discarded.

[0105] 4.7.3) If the total path loss value is less than the truncation loss value, the dual-weight wide search module (4) adds the new node and the new path total loss value to the next round of wide search queue. A new node is added to the next round at most twice, thereby achieving simultaneous recognition of logical language output and non-logical speech.

[0106] 4.8) When the audio of a speech ends, the dual-weighted wide search module 4 generates the recognition result of the path and the total path loss value;

[0107] The specific contents of the path identification result and the total path loss value generated by the dual-weighted width search module 4 in step 4.8) are:

[0108] 4.8.1) The dual-weight wide search module 4 traverses the wide search queue and retrieves the total path loss value;

[0109] 4.8.2) Take the upper words of the path and combine them into the preliminary recognition result of the path.

[0110] 4.9) The dual-weighted width search module 4 sends the recognition result and the total path loss value to the recognition output module 5

[0111] 5) The recognition output module 5 receives the recognition result and the total path loss value sent by the dual-weighted width search module 4, further processes the two, obtains the language recognition result, and outputs the language recognition result.

[0112] The specific contents of further processing the identification result and the total path loss value in step 5) include:

[0113] 5.1) Set the character number threshold of the reduction coefficient. The recognition output module 5 performs special vocabulary statistics on the preliminary recognition result in step 4.8.2), such as numbers and other non-linguistic logical words. If the character number threshold is exceeded, the total path loss value is divided by the reduction coefficient to reduce the total path loss value and improve the selection rate of the recognition result.

[0114] 5.2) The identification output module 5 sorts the adjusted total path loss values ​​from small to large; adjusted means after processing in step 5.1.

[0115] 5.3) The recognition output module 5 outputs the recognition result with the smallest total path loss value as the final language recognition result.

[0116] The above shows and describes the basic principles and main features of the present invention and the advantages of the present invention. It is obvious to those skilled in the art that the present invention is not limited to the details of the above exemplary embodiments, and the present invention can be implemented in other specific forms without departing from the spirit or basic features of the present invention. Therefore, no matter from which point of view, the embodiments should be regarded as exemplary and non-restrictive. The scope of the present invention is defined by the attached claims rather than the above description, and it is intended that all changes falling within the meaning and scope of the equivalent elements of the claims are included in the present invention. Any figure mark in the claims should not be regarded as limiting the claims involved.

[0117] In addition, it should be understood that although the present specification is described according to implementation modes, not every implementation mode contains only one independent technical solution. This description of the specification is only for the sake of clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment can also be appropriately combined to form other implementation modes that those skilled in the art can understand.

Claims

1. A speech recognition system based on a dual-weight directed graph, characterized in that: include: A language stream module (1), wherein the language stream module (1) converts speech into video frames and then sends the video frames to the acoustic module (2); An acoustic module (2), wherein the acoustic module (2) receives the video frame sent by the language stream module (1), converts the audio into an acoustic model, and then sends the acoustic model to the dual-weighted wideband search module (4); A language module (3), wherein the language module (3) converts the text knowledge base into a directed weighted decoding graph, and sends the directed weighted decoding graph to a dual-weighted wide search module (4); A dual-weighted wide search module (4), wherein the dual-weighted wide search module (4) receives the acoustic model sent by the acoustic module (2) and the directed weighted decoding graph sent by the language module (3), sets two balancing weights, uses the two balancing weights to perform a width search on the directed weighted decoding graph, finds all possible results, and sends all possible results to the recognition output module (5); The recognition output module (5) is used to receive all possible results sent by the dual-weighted wide search module (4), recognize all possible results, and output language recognition results.

2. A speech recognition method based on a dual-weight directed graph, characterized in that: The following steps are involved: 1) The language stream module (1) divides the continuous speech into multiple audio frames, and sends the obtained multiple audio frames to the acoustic module (2); 2), the acoustic module (2) converts the audio frame into an acoustic model; 2.1), the acoustic module (2) receives the audio frames sent by the language stream module (1); 2.2), the acoustic module (2) multiplies the audio frame by the Hamming window function; 2.3) Acoustic module (2) The audio frames processed in step 2.2) are converted into frequency domain data by FFT; 2.4) The acoustic module (2) processes the frequency domain data to generate an acoustic model; 2.5) The acoustic module (2) sends the acoustic model to the dual-weighted wide search module (4); 3) The language module (3) generates a directed and weighted decoding graph according to the text knowledge base, and sends the weighted and directed decoding graph to the dual-weighted wide search module (4); 4) The dual-weighted wide search module (4) processes the acoustic model and the weighted directed decoding graph to initially generate a recognition result; 4.1), the dual-weighted wide search module (4) receives the acoustic model sent by the acoustic module (2) and the weighted directed decoding graph sent by the language module (3); 4.2) The dual-weighted wide search module (4) sets two balancing weights; 4.3) The dual-weighted wide search module (4) starts the wide search from the starting node of the directed weighted decoding graph, and adds the starting node and the total path loss value to the next round of wide search queue; 4.4) The dual-weight wide search module (4) traverses the wide search queue to retrieve the current processing node and the total path loss value each time it receives the acoustic model of an audio frame, and processes them in sequence to obtain a new node; 4.5) The dual-weighted wide search module (4) uses the new node to obtain the loss value of the vocabulary, and further processes the loss value of the vocabulary to obtain the loss value of the two paths of the vocabulary; 4.6), the dual-weighted width search module (4) further processes the total path loss value in step 4.3 to obtain two total path loss values; 4.7) The dual-weighted width search module (4) compares the total loss value of the two paths with the truncation loss value, selects the path to be discarded, and further processes the path that is not discarded; 4.8) When the audio of a speech ends, the dual-weighted wide search module (4) generates the recognition result of the path and the total path loss value; 4.9) The dual-weighted width search module (4) sends the recognition result and the total path loss value to the recognition output module (5); 5) The recognition output module (5) receives the recognition result and the total path loss value sent by the dual-weighted width search module (4), further processes the two, obtains the language recognition result, and outputs the language recognition result.

3. A method for speech recognition based on a dual-weight directed graph according to claim 2, characterized in that: In step 1), the specific requirements for dividing the continuous speech into multiple audio frames are: There is a 1 / 2 overlap area between two adjacent frames.

4. A method for speech recognition based on a dual-weight directed graph according to claim 3, characterized in that: The specific contents of the step 2.4) include: 2.4.1) Acoustic module (2) applies a Mel filter bank to filter the power spectrum in the frequency domain data to simulate the hearing characteristics of the human ear; 2.4.2), the acoustic module (2) takes the logarithm of the filtered signal to obtain the Fbank feature; 2.4.3) The acoustic module (2) uses the Fbank feature input into the Transformer neural network encoder and outputs the result as the acoustic model.

5. A method for speech recognition based on a dual-weight directed graph according to claim 4, characterized in that: The specific contents processed in step 4.4) are as follows: 4.4.1) The dual-weighted wide search module (4) extracts the weights of all next nodes and the edges between two nodes from the directed weighted decoding graph according to the current processing node, and processes all next nodes as new nodes one by one; 4.4.2) The dual-weighted width search module (4) uses one minus the weights of the two node edges to obtain the loss values ​​of the two node edges.

6. A method for speech recognition based on a dual-weight directed graph according to claim 5, characterized in that: The specific contents of step 4.5) are: 4.5.1) The double-weighted width search module (4) uses a new node to represent the probability value of a phoneme of a word that can be queried in the acoustic model output. This probability value is the possibility of the word. The double-weighted width search module (4) uses one minus the probability value to obtain the loss value of the word; 4.5.2) The dual-weighted wide search module (4) uses the two balancing weights in step 4.2) to multiply the loss value of the vocabulary in step 4.5.1) to obtain the two-way loss value of the vocabulary.

7. A method for speech recognition based on a dual-weight directed graph according to claim 6, characterized in that: The specific content of step 4.6) is: the dual-weighted wide search module (4) adds the total path loss value in step 4.3) to the loss value of the two node edges in step 4.4.2), and then adds the loss values ​​of the two paths of the vocabulary of the new node in step 4.5.2) to obtain two total path loss values.

8. A method for speech recognition based on a dual-weight directed graph according to claim 7, characterized in that: The specific contents of step 4.7) are: 4.7.1), the dual-weighted width search module (4) compares the total loss values ​​of the two paths in step 4.6) with the truncated loss values; 4.7.2) Set the traverse loss value. If the total path loss value is greater than the truncation loss value, the path is discarded. 4.7.3) If the total path loss value is less than the truncation loss value, the dual-weight wide search module (4) adds the new node and the new path loss value to the next round of wide search queue. A new node can be added to the next round at most twice.

9. A method for speech recognition based on a dual-weight directed graph according to claim 8, characterized in that: The specific contents of the path identification result and the path total loss value generated by the dual-weighted width search module (4) in step 4.8) are: 4.8.1) The dual-weight wide search module (4) traverses the wide search queue and extracts the total path loss value; 4.8.2) Take the upper words of the path and combine them into the preliminary recognition result of the path.

10. A method for speech recognition based on a dual-weight directed graph according to claim 9, characterized in that: The specific contents of further processing the identification result and the total path loss value in step 5) include: 5.1) Set a character count threshold of the reduction coefficient, and the recognition output module (5) performs special vocabulary statistics on the preliminary recognition results in step 4.8.2). If the character count exceeds the set threshold, the total path loss value is divided by the reduction coefficient; 5.2), the identification output module (5) sorts the adjusted total path loss values ​​from small to large; 5.3) The recognition output module (5) outputs the recognition result with the smallest total path loss value as the final language recognition result.

Images