CN103810997B - A kind of method and apparatus for determining voice identification result confidence level - Google Patents

Abstract

The present invention provides a method and device for determining the confidence degree of speech recognition results, wherein the method includes: determining the confidence degree of each arc in the word graph obtained by decoding, and determining the optimal path in the word graph; For each arc A _i on the path, determine the arc set T that has a competitive relationship with the arc A _i in the word graph; when determining the confidence of the word represented by the arc A _i , there is no _competition The arc A _j is determined from the arc set T of the relationship, where the arc A _j and the arc A _i represent the same word, or the combination of the arc A _j and the connected arc constitutes the same word as the arc A _i ; combining the arc A _i and the arc The confidence degree of A _j , or further combine the confidence degree of the arcs connected by the arc A _j to determine the confidence degree of the words represented by the arc A _i . The present invention considers the constituent factors of compound words when determining the confidence level of the speech recognition result, so that the confidence level can more accurately reflect the real situation.

Description

A method and device for determining the confidence level of speech recognition results

【Technical field】

The invention relates to the field of speech recognition in computer application technology, in particular to a method and device for determining the confidence level of speech recognition results.

【Background technique】

In speech recognition, the confidence level is used to indicate the possibility of the recognition result being correct. The larger the value, the higher the possibility of the recognition result being the correct result. It is an important basis for speech recognition. The method for determining the confidence level of the speech recognition result is straightforward. affect the accuracy of speech recognition.

The determination of the confidence of the speech recognition result is mainly obtained by processing the word map (Aattice) generated by decoding. Word graph is a more commonly used expression form of speech recognition results in recent years. It represents multiple decoded candidate results on a directed acyclic graph, saving storage space while retaining multiple candidate information. In the word graph, arcs represent words, and nodes represent the connection relationship of words, and each word belongs to a path from the start node to the end node. Each arc in the word graph can be represented by a quintuple {W,A _w ,L _w ,S _w ,E _w }, where W represents the word corresponding to the arc, A _w represents the acoustic score of the generated word W, and L _w Represents the language score of generating word W, S _w represents the start time of generating word W, and E _w represents the end time of generating word W. Figure 1 is an example of a word graph, where <s> and </s> represent the path start character and path end character, respectively.

When the confidence degree of the existing speech recognition result is determined, the confidence degree of the word is determined according to the optimal path, as shown in Figure 1, for "Chinese people" because the optimal path is the arc "Chinese people", so the word The confidence level of is the confidence level of the arc "Chinese people". However, in Chinese, a word can be composed of two other words, so-called compound words, corresponding to this type of words, just as "Chinese people" is composed of the words "China" and "people", the confidence of the existing speech recognition results The degree determination method ignores the constituent factors of compound words, so that the confidence degree of the recognition result cannot reflect the real situation. Since the confidence degree of the recognition result may have an impact on the subsequent adaptive adjustment process of the acoustic model and language model, it will also affect the accuracy of recognition results.

【Content of invention】

In view of this, the present invention provides a method and device for determining the confidence level of the speech recognition result, so as to improve the accuracy of the confidence level of the speech recognition result.

The specific technical scheme is as follows:

A method of determining confidence in speech recognition results, the method comprising:

S1. Determine the confidence of each arc in the word graph obtained by decoding, and determine the optimal path in the word graph;

S2. For each arc A _i on the optimal path, determine the arc set T that competes with the arc A _i in the word graph;

S3. When determining the confidence level of the word represented by the arc A _i , determine the arc A _j from the arc set T that competes with the A _i , wherein the arc A _j and the arc A _i represent the same word, Or the combination of arc A _j and its connected arc constitutes the same word as arc A _i ; combine the confidence of arc A _i and arc A _j , or further determine the arc A _i in combination with the confidence of the arc connected by said arc A _j . confidence level of the word.

According to a preferred embodiment of the present invention, in the step S1, the confidence of each arc is equal to the value obtained by dividing the sum of the scores of all paths passing through the arc by the sum of the scores of all paths in the word graph.

According to a preferred embodiment of the present invention, when determining whether there is a competitive relationship between the two arcs in the step S2, the following method is adopted:

two arcs are determined to be in contention if they overlap in duration; or,

If the duration of the two arcs overlaps, and the similarity in pronunciation of the words represented by the two arcs satisfies a preset requirement, it is determined that the two arcs are in a competitive relationship.

According to a preferred embodiment of the present invention, the S3 specifically includes:

S31, the confidence degree of the words represented by the initialization arc A _i is the confidence degree of the arc A _i ;

S32. Select an unselected arc from the arc set T that competes with the arc A _i ;

S33, whether the selected arc is judged to represent the same word as the arc A _i , if so, the confidence degree of the word represented by the arc A _i is updated to the confidence degree current value of the word plus the confidence degree of the selected arc, and step S35 is performed ; Otherwise, execute step S34;

S34, judge whether the selected arc and its connected arc combination represent the same word as the arc A _i , if so, combine the current confidence value of the word represented by the arc A _i and the confidence update arc of each arc in the arc combination The degree of confidence of the word that A _i represents, executes step S35; Otherwise directly executes step S35;

S35. Judging whether there are unselected arcs in the arc set T, if yes, go to the step S32; otherwise, end the process of determining the confidence of the word represented by the arc A _i .

According to a preferred embodiment of the present invention, in step S34, the current value of the confidence degree of the word represented by the arc A _i and the confidence degree of each arc in the arc combination are updated in step S34. The confidence degree of the word represented by the arc A _i is specifically:

The confidence degree of the word represented by the arc A _i is updated as the current value of the confidence degree of the word plus the minimum value of the confidence degree of each arc in the arc combination.

A device for determining confidence in speech recognition results, the device comprising:

The initial determination unit is used to determine the confidence degree of each arc in the word graph obtained by decoding, and determine the optimal path in the word graph;

A set determination unit, configured to determine an arc set T that competes with the arc A _i in the word graph for each arc A _i on the optimal path;

Confidence degree determination unit, used to determine _the arc A _{j from the arc set T that competes with the A i} when determining the confidence degree of the word represented by the arc A _i , wherein the arc A _j is the same as the arc A _i Represent the same words, or arc A _j and its connected arcs are combined to form the same words as arc A _i ; combine the confidence of arc A _i and arc A _j , or further combine the confidence of the arcs connected by arc A _j A confidence level is determined for the term represented by the arc A _i .

According to a preferred embodiment of the present invention, the initial determination unit determines that the confidence of each arc is equal to the value obtained by dividing the sum of the scores of all paths passing through the arc by the sum of the scores of all paths in the word graph.

According to a preferred embodiment of the present invention, when the set determination unit determines whether there is a competitive relationship between two arcs, it adopts the following method:

two arcs are determined to be in contention if they overlap in duration; or,

If the duration of the two arcs overlaps, and the similarity in pronunciation of the words represented by the two arcs satisfies a preset requirement, it is determined that the two arcs are in a competitive relationship.

According to a preferred embodiment of the present invention, the confidence determination unit specifically includes:

The initialization subunit is used to initialize the confidence degree of the words represented by the arc A _i to be the confidence degree of the arc A _i , and trigger the arc selection subunit;

The arc selection subunit is configured to select an unselected arc from the arc set T that competes with the arc A _i after being triggered;

The first update subunit is used to judge whether the arc selected by the arc selection subunit expresses the same word as the arc A _i , if so, update the confidence level of the word represented by the arc A _i to the current value of the confidence level of the word Add the confidence of the selected arc to trigger the judgment subunit; otherwise trigger the second update subunit;

The second update subunit is used to judge whether the arc combination selected by the arc selection subunit and the arc it is connected to represent the same word as the arc A _i , if so, combine the current confidence value of the word represented by the arc A _i and The degree of confidence of each arc in the combination of arcs updates the degree of confidence of the word represented by the arc A _i , triggering the judgment subunit; otherwise directly triggering the judgment subunit;

The judging subunit is used to judge whether there are unselected arcs in the arc set T, and if so, trigger the arc selection subunit, otherwise end the process of determining the confidence level of the word represented by the arc A _i .

According to a preferred embodiment of the present invention, when the second update subunit updates the confidence degree of the word represented by the arc A _i , it specifically updates the confidence degree of the word represented by the arc A _i as the current value of the confidence degree of the word plus the The minimum confidence value of each arc in the above arc combination.

As can be seen from the above technical solutions, the present invention considers the constituent factors of compound words when determining the confidence degree of the speech recognition result, and for the situation that a plurality of words are combined to form a word, the confidence degree of this combination is also included in the words. The confidence level is determined, so that the confidence level reflects the real situation more accurately.

【Description of drawings】

Fig. 1 is an example diagram of word graph;

FIG. 2 is a flow chart of the method provided by Embodiment 1 of the present invention;

FIG. 3 is a specific implementation flowchart of step 204 in FIG. 2 provided by Embodiment 1 of the present invention;

FIG. 4 is a structural diagram of a device for determining the confidence level of a speech recognition result provided by Embodiment 2 of the present invention;

FIG. 5 is a structural diagram of a confidence determination unit provided in Embodiment 2 of the present invention.

【Detailed ways】

In order to make the object, technical solution and advantages of the present invention clearer, the present invention will be described in detail below in conjunction with the accompanying drawings and specific embodiments.

Embodiment one,

Fig. 2 is a flow chart of the method provided by Embodiment 1 of the present invention. As shown in Fig. 2, the method may specifically include the following steps:

Step 201: Determine the confidence level of each arc in the decoded word graph.

The confidence of each arc in this step is equal to the value obtained by dividing the sum of the scores of all paths passing through the arc by the sum of the scores of all paths in the word graph, where the path score is the sum of the acoustic score and language score of the path .

Still taking the word graph shown in Figure 1 as an example, assuming that the path "Renmin University" has a score of 5, the path "China"-"People" has a score of 3, and the path "China"-"People" has a score of 2, then can get:

The confidence of the arc "University of the People" is

The confidence of the arc "China" is

The confidence of the arc "people" is

The confidence of the arc "people" is

Step 202: Determine the optimal path in the word graph.

The so-called optimal path in the word graph is the path with the highest score among all the paths.

The above step 201 and step 202 are prior art and will not be repeated here. In addition, the above step 201 and step 202 may also be executed simultaneously, or may be executed sequentially in any order, and the above order is only one embodiment.

Step 203: For each arc A _i on the optimal path, determine the arc set T that competes with the arc A _i in the word graph.

When judging whether there is a competitive relationship between two arcs, it can be determined based on the time factor, that is, if the duration of the two arcs overlaps, it is determined that there is a competitive relationship between the two arcs. For example, two arcs A ₁ and A ₂ : A ₁ ={W ₁ ,A _w1 ,L _w1 ,S _w1 ,E _w1 }, A ₂ ={W ₂ ,A _w2 ,L _w2 ,S _w2 ,E _w2 } , if S _w2 ≤(S _w1 +E _w1 )/2<E _w2 is satisfied, arc A ₂ is considered to be in competition with arc A ₁ .

In order to describe the competitive relationship more accurately, in addition to the time factor, the similarity in pronunciation of the words represented by the two arcs must meet the preset requirements to determine the existence of a competitive relationship. The similarity in pronunciation can be edited by syllables It can be represented by the distance, or it can be represented by the Euclidean distance of the acoustic model or the language model.

Step 204: When determining the confidence of the word represented by each arc A _i on the optimal path, determine the arc A _j from the arc set T that has a competitive relationship with the arc A _i , where the arc A _j and the arc A _i represent The same word, or the combination of arc A _j and its connected arc constitutes the same word as arc A _i , combined with the confidence degree of arc A _i and A _j , or further combined with the confidence degree of the arc connected by A _j above to determine the value represented by arc A _i word confidence.

Specifically, this step can specifically execute the process shown in Figure 3 for each arc A _i on the optimal path to obtain the confidence of the words represented by each arc, as shown in Figure 3, including the following steps:

Step 301: Initialize the confidence of the word represented by the arc A _i as the confidence of the arc A _i .

Step 302: Select an unselected arc from the set of arcs competing with the arc A _i .

Step 303: Judging whether the selected arc and arc _Ai represent the same word, if yes, go to step 304; otherwise, go to step 305.

Step 304: Set the confidence degree of the word represented by the arc A _i as the current value of the confidence degree of the word plus the confidence degree of the selected arc, and execute step 307.

Step 305: Judging whether the combination of the selected arc and the connected arc represents the same word as arc A _i , if yes, go to step 306; otherwise, go to step 307.

Here the selected arc can be extended forward or backward to combine with the arcs it connects.

Step 306: Update the confidence of the word represented by the arc A _i to the current value of the confidence of the word plus the minimum value of the confidence of each arc in the above arc combination, and execute step 307.

For example, in the word diagram shown in Figure 1, the arc "China" is an arc that competes with the arc "Chinese people". Since the arc "China" expands backwards and the arc combination of the arc "People" connected to it also represents "China People", the confidence of the arc "Chinese people" is added to the minimum value between the arc "China" and the arc "People", and the obtained value is used as the confidence of the word "Chinese people".

Using the current confidence value of the word plus the minimum confidence value of each arc in the arc combination is a preferred implementation mode adopted in this embodiment. In addition to this implementation mode, confidence levels such as words can also be used The current value plus the average confidence value of each arc in the arc combination, etc., but the accuracy is not as high as the method of adding the minimum value.

Step 307: Determine whether there is an unselected arc in the set of arcs competing with the arc A _i , if yes, go to step 302; otherwise, end the process of determining the confidence of the word represented by the arc A _i .

In this way, the confidence degree of each arc on the optimal path can be obtained, which includes the fact that each word in the compound word is decoded as a separate word, so that the confidence degree more accurately reflects the word as the optimal recognition result. possibility.

The above is a detailed description of the method provided by the present invention, and the device provided by the present invention will be described in detail below in conjunction with Embodiment 2.

Embodiment two,

FIG. 4 is a structural diagram of an apparatus for determining the confidence of speech recognition results provided by Embodiment 2 of the present invention. As shown in FIG. 4 , the apparatus may include: an initial determination unit 400 , a set determination unit 410 and a confidence determination unit 420 .

First, the initial determination unit 400 determines the confidence of each arc in the decoded word graph, and determines the optimal path in the word graph. Specifically, the confidence of each arc is equal to the value obtained by dividing the sum of the scores of all paths passing through the arc by the sum of the scores of all paths in the word graph, where the path score is the sum of the acoustic score and language score of the path . The optimal path in the word graph is the path with the highest score among all paths.

Then, for each arc A _i on the optimal path, the set determination unit 410 determines the arc set T that competes with the arc A _i in the word graph.

Wherein, when the set determination unit 410 determines whether there is a competitive relationship between the two arcs, the following method can be adopted: if the two arcs overlap in duration, then determine that the two arcs have a competitive relationship; or, if the two arcs are in the If there is an overlap in the duration, and the similarity in pronunciation of the words represented by the two arcs meets the preset requirements, then it is determined that the two arcs have a competitive relationship. The similarity in pronunciation can be reflected by the edit distance of syllables, or the Euclidean distance of the acoustic model or language model.

Finally, the confidence determination unit 420 determines the arc A _j from the arc set T that competes with A _i when determining the confidence of the word represented by the arc A i, wherein the arc _{A j and} the arc A _i represent the same word, Or the combination of arc A _j and its connected arc constitutes the same word as arc A _i ; combine arc A _i and arc A _j , or further combine the confidence of the arc connected by arc A _j to determine the confidence of the word represented by arc A _i .

The specific structure of the confidence degree determination unit 420 is described in detail below. As shown in FIG. Two update subunit 424 and judgment subunit 425.

Among them, the initialization subunit 421 is used to initialize the confidence degree of the word represented by the arc A _i as the confidence degree of the arc A _i , and then trigger the arc selection subunit 422 .

The arc selection subunit 422 is configured to select an unselected arc from the arc set T that competes with the arc A _i after being triggered.

The first update subunit 423 is used to judge whether the arc selected by the arc selection subunit 422 represents the same word as the arc A _i , if so, update the confidence level of the word represented by the arc A _i to the current value of the confidence level of the word Add the confidence of the selected arc, trigger the judgment subunit 425; otherwise trigger the second update subunit 424.

The second update subunit 424 is used to judge whether the arc combination selected by the arc selection subunit 422 and the arc it is connected to represent the same word as the arc A _i , if so, combine the current confidence value of the word represented by the arc A _i and The confidence of each arc in the arc combination updates the confidence of the word represented by the arc A _i , triggering the judging subunit 425; otherwise, directly triggering the judging subunit 425.

Preferably, when the second update subunit updates the confidence of the word represented by the arc A _i , the confidence of the word represented by the arc A _i can be updated as the current value of the confidence of the word plus the confidence of each arc in the arc combination min.

The judging subunit 425 is used for judging whether there are unselected arcs in the arc set T, if so, triggering the arc selection subunit 422, otherwise ending the process of determining the confidence of the word represented by the arc A _i .

After using the above method and device to determine the confidence of the word represented by each arc on the optimal path, it may include but not limited to the following applications:

1) If the confidence of a word on the optimal path is lower than the preset confidence threshold, it means that there are inaccurate recognition results in the recognition results determined by the optimal path. In order to avoid recognition errors and bring poor results to users For user experience, you can refuse to return the recognition result, and you can further prompt the user to input the voice again.

2) The confidence of the determined words is applied to the unsupervised adaptive technology of speech recognition, that is, it is used in the subsequent adaptive adjustment process of the acoustic model and the speech model, so that the speech recognition is more accurate.

3) It can be used to correct the recognition results. If the confidence of a word is lower than the preset confidence threshold, it indicates that there is an error in the recognition of the word, which provides a basis for error correction of the recognition results.

The above descriptions are only preferred embodiments of the present invention, and are not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present invention shall be included in the present invention. within the scope of protection.

Claims (8)

1. A method for determining the degree of confidence in speech recognition results, characterized in that the method comprises: S1. Determine the confidence of each arc in the word graph obtained by decoding, and determine the optimal path in the word graph; S2. For each arc A _i on the optimal path, determine the arc set T that competes with the arc A _i in the word graph; S31, the confidence degree of the words represented by the initialization arc A _i is the confidence degree of the arc A _i ; S32. Select an unselected arc from the arc set T that competes with the arc A _i ; S33, whether the selected arc is judged to represent the same word as the arc A _i , if so, the confidence degree of the word represented by the arc A _i is updated to the confidence degree current value of the word plus the confidence degree of the selected arc, and step S35 is performed ; Otherwise, execute step S34; S34, judge whether the selected arc and its connected arc combination represent the same word as the arc A _i , if so, combine the current confidence value of the word represented by the arc A _i and the confidence update arc of each arc in the arc combination The degree of confidence of the word that A _i represents, executes step S35; Otherwise directly executes step S35; S35. Judging whether there are unselected arcs in the arc set T, if yes, go to the step S32; otherwise, end the process of determining the confidence of the word represented by the arc A _i . 2. The method according to claim 1, characterized in that, in said step S1, the confidence of each arc is equal to the sum of the scores of all paths passing through the arc divided by the sum of the scores of all paths in the word graph the resulting value. 3. The method according to claim 1, characterized in that, when determining whether there is a competitive relationship between the two arcs in the step S2, the following method is adopted: two arcs are determined to be in contention if they overlap in duration; or, If the duration of the two arcs overlaps, and the similarity in pronunciation of the words represented by the two arcs satisfies a preset requirement, it is determined that the two arcs are in a competitive relationship. 4. The method according to claim 1, characterized in that, in step S34, the confidence level current value of the words represented by the combined arc _Ai and the confidence level of each arc in the combination of arcs are updated in step S34 _. The word confidence is specifically: The confidence degree of the word represented by the arc A _i is updated as the current value of the confidence degree of the word plus the minimum value of the confidence degree of each arc in the arc combination. 5. A device for determining the degree of confidence in speech recognition results, characterized in that the device comprises: The initial determination unit is used to determine the confidence degree of each arc in the word graph obtained by decoding, and determine the optimal path in the word graph; A set determination unit, configured to determine an arc set T that competes with the arc A _i in the word graph for each arc A _i on the optimal path; The confidence determination unit specifically includes: The initialization subunit is used to initialize the confidence degree of the words represented by the arc A _i to be the confidence degree of the arc A _i , and trigger the arc selection subunit; The arc selection subunit is configured to select an unselected arc from the arc set T that competes with the arc A _i after being triggered; The first update subunit is used to judge whether the arc selected by the arc selection subunit expresses the same word as the arc A _i , if so, update the confidence level of the word represented by the arc A _i to the current value of the confidence level of the word Add the confidence of the selected arc to trigger the judgment subunit; otherwise trigger the second update subunit; The second update subunit is used to judge whether the arc combination selected by the arc selection subunit and the arc it is connected to represent the same word as the arc A _i , if so, combine the current confidence value of the word represented by the arc A _i and The degree of confidence of each arc in the combination of arcs updates the degree of confidence of the word represented by the arc A _i , triggering the judgment subunit; otherwise directly triggering the judgment subunit; The judging subunit is used to judge whether there are unselected arcs in the arc set T, and if so, trigger the arc selection subunit, otherwise end the process of determining the confidence level of the word represented by the arc A _i . 6. The device according to claim 5, wherein the initial determination unit determines that the confidence of each arc is equal to the sum of the scores of all paths passing through the arc divided by the sum of the scores of all paths in the word graph. get the value. 7. The device according to claim 5, wherein the set determination unit adopts the following method when determining whether there is a competitive relationship between two arcs: two arcs are determined to be in contention if they overlap in duration; or, If the duration of the two arcs overlaps, and the similarity in pronunciation of the words represented by the two arcs satisfies a preset requirement, it is determined that the two arcs are in a competitive relationship. 8. The device according to claim 5, wherein when the second update subunit updates the confidence of the word represented by the arc A _i , it specifically updates the confidence of the word represented by the arc A _i to the value of the word. The current value of the confidence plus the minimum confidence value of each arc in the combination of arcs.