Recognition rejection method and device for out-of-set words, equipment and medium

Abstract

The invention relates to the technical field of speech recognition, and discloses an out-of-set word rejection recognition method and device, equipment and a medium, and the method comprises the steps: carrying out the forced alignment of a current speech segment recognition result and a target command word, and obtaining an alignment path, and the process is based on dynamic planning and a Viterbi algorithm, and is combined with a CTC criterion; determining a key interval according to the alignment path, and judging and counting wrong line counts according to two types of alignment results of a bank phoneme and a target command word phoneme; positioning a key phoneme pair containing position association in combination with the alignment path and the phoneme sequence characteristics of the target command word; calculating the position distance of the key phoneme pair and judging whether the position distance is abnormal; and if the error line count exceeds the set threshold value or the position distance is abnormal, determining the word as an out-of-set word and refusing to identify, otherwise, determining the word as an effective command word. The method does not need to train a model again, is adaptive to low-resource equipment, is simple in calculation, is small in resource occupation, can precisely recognize the similar set external words, reduces the error recognition rate, improves the voice interaction accuracy, and meets the precise control demands of a smart home.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of speech recognition, in particular, to a method and device for rejecting out-of-set words, and a related equipment and medium. BACKGROUND

[0002] Command word recognition is a key application direction in the field of speech recognition, and is widely used in the field of smart home. Typical scenarios include voice interaction control of devices such as smart voice sound boxes, smart voice earphones, smart voice lamps, and smart voice fans. At present, low-resource smart home devices are usually equipped with a fixed command word list for recognition due to hardware performance limitations. However, this mode has significant defects: the devices are prone to misrecognition of out-of-set words, especially those similar to the pronunciation or semantics of the command words in the list. For example, when the command word list only contains "set time for twenty minutes", out-of-set words such as "set time for twenty-five minutes" and "set time for twenty-eight minutes" should not be recognized, but in actual application, they are often misrecognized as "set time for twenty minutes", which seriously affects the interaction accuracy of the device and makes it difficult to meet the needs of users for precise voice control of smart home devices. SUMMARY

[0003] The main purpose of the present application is to provide a method for rejecting out-of-set words, aiming to solve the technical problem of high misrecognition rate of similar out-of-set words when low-resource smart home devices use a fixed command word list for recognition, which affects the interaction accuracy and makes it difficult to meet the needs of precise control.

[0004] The first aspect of the present application provides a method for rejecting out-of-set words, comprising: performing forced alignment on the current speech segment recognition result and the target command word in the command word list to obtain an alignment path, wherein the forced alignment is based on dynamic programming and Viterbi algorithm, and the decoding matrix generated by decoding the speech segment is aligned with the standard phoneme path of the target command word in combination with the CTC criterion; based on the alignment path, determining a key interval and judging the alignment result in the interval according to different cases, and counting the number of error rows, wherein when the alignment result is a blank phoneme or a target command word phoneme, the proportion is judged respectively; based on the alignment path and the phoneme sequence features of the target command word, locating at least one group of key phoneme pairs, each group of key phoneme pairs containing two key phonemes with positional association; calculating the positional distance of the two key phonemes in each group of key phoneme pairs, and determining whether the positional distance is abnormal; if the number of error rows exceeds a set threshold or the positional distance is abnormal, it is determined that the current speech segment is an out-of-set word and the rejection recognition is performed, otherwise it is determined to be a valid command word.

[0005] Further, the step of performing forced alignment between the current speech segment recognition result and the target command word in the command word list to obtain an alignment path comprises the following steps: extracting a standard phoneme path corresponding to the target command word, and obtaining a decoding matrix generated by decoding the current speech segment; based on the CTC criterion, calculating the maximum probability and the source position of each position in the decoding matrix by using the modified forward algorithm, and determining the maximum probability path by backtracking; within a preset recognition window interval, mapping and aligning the maximum probability path and the standard phoneme path of the target command word based on dynamic programming and the Viterbi algorithm to generate an alignment path.

[0006] Further, the step of calculating the maximum probability and the source position of each position in the decoding matrix by using the modified forward algorithm based on the CTC criterion to determine the maximum probability path comprises the following steps: based on the CTC criterion, replacing the probability summation operation in the traditional CTC forward algorithm with a maximum value operation to obtain a modified forward algorithm; using the modified forward algorithm to perform dynamic programming calculation on the decoding matrix to determine the maximum probability of reaching each position; recording the previous time source position corresponding to the maximum probability to form path information; starting from the end of the decoding matrix, backtracking to the starting end according to the path information, matching the continuous path obtained by backtracking with the standard phoneme path of the target command word, and determining the maximum probability path.

[0007] Further, the step of determining a key interval based on the alignment path and determining the alignment result in the interval based on the situation to count the error line count comprises: determining the range of the key interval based on the alignment path; classifying the alignment result in the key interval as a blank phoneme alignment result or a target command word phoneme alignment result; if it is a blank phoneme alignment result, determining whether it occupies a dominant position, if not, the error line count is increased by 1; if it is a target command word phoneme alignment result, determining whether the phoneme occupies a dominant position and whether the non-target command word phoneme does not occupy a dominant position, if any of the determinations is false, the error line count is increased by 1; counting the total number of alignment results in the key interval that do not meet the determination condition as the error line count.

[0008] Further, the step of locating at least one group of key phoneme pairs based on the alignment path and the phoneme sequence features of the target command word, each group of key phoneme pairs containing two key phonemes with a position association comprises: Extract the phoneme sequence features of the target command word and determine the type and position relationship of the key phonemes to be located; Based on the phoneme sequence features, identifying the uniqueness or recurrence attribute of each key phoneme in the sequence; Traversing the alignment path, directly locating the position of the unique key phoneme, and screening and locating the repeated key phoneme by combining the preceding and following characteristic phoneme sequences; The two key phonemes with positional association obtained by positioning are combined to form a set of key phoneme pairs.

[0009] Furthermore, the step of calculating the position distance between two key phonemes in each group of key phoneme pairs and determining whether the position distance is abnormal includes the following steps: Obtaining the positioning positions of the two key phonemes in each group of key phoneme pairs in the alignment path; Based on the positioning positions, calculating the position distance between two key phonemes; Determining the position distance corresponding judgment criteria based on the phoneme sequence characteristics of the target command word; The position distance is compared with the judgment standard, and if it does not meet the judgment standard, the position distance is determined to be abnormal.

[0010] Furthermore, if the error line count exceeds a set threshold or the position distance is abnormal, the step of determining that the current speech segment is an out-of-word and rejecting recognition is performed; otherwise, the step of determining that the current speech segment is a valid command word includes the following steps: Obtaining the error row count and position distance determination results of each key phoneme pair; Determining whether the error row count exceeds a set threshold, or whether the position distance of any key phoneme pair is abnormal; If any of the above conditions is met, the current speech segment is determined to be an out-of-word and a rejection recognition operation is performed; otherwise, it is determined to be a valid command word.

[0011] The second aspect of the present application further provides a device for rejecting recognition of out-of-words, comprising: An alignment module is configured to perform forced alignment between the current speech segment recognition result and the target command word in the command word list to obtain an alignment path. The forced alignment is based on dynamic programming and the Viterbi algorithm, combined with the CTC criterion to align the decoding matrix generated by decoding the speech segment with the standard phoneme path of the target command word. A statistics module is used to determine a key interval based on the alignment path, and to judge the alignment results within the interval by situation, and to count the number of error lines, wherein the proportion of alignment results of blank phonemes or target command word phonemes is judged respectively; The positioning module is configured to position at least one set of key phoneme pairs based on the alignment path and a phoneme sequence feature of the target command word, each set of key phoneme pairs including two key phonemes with a position association; The distance judgment module is configured to calculate a position distance of the two key phonemes in each set of key phoneme pairs, and judge whether the position distance is abnormal. The judgment module is configured to determine that the current speech segment is an out-of-vocabulary word and perform rejection recognition if the error line count exceeds a set threshold or the position distance is abnormal, and otherwise determine that the current speech segment is a valid command word.

[0012] The third aspect of the present application provides an out-of-vocabulary word rejection recognition device, comprising: A processor; A memory for storing a computer program; When the processor executes the computer program, the out-of-vocabulary word rejection recognition method of any one of the above embodiments is implemented.

[0013] The fourth aspect of the present application provides a computer readable storage medium, the computer readable storage medium stores a computer program, and the computer program is executed by a processor to implement the out-of-vocabulary word rejection recognition method of any one of the above embodiments.

[0014] The first aspect of the present application brings the following beneficial effects: The present application uses dynamic programming, Viterbi algorithm and CTC criterion for forced alignment, without the need for retraining the model, adapting to the hardware limitations of low-resource devices, reducing deployment costs, and using key interval alignment judgment and error line statistics to accurately identify similar out-of-vocabulary words with alignment abnormalities. In addition, the positioning of key phoneme pairs and the verification of position distance further lock the position deviation of semantically related phonemes, double-protect against misjudgment, and further simplify the overall process calculation, reduce resource occupation, and efficiently run on low-configuration devices. This greatly reduces the misrecognition rate of similar out-of-vocabulary words, improves the accuracy of voice interaction, and meets the needs of users for precise control of smart homes. BRIEF DESCRIPTION OF DRAWINGS

[0015] Figure 1 is a flowchart of an out-of-vocabulary word rejection recognition method according to an embodiment of the present application; Figure 2 is a structural diagram of an out-of-vocabulary word rejection recognition device according to an embodiment of the present application; Figure 3 is a structural diagram of a computer device according to an embodiment of the present application; The purpose of the present application, functional characteristics and advantages will be further described with reference to the embodiments and the accompanying drawings. DETAILED DESCRIPTION

[0016] In order to make the purposes, technical solutions and advantages of the present application clearer, further detailed description will be made to the present application in combination with the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application, and are not used to limit the present application.

[0017] Those skilled in the art can understand that the singular forms "a", "an" and "the" used herein include plural forms unless specifically stated otherwise. It should be further understood that the use of the term "include" in the specification of the present application means that the presence of a feature, integer, step, operation, element, module and / or assembly, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, modules, assemblies and / or their combinations. It should be understood that when an element is "connected" or "coupled" to another element, it can be directly connected or coupled to the other element, or there can be an intermediate element. In addition, "connected" or "coupled" used herein can include wireless connection or wireless coupling. The phrase "and / or" used herein includes all or any module of the associated list and all combinations.

[0018] Those skilled in the art can understand that, unless otherwise defined, all terms (including technical and scientific terms) used herein have the same meaning as that generally understood by those skilled in the art to which the present application belongs. It should also be understood that terms such as those defined in a general dictionary should be understood to have meanings consistent with those in the context of the prior art, and should not be interpreted in an idealized or overly formal sense unless specifically defined as such.

[0019] Reference Figure 1 The embodiment of the present application provides a method for rejecting an out-of-vocabulary word, comprising: S1: performing forced alignment on a current speech segment recognition result and a target command word in a command word list to obtain an alignment path, wherein the forced alignment is based on dynamic programming and Viterbi algorithm, and the alignment path is obtained by aligning a decoding matrix generated by speech segment decoding with a standard phoneme path of the target command word in combination with CTC criterion; S2: determining a key interval based on the alignment path and determining the alignment result in the interval according to different cases, and counting the number of error rows, wherein when the alignment result is a blank phoneme or a target command word phoneme, the proportion is determined respectively; S3: positioning at least one group of key phoneme pairs based on the alignment path and phoneme sequence features of the target command word, each group of key phoneme pairs containing two key phonemes with position association; S4: calculating the position distance of the two key phonemes in each group of key phoneme pairs, and determining whether the position distance is abnormal; S5: If the error line count exceeds a set threshold or the position distance is abnormal, determining that the current voice segment is an out-of-set word and performing rejection recognition, otherwise determining that it is a valid command word.

[0020] In step S1, the recognition result of the current voice segment is forcibly aligned with the target command word in the command word list by dynamic programming and Viterbi algorithm combined with CTC (Connectionist Temporal Classification) criterion, and finally an alignment path is generated. Taking the target command word "Timing Twenty Minutes" as an example, its standard phoneme path is 'd', 'ing','sh', 'ix', 'ee', 'er','sh', 'ix', 'f', 'en', 'zh', 'ong'. When a user inputs a voice segment (such as "Timing Twenty Five Minutes" which may be misrecognized), the system first decodes the voice segment to generate a decoding matrix, where each element in the matrix represents the probability value of the position belonging to a certain phoneme (vertical axis) at the corresponding time step (horizontal axis). Then, based on the CTC criterion, the traditional forward algorithm is optimized: the original algorithm "sums the probabilities of all possible paths" is changed to "only retains the maximum probability of each position", and the source position of the maximum probability at the previous time step is recorded, which reduces the calculation amount to adapt to low-resource devices and accurately locks the optimal path. Through dynamic programming to traverse the decoding matrix, the maximum probability and path information of each position are determined from the initial time step to the end time step. Then, through the backtracking algorithm, the maximum probability path that best matches the standard phoneme path of "Timing Twenty Minutes" is obtained by tracing back from the end of the matrix to the beginning. Finally, within the preset recognition window interval, the maximum probability path is forcibly mapped to the standard phoneme path by Viterbi algorithm to determine whether each time step corresponds to a certain phoneme (such as 'd', 'ing', etc.) or blank phoneme (blank phoneme without actual semantics) in the standard phoneme path, forming a complete alignment path. This step is based on the existing model output and does not need to be retrained, which is suitable for low-resource devices. Through the combination of dynamic programming and Viterbi algorithm, the accuracy of alignment is guaranteed while the efficiency is improved, providing reliable data for subsequent judgment and laying the foundation for rejection recognition.

[0021] In step S2, based on the alignment path generated in step S1, the key interval determination and error line count statistics are completed, which is the core verification link of out-of-set word determination. First, the key interval needs to be determined: according to the phoneme sequence characteristics, the subsequence carrying the key semantics "Twenty" corresponding to "ee, er, sh, ix" and "Minutes" corresponding to "f, en, zh, ong" are extracted, and then the continuous column interval in the alignment path corresponding to the two sub-sequences (assuming that the 5th-8th column and the 9th-12th column of the alignment path) are located, which is the key interval. Subsequently, the alignment results in the key interval are determined on a case-by-case basis: if the alignment result of a column is a blank phoneme, it is determined whether the probability of the phoneme accounts for ≥60% of the total probability of the column (dominant position), and if not, it is recorded as 1 error line; if the alignment result is a target command word phoneme (such as “ee”), it is necessary to simultaneously satisfy the conditions that the phoneme probability accounts for ≥60% and the highest probability of non-target phonemes accounts for ≤30%, and if either condition is not met, it is recorded as 1 error line. For example, if the alignment result of the 6th column in the interval corresponding to the “20” subsequence is “er”, but the phoneme probability is only 55% (lower than 60%), this column is counted as an error line. Finally, the total number of error lines in the key interval is counted, providing a quantitative basis for subsequent determination. By focusing on the key semantic interval, the interference of non-core areas is reduced, and the accuracy of determination is improved; the case-by-case proportion determination rule realizes the quantitative capture of alignment abnormalities, provides reliable data support for out-of-vocabulary recognition, and adapts to the computing power of low-resource devices.

[0022] In step S3, step S3 locates at least one group of key phonemes based on the alignment path generated in step S1 and the phoneme sequence features of the target command word, providing a basis for subsequent position distance verification. First, analyze the phoneme sequence features to determine key phoneme pairs with positional correlation: “ix” (the eighth phoneme) in “20” and “f” (the ninth phoneme) in “minute” have semantic correlation, the former is the core phoneme of the time value, and the latter is the starting phoneme of the time unit, forming a group of key phoneme pairs (ix, f). In the positioning process, for phonemes like “ix” that appear repeatedly in the path (at the fourth and eighth positions, respectively), distinguish them based on their features before and after: the fourth “ix” corresponds to “time”, and its previous phoneme is “sh”; the eighth “ix” corresponds to “20”, and its previous phoneme is “sh” and there is “ee, er” in the more front position, so by traversing the alignment path, find the “ix” after “ee, er, sh” to determine the position of the target “ix”. For unique phonemes like “f”, directly traverse the alignment path to find their corresponding positions. Finally, lock the specific column index of “ix” and “f” in the alignment path, complete the positioning of this group of key phoneme pairs, and simultaneously locate other related phoneme pairs as needed (such as “ix” of “time” and “er” of “two”). By accurately positioning the key phoneme pairs, the positional correlation features between phonemes are captured, providing a core basis for distinguishing similar out-of-vocabulary words; the positioning method based on phoneme sequence features improves the accuracy of key phoneme recognition and adapts to various command word scenarios.

[0023] In step S4, the core of step S4 is to calculate the position distance of the key phoneme pair located in the alignment path in step S3, and to determine whether the distance is abnormal according to the standard phoneme sequence characteristics of the target command word, which is a key check link for distinguishing valid command words from out-of-set words. First, the reference information of the phoneme pair in the standard phoneme sequence needs to be clarified: the standard phoneme path of "twenty minutes" is 'd', 'ing','sh', 'ix', 'ee', 'er','sh', 'ix', 'f', 'en', 'zh', 'ong', wherein the ix of "twenty" corresponds to the 8th phoneme (index 7, counted from 0), and the f of "minutes" corresponds to the 9th phoneme (index 8). The difference in the number of phonemes in the standard sequence is 1, so the position distance determination threshold is 1±2, that is, the effective range is -1 to 3 (the actual position distance is a non-negative value, so the effective range is 0 to 3); Then the positioning position of the key phoneme pair in the alignment path is obtained: assuming that through step S3, the ix of "twenty" corresponds to the 15th column (time step) and the f corresponds to the 18th column (time step) in the alignment path, the position distance of the two is calculated as 18-15=3. Comparing the actual distance with the determination threshold (0-3), 3 is within the effective range, and it is determined that the position distance of the key phoneme pair is normal. If the user input is the out-of-set word "twenty-five minutes", the phonetic segment processed by steps S1-S3, the ix of "twenty-five" in the alignment path may correspond to the 14th column, and the f corresponds to the 20th column, with a position distance of 6, which is outside the effective range of 0-3. At this time, it is determined that the position distance of the key phoneme pair is abnormal. If there are multiple groups of key phoneme pairs (such as the ix of "time" and the er of "two"), the position distance of each group needs to be calculated and judged one by one. The position distance is determined by a dynamic threshold (based on the difference in the standard phoneme sequence ±2), which takes into account the difference in the phoneme sequence of the command word and the accuracy of the check. The quantitative distance check method can accurately capture the phoneme position deviation between out-of-set words and valid command words, providing an objective basis for subsequent rejection, and the calculation logic is simple and suitable for low-resource devices.

[0024] In step S5, step S5 is the final determination link of the out-of-set word rejection, which determines whether the current phonetic segment is an out-of-set word by comprehensively considering the error line count of step S2 and the position distance judgment result of step S4. Taking the target command word "twenty minutes" as an example, the key interval includes the 8-column alignment path corresponding to "twenty" and "minutes" (columns 5-12). According to the rule, the length of the phoneme sub-sequence is ≤5, and the error line count threshold is set to 50% of the total number of columns in the key interval, i.e. 8x50%=4 lines. If the user inputs a valid command word "twenty minutes", after step S2, it is determined that the error line count in the key interval is 1 (which does not exceed the threshold of 4). In step S4, the position distance of the key phoneme pair (ix, f) is 3 (which is within the effective range of 1±2), so it is determined to be a valid command word, and the device performs the timing operation; If the user inputs the out-of-set word "set time twenty-five minutes", in step S2, the error line count in the key interval reaches 5 (exceeding the threshold value 4) due to the phoneme difference between "twenty-five" and "twenty"; meanwhile, the position distance of (ix, f) in step S4 is 6 (exceeding the range of 1±2), satisfying any abnormal condition, and the out-of-set word is determined and the recognition is rejected; if another out-of-set word "set time twenty minutes" has an error line count of 3 (not exceeding the threshold value), but the position distance of the key phoneme pair (ix, f) is 5 (exceeding the range), it is still determined as an out-of-set word. Only when the error line count does not exceed the threshold value and the position distance of all key phoneme pairs is normal, it is confirmed to be valid. Through the double determination conditions (error line count + position distance), the recognition accuracy of the out-of-set word is greatly improved; the dynamic threshold value adapts to the characteristics of different command words, and takes into account flexibility and rigor, and the logic is simple, which is suitable for efficient operation of low-resource devices.

[0025] In one embodiment, the step of performing forced alignment between the current speech segment recognition result and the target command word in the command word list to obtain an alignment path comprises the following steps: S10, extracting a standard phoneme path corresponding to the target command word, and obtaining a decoding matrix generated by decoding the current speech segment; S11, based on the CTC criterion, using the modified forward algorithm to calculate the maximum probability and the source position of each position of the decoding matrix, and determining the maximum probability path by backtracking; S12, within a preset recognition window interval, based on dynamic programming and Viterbi algorithm, mapping and aligning the maximum probability path with the standard phoneme path of the target command word to generate an alignment path.

[0026] In this embodiment, in the scenario of out-of-set word rejection recognition for "timing for twenty minutes," this embodiment completes forced alignment and generates an alignment path in three steps. First, the standard phoneme path of the target command word "timing for twenty minutes" is extracted: 'd', 'ing', 'sh', 'ix', 'ee', 'er', 'sh', 'ix', 'f', 'en', 'zh', 'ong'. At the same time, a decoding matrix is ​​generated by decoding the current speech segment (e.g., "timing for twenty-five minutes" spoken by the user). The horizontal axis of the matrix represents time steps (assuming 20 time steps) and the vertical axis represents phonemes (including all possible phonemes and blank). The value of each cell represents the recognition probability of the corresponding phoneme at that time step. Secondly, the forward algorithm is modified based on the CTC criterion: the traditional forward algorithm calculates the sum of probabilities at each position, but this algorithm calculates the maximum value instead, and records the position of the maximum value in the previous time step. Taking the 5th time step of the decoding matrix as an example, if the phoneme probability of "ee" is 0.8, which is the maximum value, and it comes from "ix" at the 4th time step, then record the maximum probability of 0.8 at that position and the source position 4. After traversing the entire matrix, backtrack from the last time step (step 20) to step 1, trace the source of the maximum probability at each position, and determine the maximum probability path that best matches the standard phoneme path (for example, time steps 1-3 correspond to 'd', 'ing', 'sh', time steps 4-8 correspond to 'ix', 'ee', 'er', 'sh', 'ix', etc.).

[0027] Finally, within a preset recognition window (e.g., time steps 1-20), the maximum probability path is aligned with the standard phoneme path using a combination of dynamic programming and the Viterbi algorithm. Dynamic programming calculates path matching, while the Viterbi algorithm optimizes path selection, ultimately determining which phoneme or blank corresponds to each time step in the standard phoneme path (e.g., time step 5 corresponds to 'ee', time step 9 corresponds to 'f'). This generates an alignment path containing the mapping between time steps and phonemes, providing accurate data for subsequent key interval determination and phoneme location. Through step-by-step extraction, optimized calculation, and precise mapping, alignment paths are efficiently generated on low-resource devices. A modified forward algorithm reduces computational effort, while the combination of dynamic programming and the Viterbi algorithm improves alignment accuracy, providing a reliable foundation for subsequent rejection recognition.

[0028] In one embodiment, the step of calculating the maximum probability and source position of each position of the decoding matrix using a modified forward algorithm based on the CTC criterion and determining the maximum probability path by backtracking includes the following steps: S110, based on the CTC criterion, replacing the probability summation operation in the traditional CTC forward algorithm with a maximum value operation to obtain a modified forward algorithm; S111, using the modified forward algorithm to perform dynamic programming calculation on the decoding matrix to determine the maximum probability of reaching each position; S112, record the previous time source position corresponding to the maximum probability to form path information; S113, start from the end of the decoding matrix, backtrack to the starting end according to the path information, match the coherent path obtained by backtracking with the standard phoneme path of the target command word, and determine the maximum probability path.

[0029] In the embodiment, first, the forward algorithm is modified based on the CTC criterion: the traditional CTC forward algorithm sums all possible path probabilities at each position, and here the maximum value is obtained, only the maximum probability of a single path is retained (such as phoneme x at time step t, only the highest probability in all paths reaching this position is recorded), which greatly reduces the calculation load of low-resource devices; second, the modified forward algorithm is used to dynamically plan the decoding matrix: taking a decoding matrix of “timing twenty minutes” (20 time steps) as an example, starting from time step 1, the maximum probability of each position is calculated in turn. For example, when time step 4 corresponds to the “ix” phoneme, compared with all previous paths (time step 3 “sh” or blank) that can reach this position, if the probability from time step 3 “sh” is the maximum, that is, 0.7, then the maximum probability of time step 4 “ix” is recorded as 0.7. In this way, the maximum probability calculation of the whole matrix is completed by traversing to time step 20. Then, record the previous time source position of each maximum probability: continuing the above example, the maximum probability of time step 4 “ix” comes from time step 3 “sh”, so the source position is recorded as 3; if the maximum probability of time step 5 “ee” comes from time step 4 “ix”, the source position is recorded as 4, forming a complete path information chain; finally, backtracking from the end of the decoding matrix (time step 20): first, determine the maximum probability phoneme at time step 20 (such as “ong”), find its source as “zh” at time step 19 through the path information, and then trace back to “en” at time step 18 … until the starting end (time step 1). Match the path obtained by backtracking (such as 1→2→3→4→…→20 corresponding to 'd'→'ing'→'sh'→'ix'→…→'ong') with the standard phoneme path of “timing twenty minutes”, and finally determine that this path is the maximum probability path. The modified forward algorithm reduces the amount of calculation and adapts to low-resource devices; the combination of dynamic planning and backtracking accurately locks the maximum probability path matched with the standard phoneme path, providing high-reliability basic data for subsequent alignment.

[0030] In one embodiment, the step of determining a key interval based on the alignment path and determining the alignment result in the interval according to different cases and counting the number of error rows comprises: S20, determining the range of the key interval based on the alignment path; S21, classifying the alignment result in the key interval as a blank phoneme alignment result or a target command word phoneme alignment result; S22, if it is a blank phoneme alignment result, determine whether it is in the dominant position, if not, the error line count is incremented by 1; if it is a target command word phoneme alignment result, determine whether the phoneme is in the dominant position and whether the non-target command word phoneme is not in the dominant position, if either condition is not met, the error line count is incremented by 1; S23, count the total number of alignment results in the key interval that do not meet the determination condition as the error line count.

[0031] In this embodiment, first, the key interval range is determined based on the alignment path generated in step S1. Taking the target command word "set the time for 20 minutes" as an example, the "ee, er, sh, ix" corresponding to "20" and "f, en, zh, ong" corresponding to "minutes" in the standard phoneme path are sub-sequences carrying key semantics, which are mapped to columns 5-8 and 9-12 (a total of 8 columns) in the alignment path. This continuous column interval is the key interval; second, the alignment results in the key interval are classified. Traverse columns 5-12, if the alignment result of a column is a blank phoneme (no actual semantics), it is classified as a blank phoneme alignment result; if the alignment result is a target command word phoneme such as "ee", "er", "f", etc., it is classified as a target command word phoneme alignment result, for example, the alignment result of column 5 is "ee", which belongs to the target command word phoneme alignment result; the alignment result of column 7 is blank, which belongs to the blank phoneme alignment result; Then, the two types of alignment results are determined and the error lines are counted. For blank phoneme alignment results (such as column 7), calculate the proportion of its probability in the total probability of all phonemes in the column, if the proportion is 55% (<60%), it is determined that it does not occupy the dominant position, and the error line count is incremented by 1. For target command word phoneme alignment results (such as "ee" in column 5), two conditions must be met: the "ee" probability ratio is ≥60% and the highest probability ratio of non-target phonemes is ≤30%. If the "ee" probability ratio is 58% (<60%) or the non-target phoneme "eh" probability ratio is 35% (>30%), the error line count is incremented by 1 if either condition is not met; finally, count the total number of alignment results in the key interval that do not meet the determination condition. Assuming that columns 5, 7, and 10 do not meet the conditions, the error line count is 3, which will be used as an important basis for subsequent out-of-set word determination. By determining the key interval and classifying and determining the alignment results in steps, the semantic core area is accurately focused; the quantitative proportion judgment rule ensures that the error line count is objective and reliable, providing accurate data for out-of-set word recognition and adapting to the operation requirements of low-resource devices.

[0032] In an embodiment, the step of locating at least one group of key phoneme pairs based on the alignment path and the phoneme sequence features of the target command word, each group of key phoneme pairs containing two key phonemes with a positional association, comprises: S30, extract the phoneme sequence feature of the target command word, determine the type and position correlation of the key phonemes to be positioned; S31, based on the phoneme sequence feature, identify the uniqueness or repeated occurrence attribute of each key phoneme in the sequence; S32, traverse the alignment path, directly position the position of the unique key phoneme, and screen and position the repeated key phoneme combined with the front and rear feature phoneme sequence; S33, combine the two key phonemes with position correlation obtained by positioning to form a group of key phoneme pairs.

[0033] In this embodiment, first, the phoneme sequence feature of the target command word is extracted, and the type and position correlation of the key phonemes are determined. The standard phoneme path of "twenty minutes" is 'd', 'ing','sh', 'ix', 'ee', 'er','sh', 'ix', 'f', 'en', 'zh', 'ong'. Extracting the phoneme sequence feature can know that "twenty" is the time value core, corresponding to the phonemes "ee, er, sh, ix"; "minutes" is the time unit core, corresponding to the phonemes "f, en, zh, ong". From the semantic logic, "ix" (the eighth phoneme) of "twenty" is the semantic preposition component of "f" (the ninth phoneme) of "minutes" - only when "twenty" is clear, the unit of "minutes" has meaning, and the two constitute a position correlation relationship. It is determined that the type of key phoneme to be positioned is "value core phoneme (ix)" and "unit start phoneme (f)"; Then, based on the phoneme sequence feature, the uniqueness attribute of the key phoneme is identified, and the standard phoneme path is traversed to find that "f" appears only once (at the ninth position) in the entire sequence, which is a unique key phoneme; "ix" appears twice (the fourth position corresponds to "time", and the eighth position corresponds to "twenty"), which is a repeated key phoneme. Secondly, the key phoneme position is positioned by traversing the alignment path. For the unique key phoneme "f", directly traverse the alignment path (assuming 20 columns of time steps) generated in step S1 to find the column index (such as the 18th column) corresponding to "f". For the repeated key phoneme "ix", combined with the front and rear feature phoneme sequence screening: the front phoneme of the fourth "ix" is "sh" (corresponding to "time"), without "ee, er"; the front phoneme of the eighth "ix" is "sh", and the more forward column (such as the 12th-13th column) corresponds to "ee, er", so when traversing the alignment path, the column (such as the 15th column) corresponding to "ix" is found after "ee, er, sh" (such as the 12th-14th column), and it is determined as the position of "ix" corresponding to "twenty"; Finally, the key phoneme pairs are formed by combining the key phonemes with position association, the located "ix" (column 15) is combined with "f" (column 18) to obtain a set of key phoneme pairs (ix, f); if the verification dimension needs to be increased, the located "ix" of "time" (such as column 8) and the located "er" of "two" (such as column 13) can also be combined to form another set of key phoneme pairs (ix, er), and the positioning of the key phoneme pairs is completed. The phoneme association relationship is determined through semantic logic to ensure that the association of the key phoneme pairs conforms to the semantic rules of the command words; the position accuracy of the key phonemes is improved by combining the unique attribute differentiation positioning; and the multi-dimensional basis for subsequent verification is provided by the multi-group phoneme pair positioning, which adapts to the efficient operation requirements of low-resource devices.

[0034] In one embodiment, the step of calculating the position distance between the two key phonemes in each set of key phoneme pairs and determining whether the position distance is abnormal includes the following steps: S40, obtaining the positioning positions of the two key phonemes in each set of key phoneme pairs in the alignment path; S41, calculating the position distance between the two key phonemes based on the positioning positions; S42, determining the judgment standard corresponding to the position distance according to the phoneme sequence characteristics of the target command word; S43, comparing the position distance with the judgment standard, and determining that the position distance is abnormal if it does not conform to the judgment standard.

[0035] In this embodiment, first, the positioning positions of each set of key phoneme pairs in the alignment path are obtained, taking the target command word "time twenty minutes" and its key phoneme pair (ix, f) as an example, and combining the positioning results of step S3 in the foregoing, it can be known that the ix corresponding to "twenty" is mapped to the column index 15 (i.e., the 15th time step) in the alignment path, and the f corresponding to "minute" is mapped to the column index 18 (i.e., the 18th time step) in the alignment path; if there is another set of key phoneme pairs (ix_time, er) (the ix of "time" and the er of "two"), the positioning column index of the ix of "time" is 8, and the positioning column index of the er of "two" is 13, and the specific positioning positions of the two phonemes of each set of key phoneme pairs need to be recorded respectively; second, the position distance between the two key phonemes is calculated based on the positioning positions, for the key phoneme pair (ix, f), the positioning column index of the latter phoneme is subtracted from the positioning column index of the former phoneme, i.e., 18-15=3, and the position distance of this set of key phoneme pairs is obtained as 3; for the key phoneme pair (ix_time, er), 13-8=5 is calculated, and the position distance thereof is obtained as 5, and in the calculation process, it needs to be ensured that the index order is consistent with the semantic order of the phonemes in the command word (such as "ix" first and "f" second, "ix_time" first and "er" second), so as to avoid distance calculation errors caused by the order being reversed.

[0036] Then, the position distance determination criterion is determined according to the phoneme sequence feature of the target command word. First, the standard phoneme path of "determine twenty minutes" is checked ('d', 'ing','sh', 'ix', 'ee', 'er','sh', 'ix', 'f', 'en', 'zh', 'ong'), and the position index of each group of key phoneme pairs in the standard sequence is determined: the ix of "twenty" corresponds to index 7, the f of "minutes" corresponds to index 9, and the index difference is 9-7=2. According to the rule determination criterion, the difference is ±2, that is, 2-2=0 to 2+2=4, and the effective range of the position distance of this group of key phoneme pairs is 0-4; the ix of "time" corresponds to index 3, and the er of "two" corresponds to index 5. The index difference is 5-3=2, and the determination criterion is also 0-4, and the effective range is also 0-4. Finally, the calculated position distance is compared with the determination criterion to determine whether it is abnormal. For the key phoneme pair (ix, f), the position distance 3 is within the effective range of 0-4, and the distance is determined to be normal; if the user inputs the out-of-set word "determine twenty-five minutes", the position distance of the key phoneme pair (ix, f) is calculated as 6, which is outside the range of 0-4, and the distance is determined to be abnormal; for the key phoneme pair (ix_time, er), the position distance 5 is outside the effective range of 0-4, and the distance is also determined to be abnormal. As long as there is a group of key phoneme pairs with abnormal distance, the speech segment is marked as a suspicious out-of-set word. Through the steps of obtaining the position, calculating the distance, determining the criterion and comparing and judging, the position relationship of the key phonemes is accurately checked; the determination threshold is dynamically set based on the standard phoneme sequence, the feature differences of different command words are considered, the accuracy of abnormal judgment is improved, and the lightweight operation demand of low-resource devices is adapted.

[0037] In one embodiment, if the error line count exceeds a set threshold or the position distance is abnormal, the current speech segment is determined to be an out-of-set word and the recognition is rejected, otherwise it is determined to be a valid command word. The step includes the following steps: S50, obtaining the error line count and the position distance determination result of each key phoneme pair; S51, determining whether the error line count exceeds a set threshold or the position distance of any key phoneme pair is abnormal; S52, if any of the above conditions is met, the current speech segment is determined to be an out-of-set word and the recognition is rejected, otherwise it is determined to be a valid command word.

[0038] In the embodiment, firstly, the error line count and the position distance determination result of each key phoneme pair are obtained. Taking the target command word "timing twenty minutes" as an example, the key interval is 8 columns (5th-12th columns), according to the rule that the length of the phoneme subsequence is less than or equal to 5, the error line count threshold is set to 8*50% = 4 lines, and it is assumed that the error line count is 3 after step S2 statistics; after step S4 determination, the position distance of the key phoneme pair (ix, f) is 3 (within the effective range of 0-4, normal), and the position distance of the other key phoneme pair (ix_time, er) is 5 (out of the range of 0-4, abnormal), which needs to be summarized as the determination basis; secondly, whether the error line count exceeds the set threshold or whether the position distance of any key phoneme pair is abnormal is determined. For the error line count, 3 does not exceed the threshold 4, and this condition is not met. For the position distance, the key phoneme pair (ix_time, er) is determined to be abnormal, and the condition of "the position distance of any key phoneme pair is abnormal" is met. If the user inputs the valid command word "timing twenty minutes", the error line count is 1 (does not exceed the threshold), and the position distances of all key phoneme pairs are within the effective range (such as (ix, f) distance 3, (ix_time, er) distance 4), both conditions are not met, and if the input is the out-of-set word "timing twenty-eight minutes", the error line count is 5 (exceeds the threshold 4), and the key phoneme pair (ix, f) distance 6 (abnormal), both conditions are met. Then, corresponding operations are performed according to the determination result. When any one of the conditions is met, it is determined that the current speech segment is an out-of-set word and recognition is refused. For example, in the above example of "timing twenty minutes", because there is a key phoneme pair with abnormal position distance, it is determined that it is an out-of-set word, and the device does not respond. If both conditions are not met (such as a valid command word), it is determined that it is a valid command word, and the device performs a timing operation. If only the error line count exceeds the threshold (such as error line count 5 and all position distances are normal), it is still determined to be an out-of-set word and recognition is refused. This logic ensures that as long as there is a key semantic deviation or phoneme position abnormality, the out-of-set word can be effectively filtered. Through double condition determination, both semantic deviation and position abnormality of out-of-set words are covered, which greatly improves the recognition accuracy. The dynamic threshold adapts to different command word characteristics, and the flexibility and rigor are considered. The logic is simple and suitable for low-resource devices to run efficiently.

[0039] Reference Figure 2 A device for refusing recognition of an out-of-set word, comprising: An alignment module 100 is configured to perform forced alignment on a current speech segment recognition result and a target command word in a command word list to obtain an alignment path. The forced alignment is based on dynamic programming and Viterbi algorithm, and combines CTC criterion to align the decoding matrix generated by speech segment decoding with the standard phoneme path of the target command word. The statistical module 200 is configured to determine a key interval based on the alignment path, and determine a case of the alignment result in the interval, and count an error line, wherein the alignment result is a blank phoneme or a target command phoneme, and the proportion judgment is performed respectively; The positioning module 300 is configured to position at least one group of key phoneme pairs based on the alignment path and a phoneme sequence feature of the target command, wherein each group of key phoneme pairs comprises two key phonemes with a position association; The distance judgment module 400 is configured to calculate a position distance of the two key phonemes in each group of the key phoneme pairs, and judge whether the position distance is abnormal. The judgment module 500 is configured to judge that the current speech segment is an out-of-set word and perform rejection recognition if the error line count exceeds a set threshold or the position distance is abnormal, and otherwise, judge that it is a valid command word.

[0040] Further, the alignment module 100 comprises: The standard phoneme and decoding matrix acquisition unit is configured to extract a standard phoneme path corresponding to the target command, and acquire a decoding matrix generated by decoding the current speech segment; The maximum probability path determination unit is configured to calculate a maximum probability and a source position of each position of the decoding matrix based on a CTC criterion by using a modified forward algorithm, and determine a maximum probability path by backtracking. The alignment path generation unit is configured to map and align the maximum probability path and the standard phoneme path of the target command based on dynamic programming and Viterbi algorithm in a preset recognition window interval, and generate an alignment path.

[0041] Further, the maximum probability path determination unit comprises: The modified forward algorithm generation unit is configured to replace a probability summation operation in a traditional CTC forward algorithm with a maximum value operation based on a CTC criterion, and obtain a modified forward algorithm. The matrix maximum probability calculation unit is configured to perform dynamic programming calculation on the decoding matrix by using the modified forward algorithm, and determine a maximum probability of reaching each position. The path information recording unit is configured to record a previous time source position corresponding to the maximum probability, and form path information. The maximum probability path matching unit is configured to start from an end of the decoding matrix, backtrack to a start end according to the path information, match a coherent path obtained by backtracking with the standard phoneme path of the target command, and determine the maximum probability path.

[0042] Further, the statistical module 200 comprises: The key interval range determination unit is configured to determine a range of the key interval based on the alignment path. an alignment result classification unit, configured to classify the alignment result in the key interval as a blank phoneme alignment result or a target command word phoneme alignment result; an alignment result judgment and error line count unit, configured to, if the alignment result is the blank phoneme alignment result, judge whether the blank phoneme alignment result is dominant, and if not, increase the error line count by 1; if the alignment result is the target command word phoneme alignment result, judge whether the target command word phoneme is dominant and whether the non-target command word phoneme is not dominant, and if either of the judgments is no, increase the error line count by 1; an error line total number statistics unit, configured to count the total number of alignment results in the key interval that do not meet the judgment condition as the error line count.

[0043] Further, the positioning module 300 described above comprises: a phoneme feature and association relationship determination unit, configured to extract a phoneme sequence feature of a target command word, and determine a type and a position association relationship of a key phoneme to be positioned; a key phoneme attribute recognition unit, configured to recognize, based on the phoneme sequence feature, a unique or repeated attribute of each key phoneme in the sequence; a key phoneme position positioning unit, configured to traverse the alignment path, directly position the position of a unique key phoneme, and combine a preceding and a following feature phoneme sequence to filter and position a repeated key phoneme; a key phoneme pair combination unit, configured to combine two key phonemes with a position association obtained by positioning to form a key phoneme pair.

[0044] Further, the distance judgment module 400 described above comprises: a key phoneme positioning position acquisition unit, configured to acquire a positioning position of two key phonemes in each key phoneme pair in the alignment path; a position distance calculation unit, configured to calculate a position distance between the two key phonemes based on the positioning positions; a judgment standard determination unit, configured to determine a judgment standard corresponding to the position distance according to a phoneme sequence feature of a target command word; a position distance abnormality judgment unit, configured to compare the position distance with the judgment standard, and if the position distance does not meet the judgment standard, judge that the position distance is abnormal.

[0045] Further, the judgment module 500 described above comprises: a judgment basis acquisition unit, configured to acquire the error line count and a position distance judgment result of each key phoneme pair; a judgment condition judgment unit, configured to judge whether the error line count exceeds a set threshold value, or whether the position distance of any key phoneme pair is abnormal; The out-of-set word and valid command word determination unit is configured to determine that the current speech segment is an out-of-set word and perform a rejection operation if any of the conditions is met, and otherwise determine that the current speech segment is a valid command word.

[0046] Referring to Figure 3 In the embodiments of the present application, a computer device is also provided, which can be a server, and the internal structure of the computer device can be as shown in Figure 3 The computer device includes a processor, a memory, a network interface and a database connected through a bus. The processor of the computer device is configured to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores operations, a computer program and a database. The internal memory provides an environment for the operation and computer program in the non-volatile storage medium. The database of the computer device is configured to store data such as the rejection method of the out-of-set word. The network interface of the computer device is configured to communicate with an external terminal through a network connection. The computer program is executed by the processor to implement a rejection method of an out-of-set word, including: performing forced alignment on a current speech segment recognition result and a target command word in a command word list to obtain an alignment path, wherein the forced alignment is based on dynamic programming and Viterbi algorithm, and the alignment is performed between a decoding matrix generated by speech segment decoding and a standard phoneme path of the target command word based on CTC criteria; determining a key interval based on the alignment path and determining the alignment result in the interval based on different cases, and counting the number of error rows, wherein the alignment result is blank phonemes or target command word phonemes, and the proportion is determined respectively; based on the alignment path and the phoneme sequence features of the target command word, locating at least one group of key phoneme pairs, each group of key phoneme pairs including two key phonemes with position association; calculating the position distance of the two key phonemes in each group of key phoneme pairs, and determining whether the position distance is abnormal; if the number of error rows exceeds a set threshold or the position distance is abnormal, determining that the current speech segment is an out-of-set word and performing a rejection operation, otherwise determining that the current speech segment is a valid command word.

[0047] The embodiment of the present application further provides a computer readable storage medium, which stores a computer program, and the computer program is executed by a processor to implement a method for rejecting an out-of-set word, comprising: performing forced alignment on a current speech segment recognition result and a target command word in a command word list to obtain an alignment path, wherein the forced alignment is based on dynamic programming and Viterbi algorithm, and a decoding matrix generated by speech segment decoding is aligned with a standard phoneme path of the target command word in combination with a CTC criterion; determining a key interval based on the alignment path and determining the alignment result in the interval based on a case, and counting an error line count, wherein the alignment result is a blank phoneme or a target command word phoneme, and a proportion judgment is performed respectively; positioning at least one group of key phoneme pairs based on the alignment path and a phoneme sequence feature of the target command word, each group of key phoneme pairs comprising two key phonemes with a position correlation; calculating a position distance of the two key phonemes in each group of the key phoneme pairs, and determining whether the position distance is abnormal; if the error line count exceeds a set threshold or the position distance is abnormal, determining that the current speech segment is an out-of-set word and performing rejection recognition, otherwise, determining that it is a valid command word.

[0048] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by a computer program instructing related hardware, and the computer program can be stored in a computer readable storage medium. When the computer program is executed, the processes of the above-mentioned embodiments can be included. Any reference to memory, storage, database or other medium provided by the present application and used in the embodiments can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. As an illustration but not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (SSRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM) and memory bus dynamic RAM (RDRAM) and the like.

[0049] The above only describes the preferred embodiments of the present application, and does not limit the patent scope of the present application. Any equivalent structure or equivalent process transformation based on the content of the specification and drawings, or direct or indirect application in other related technical fields, is also included in the patent protection scope of the present application.

Claims

1. A method for rejecting out-of-word recognition, characterized in that: include: Performing forced alignment on the current speech segment recognition result and the target command word in the command word list to obtain an alignment path, wherein the forced alignment is based on dynamic programming and the Viterbi algorithm, combined with the CTC criterion to align the decoding matrix generated by decoding the speech segment with the standard phoneme path of the target command word; Based on the alignment path, a key interval is determined and the alignment results within the interval are judged by situation, and the number of error lines is counted. In particular, the proportion of alignment results of blank phonemes or target command word phonemes is judged respectively; Based on the alignment path and the phoneme sequence features of the target command word, at least one group of key phoneme pairs is located, each group of key phoneme pairs includes two key phonemes with positional association; Calculating the position distance between two key phonemes in each group of the key phoneme pairs, and determining whether the position distance is abnormal; If the error line count exceeds a set threshold or the position distance is abnormal, the current speech segment is determined to be an out-of-set word and rejection recognition is performed; otherwise, it is determined to be a valid command word.

2. The method for rejecting out-of-word recognition according to claim 1, characterized in that: The step of forcibly aligning the current speech segment recognition result with the target command word in the command word list to obtain an alignment path includes the following steps: Extracting the standard phoneme path corresponding to the target command word, and obtaining a decoding matrix generated by decoding the current speech segment; Based on the CTC criterion, a modified forward algorithm is used to calculate the maximum probability and source position of each position in the decoding matrix, and the maximum probability path is determined by backtracking; Within a preset recognition window, based on dynamic programming and the Viterbi algorithm, the maximum probability path is mapped and aligned with the standard phoneme path of the target command word to generate an aligned path.

3. The method for rejecting out-of-word recognition according to claim 2, characterized in that: The step of calculating the maximum probability and source position of each position of the decoding matrix using a modified forward algorithm based on the CTC criterion and determining the maximum probability path by backtracking includes the following steps: Based on the CTC criterion, the probability summation operation in the traditional CTC forward algorithm is replaced by the maximum value operation to obtain a modified forward algorithm; Performing dynamic programming calculations on the decoding matrix using the modified forward algorithm to determine the maximum probability of reaching each position; Recording the source position at the previous time corresponding to the maximum probability to form path information; Starting from the end of the decoding matrix, the path information is traced back to the starting end, and the coherent path obtained by tracing back is matched with the standard phoneme path of the target command word to determine the maximum probability path.

4. The method for rejecting out-of-word recognition according to claim 1, characterized in that: The step of determining a critical interval based on the alignment path, judging the alignment results within the interval by situation, and counting the number of error rows includes: Determining a range of a critical interval based on the alignment path; Classify the alignment results in the key interval into blank phoneme alignment results or target command word phoneme alignment results; If it is a blank phoneme alignment result, determine whether it occupies a dominant position. If not, the error row count is increased by 1. If it is a target command word phoneme alignment result, determine whether the phoneme occupies a dominant position and whether the non-target command word phoneme does not occupy a dominant position. If any of the judgments is no, the error row count is increased by 1. The total number of alignment results that do not meet the judgment conditions in the key interval is counted as the error row count.

5. The method for rejecting out-of-word recognition according to claim 1, characterized in that: The step of locating at least one group of key phoneme pairs based on the alignment path and the phoneme sequence features of the target command word, each group of key phoneme pairs including two key phonemes with positional association, includes: Extract the phoneme sequence features of the target command word and determine the type and position relationship of the key phonemes to be located; Based on the phoneme sequence features, identifying the uniqueness or recurrence attribute of each key phoneme in the sequence; Traversing the alignment path, directly locating the position of the unique key phoneme, and screening and locating the repeated key phoneme by combining the preceding and following characteristic phoneme sequences; The two key phonemes with positional association obtained by positioning are combined to form a set of key phoneme pairs.

6. The method for rejecting out-of-word recognition according to claim 1, characterized in that: The step of calculating the position distance between two key phonemes in each group of key phoneme pairs and determining whether the position distance is abnormal comprises the following steps: Obtaining the positioning positions of the two key phonemes in each group of key phoneme pairs in the alignment path; Based on the positioning positions, calculating the position distance between two key phonemes; Determining the position distance corresponding judgment criteria based on the phoneme sequence characteristics of the target command word; The position distance is compared with the judgment standard, and if it does not meet the judgment standard, the position distance is determined to be abnormal.

7. The method for rejecting out-of-word recognition according to claim 1, characterized in that: If the error line count exceeds a set threshold or the position distance is abnormal, the step of determining that the current speech segment is an out-of-set word and rejecting recognition is performed; otherwise, the step of determining that the current speech segment is a valid command word includes the following steps: Obtaining the error row count and position distance determination results of each key phoneme pair; Determining whether the error row count exceeds a set threshold, or whether the position distance of any key phoneme pair is abnormal; If any of the above conditions is met, the current speech segment is determined to be an out-of-word and a rejection recognition operation is performed; otherwise, it is determined to be a valid command word.

8. A device for rejecting out-of-word recognition, characterized in that: include: An alignment module is configured to perform forced alignment between the current speech segment recognition result and the target command word in the command word list to obtain an alignment path. The forced alignment is based on dynamic programming and the Viterbi algorithm, combined with the CTC criterion to align the decoding matrix generated by decoding the speech segment with the standard phoneme path of the target command word. A statistics module is used to determine a key interval based on the alignment path, and to judge the alignment results within the interval by situation, and to count the number of error lines, wherein the proportion of alignment results of blank phonemes or target command word phonemes is judged respectively; a positioning module, configured to locate at least one group of key phoneme pairs based on the alignment path and the phoneme sequence features of the target command word, each group of key phoneme pairs comprising two key phonemes with positional association; a distance judgment module, configured to calculate the position distance between two key phonemes in each group of key phoneme pairs and determine whether the position distance is abnormal; A judgment module is used to judge that the current speech segment is an out-of-set word and reject recognition if the error line count exceeds a set threshold or the position distance is abnormal; otherwise, it is judged to be a valid command word.

9. A device for rejecting out-of-word recognition, characterized in that: include: processor; memory for storing computer programs; Wherein, when the processor executes the computer program, the method for rejecting out-of-word recognition according to any one of claims 1 to 7 is implemented.

10. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the method for rejecting recognition of out-of-words according to any one of claims 1 to 7 is implemented.

Images