TWI386918B - Sound recognition method - Google Patents

Description

Method of voice recognition

The present invention is a method for voice recognition, and more particularly to a method for determining which sound source is an audio source.

Because of the uniqueness of the sounds emitted by various creatures or people, it is possible to determine which creature or person the sound is emitted by the recognition of the sound, and by the uniqueness of the sound, the sound recognition can be applied to the anti-theft system, the creature. Identification system, etc.
Generally, the technique of sound recognition that is commonly used is to use the distance between the two parameters of the two sounds for identification. It is more common for the metirc-based voice recognition method. The same sample and audio are used in the multiple sample audio of an audio database. The audio to be authenticated is compared. (The audio signal referred to here refers to the sound digital signal. Since the sound source is an analog signal, it is converted into a digital signal by an analog digital converter. Therefore, unless otherwise explained, The audio referred to in this specification refers to a sound digital signal. By comparing the distance between each audio parameter in the to-be-identified audio and each audio parameter in the sample audio, a similarity is obtained by calculating the distance of each audio parameter. The value is used to determine whether the to-be-identified audio and the sample audio are the same sound source.
Since the metric-based voice recognition method calculates the distance between each audio parameter in the to-be-identified audio and each audio parameter in the sample audio, since the number of audio parameters is large, the distance for calculating each audio parameter takes more time to calculate. The distance between each audio parameter, and the sound source emitted by the creature or the person or the item, is not the same every time the sound is generated, and the audio parameter is changed more irregularly because the energy of the sound generated at that time or the environment is noisy. The metric-based voice recognition method has a high recognition error rate in voice recognition, and the metric-based voice recognition method only uses the similarity value to determine whether the to-be-identified audio and the sample audio are the same audio, and no Other methods of verification result in a higher recognition error rate for metric-based voice recognition methods.
Since the conventional technology has the above problems, and the correct rate of sound recognition is insufficient, the inventor has high knowledge of the professional knowledge in the field of sound science, and upholds the enthusiasm for sound recognition, so he began to study and think about how to solve it. The above-mentioned conventional techniques are insufficient.

In view of the problems arising from the prior art, the inventors believe that there should be a method for improvement. After many designs, experiments and reflections, a method of voice recognition is finally obtained, thereby improving the prior art deficiencies.
The invention is a method for voice recognition, which determines the sound of the audio system to be authenticated by the following steps, including the following steps:
(A) taking an audio database, the audio database comprising a plurality of sample audio;
(B) taking the same audio from the plurality of sample audio, and classifying the to-be-identified audio and the sample audio by using a to-be-identified audio and the sample audio to obtain a similarity value;
(C) setting a similarity threshold, when the similarity value is higher than the similarity threshold, determining that the to-be-identified audio is the same audio as the sample audio, if the similarity value is lower than the similarity threshold Then, return to step (B).
The invention compares the to-be-identified audio with the similarity of each sample audio in the audio database, and the sample audio includes sounds spoken by different people, sounds of various animals, sounds of object collisions, etc., thereby judging the What kind of sound is to be identified, because similar or identical audio is less easy to classify, the classifier is used to classify the audio to be identified and any sample audio in the audio database as the similarity value. The basis for judgment.
The classifier may be a support vector machine, a nearest neighbor classifier, a GMM, a K-means, etc., and in the preferred embodiment, the classifier performs classification processing on the to-be-identified audio and the sample audio for the support vector machine. A sorting line can classify each audio parameter in the to-be-identified audio and each audio parameter in the sample audio, and then perform a classification test on each audio parameter in the to-be-identified audio and each audio parameter in the sample audio, and check In the process, after the audio parameter to be identified is verified as the sample audio parameter, the first classification error rate is obtained. Unless otherwise specified, the number of sample audio parameters is marked as A, and the number of the audio parameter to be identified is marked. Is B, so the first classification error rate calculation formula is: When the sample audio parameter is verified to be the to-be-identified audio parameter during the verification process, the second classification error rate is obtained, and the second classification error rate calculation formula is: Since the audio parameters of the same sound are relatively the same, the classifier is less able to accurately find a classification line to classify the to-be-identified audio parameters and the sample audio parameters, so that the first classification error rate and the first The binary error rate has a higher error rate value.
In the preferred embodiment, in step (C), the similarity threshold is a classification error rate threshold. When the first classification error rate and the second classification error rate are both higher than the classification error rate threshold, then Determining that the to-be-identified audio is the same audio as the sample audio, thereby finding out what kind of sound the to-be-identified audio system is; if only the first classification error rate is higher than the classification error rate threshold, or only the second classification error If the rate is higher than the classification error rate threshold, or the first classification error rate and the second classification error rate are not higher than the classification error rate threshold, it is determined that the to-be-identified audio is different from the sample audio system. The to-be-identified audio continues to be compared with other sample audios until the sound of the to-be-identified audio system is determined.
Since the present invention uses the classification line of the to-be-identified audio parameter and the sample audio parameter to determine whether it is the same sound, the conventional technique uses the distance between the to-be-identified audio parameter and each parameter of the sample audio parameter to be the smallest. The value is classified, has a better similarity judgment effect, and takes less time to judge, and since the classification line is found by the method of the present invention, each audio parameter is tested by the classification line, and then The accuracy of the sound judgment of the present invention is improved.

BRIEF DESCRIPTION OF THE DRAWINGS The constructions, features, and embodiments of the present invention are illustrated by the accompanying drawings, which are set forth in the claims.
The invention is a method for voice recognition, which determines the sound of a sound source system by the following steps, including the following steps:
Referring to the first figure, in step (A), an audio database is included. The sound database includes a plurality of sample audio, and the plurality of sample audio systems includes sound sources emitted by humans, animals, machinery, or other items. The digital analog converter converts these sound sources into audio, which is stored as sample audio, and is used as a basis for judging which sound is to be identified.
Referring to the first figure, step (B), taking an audio to be identified, the sound to be identified may be a sound source for human, animal, mechanical or other objects, and then the sound is amplified by a digital analog converter. Converting the source into audio, taking the same audio in the audio database, comparing the to-be-identified audio with the sample audio, thereby determining whether the to-be-identified audio and the sample audio are the same sound, the similarity The method of comparing degrees uses a classifier to classify the to-be-identified audio and the sample audio, and since two similar or the same sounds are less likely to be classified by the classifier, the classifier will When the to-be-identified audio and the sample audio are classified, a similarity value is generated to determine whether the sound is the same sound, and the classification machine can be a nearest neighbor classifier, a support vector machine, a GMM, a K-means, etc. A machine that classifies two or more pieces of data.
Referring to the first figure, in step (C), a similarity threshold is set. When the similarity value is higher than the similarity threshold, it is determined that the to-be-identified audio is the same audio as the sample audio. To achieve the effect of voice recognition.
Since the operation mode of the support vector machine inputs two different data, the support vector machine is used to find the minimum boundary between each parameter to obtain a classification line, and the parameters of the two data are distinguished and classified. Therefore, in the first embodiment of the present invention, the classifier is a support vector machine in step (B), as shown in FIG. 3A, the to-be-identified audio includes a plurality of to-be-identified audio parameters (2), The sample audio includes a plurality of sample audio parameters (3), and the support vector machine is used to find the classification line (1), and the respective to-be-identified audio parameters (2) and the sample audio parameters (3) are divided into two categories, and then the The classification line (1) checks the respective to-be-identified audio parameters (2) and the respective sample audio parameters (3), and if the audio parameters (2) are to be identified in the verification process, the sample audio parameters are determined after the test. (3), please refer to the third-B diagram, the right side of the classification line (1) is the audio parameter to be identified (2) is represented by a circle, and the left side of the classification line (1) is the sample audio parameter (3) The square shape indicates that the circular line on the right side of the classification line (1) moves to the direction indicated by the dotted arrow on the left side of the classification line (1), which indicates that the audio parameter to be identified (2) is originally determined, and the sample audio is determined after the test. Parameter (3), counting these parameters that are incorrectly judged, then obtaining a first classification error rate, the first classification error The error rate calculation formula is: . If the sample audio parameter (3) is originally in the test, it is determined as the to-be-identified audio parameter (2) after inspection, please refer to the third-B figure, and the left side square of the classification line (1) moves to the classification line ( 1) The direction indicated by the dotted arrow on the right side indicates that the original audio parameter (3) is determined. After the test, it is determined as the audio parameter to be identified (2). If these parameters are incorrectly determined, a second classification is obtained. The error rate, the second classification error rate calculation formula is: .
Referring to the second figure, in the first embodiment, the similarity value is the first classification error rate and the second classification error rate, because the to-be-identified audio is matched by the support vector machine. When the sample audio is classified, if the to-be-identified audio and the sample audio are the same sound, it is difficult to find the classification line when sorting (1) to classify the to-be-identified audio and the sample audio, and therefore, in the inspection process A higher classification error rate is generated, so the classification error rate can be used as the judgment of similarity. Therefore, the threshold of similarity in step (C) is a classification error rate threshold (0), and the first classification error rate and the second classification error rate are both higher than the classification error rate threshold (0). And determining that the to-be-identified audio is the same sound as the sample audio, if only the first classification error rate is higher than the classification error rate threshold (0), or only the second classification error rate is higher than the classification error rate If the threshold value (0), or the first classification error rate and the second classification error rate are not higher than the classification error rate threshold (0), it is determined that the to-be-identified audio is different from the sample audio, The first classification error rate and the second classification error rate are both higher than the classification error rate threshold (0) to determine that the to-be-identified audio and the sample audio are the same sound, so that the invention has a higher sound recognition. rate.
Referring to the second figure, the second embodiment of the present invention is further provided with a pre-step before the step (A). The pre-step is to first take a conversation audio (6), and the conversation audio (6) can be A record of a dialogue between a plurality of different speakers, such as a radio broadcast, a conference record, or a general conversation record, and then converting the dialog record into an audio signal (6) using an analog digital converter, first extracting the dialogue audio (6) The audio overlap segment and the audio mute segment of different speakers are used to find out the audio boundary of one of the different speakers (7), and the method for finding the audio boundary (7) can be using the nearest neighbor classifier, support vector machine, etc. The preferred embodiment of the present invention utilizes the support vector machine to find the audio boundary (7).
Referring to the second figure, a first detection window (4) and a second detection window (5) are provided. The first detection window (4) and the second detection window (5) are Detecting the conversation audio (6) in the same unit time, respectively, and obtaining a first audio parameter and a second audio parameter, and then using the support vector machine to classify the first audio parameter and the second audio parameter. By determining whether the first detection window (4) and the second detection window (5) detect whether the conversation audio (6) is the same language, the principle of the judgment is explained in the embodiment of the present invention. Therefore, the first classification error rate curve (8) and the second classification error rate curve (9) are obtained. If it is determined that the dialogue audio (6) detected by the first detection window (4) and the conversation audio (6) detected by the second detection window (5) are different language audio, then The audio boundary (7), the audio boundary (7) is adjacent to the edge of the first detection window (4) through the second detection window (5), and the conversation audio (6) is divided into different speakers. The first detection window (4) is detected by the start time of the conversation audio (6), and the second detection window (5) is adjacent to the first detection window (4). Detecting, detecting the dialogue audio (6) by the first detection window (4) and the second detection window (5), and obtaining at least one audio boundary (7) to record the conversation (6) The audio to be authenticated is divided into a plurality of different words, and then the sound of the speaker to be identified is determined by the steps (A) to (C).
The third embodiment of the present invention is in accordance with the second embodiment. Referring to the second figure, any audio to be authenticated is selected in the plurality of to-be-identified audio signals, and the to-be-identified audio and other complex numbers are used by the support vector machine. Identifying the audio to make a similarity comparison, and comparing the to-be-identified audio whose similarity value is higher than the similarity threshold value to the voice of the same speaker, and the audio zone of each of the same speakers in the dialogue audio (6) The block is marked so that step (A) to step (C) only need to identify the to-be-identified audio of the different speakers, and reduce the time required to judge the voice of each speaker to be authenticated.
A fourth embodiment of the present invention further includes a step (D), wherein the step (D) is provided with a language recognizer that converts the dialogue content of the dialog audio (6) into a text record, so that the present invention The embodiment is applied to the automatic meeting record, and the content of the meeting is recorded by the recorder. By the method of the present invention, the content of each person's speech is found, and then the language recognizer is used to convert the content of each person's speech into a text file. In order to speed up the meeting, it is not necessary to manually record the content of the meeting, and accelerate the time required for the meeting.
In summary, the present invention is indeed in line with industrial utilization, and is not found in publications or publicly used before application, nor is it known to the public, and has non-obvious knowledge, conforms to patentable requirements, and patents are filed according to law. .
However, the foregoing is a preferred embodiment of the present invention in the industry, and all the equivalent changes made by the scope of the patent application of the present invention belong to the scope of the claim.

(0)‧‧‧Classification error rate threshold

(1)‧‧‧ classification line

(2) ‧‧‧ audio parameters to be authenticated

(3) ‧‧‧sample audio parameters

(4) ‧‧‧ first detection window

(5) ‧‧‧Second detection window

(6) ‧‧‧Dialog audio

(7) ‧ ‧ audio boundaries

(8) ‧‧‧First classification error rate curve

(9) ‧‧‧Second classification error rate curve

The first diagram is a flow chart of the steps of the present invention. The second diagram is a schematic diagram of an embodiment of the present invention.

The third-A diagram is a schematic diagram of the operation of the support vector machine of the present invention (1)
The third-B diagram is a schematic diagram of the operation of the support vector machine of the present invention (2)

Claims (5)

A method of voice recognition, comprising:
(A) taking an audio database, the audio database comprising a plurality of sample audio;
(B) taking the same audio from the plurality of sample audio, and classifying the to-be-identified audio and the sample audio by using a to-be-identified audio and the sample audio to obtain a similarity value;
(C) setting a similarity threshold, when the similarity value is higher than the similarity threshold, determining that the to-be-identified audio is the same audio as the sample audio, if the similarity value is lower than the similarity threshold Then, return to step (B). The method for voice recognition according to the first aspect of the patent, wherein the classifier used in the step (B) is a support vector machine, and when the support vector machine classifies the to-be-identified audio and the sample audio, The two similarity values are respectively a first classification error rate and a second classification error rate, and the similarity threshold is a classification error rate threshold, when the first classification error rate and the second classification error rate are both When the classification error rate threshold is higher than the classification error rate, it is determined that the to-be-identified audio is the same sound as the sample audio. The method for voice recognition according to claim 1, wherein before the step (A), a pre-step is set, the pre-step adopts a dialogue audio, and the dialogue audio is a dialogue of the plural speakers. The first detection window and the second detection window are respectively configured to detect the conversation audio in the same unit time, and the at least one audio interface has a first detection window and a second detection window. A first signal parameter and a second signal parameter are respectively obtained, and the first detection window is detected by the start time of the conversation audio, and the second detection window is adjacent to the first detection on the time axis. The window starts to detect, and the first signal parameter and the second signal parameter are classified by using a support vector machine to obtain a first classification error rate and a second classification error rate respectively; and setting a classification error rate threshold When the first classification error rate and the second classification error rate are both lower than the classification error rate threshold, an audio dividing line is set, and the audio dividing line is vertically detected by the second detection window. Point and time When the first classification error rate and the second classification error rate are not lower than the classification error rate threshold, the first detection window does not move, and the second detection window slides for one unit time. And detecting the conversation audio until the first classification error rate and the second classification error rate are lower than the classification error rate threshold; by repeating the above steps, the conversation audio is divided into multiple to-be-identified audio signals. . The method for voice recognition according to claim 3, wherein the pre-steps take an audio to be authenticated in the plurality of to-be-identified audios, and use the classifier to make the to-be-identified audio similar to other plural to-be-identified audios. For comparison, the to-be-identified audio with the similarity value higher than the similarity threshold is marked as the same sound. The method for voice recognition according to claim 3, wherein a step (D) is further provided, wherein the step (D) is a language identifier, and the language recognizer converts the audio of each speaker into a text file. .