KR20170109178A - Method of detecting a misperception section in speech recognition on natural language - Google Patents

Abstract

According to an embodiment of the present invention, a method for detecting a misrecognized section in a natural language voice recognition comprises the following steps: extracting a characteristic vector from voice input from outside; performing a first viterbi decoding by using a sound model and a language model with respect to the characteristic vector; performing a second viterbi decoding by using the sound model and the language model with respect to the characteristic vector; and comparing a first string obtained according to the first viterbi decoding and first time information, and a second string obtained according to the second viterbi decoding and second time information. A weight for the language model at a time of performing the second viterbi decoding can be zero.

Description

FIELD OF THE INVENTION [0001] The present invention relates to a method for detecting a false-

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to speech recognition, and more particularly, to a method for detecting a false recognition period in natural language speech recognition.

Generally, in speech recognition, a speech recognition verification method is applied to determine whether or not a speech recognition result is applied. Generally speaking, a method of calculating the reliability of a speech recognition result string can be used in verification of speech. For example, various methods such as a likelihood-based calculation method using anti-phoneme and a method using a string probability in a speech recognition search space can be used.

Generally, when speech recognition is performed, the results of the input speech are output using the probabilities of the acoustic model and the language model. In this case, if the input of the voice having different characteristics and characteristics of the incorrect voice or the learned acoustic model is low, the probability of the acoustic model is low, and thus it is affected by the language model. In this case, a string with a high probability of a language model is outputted as a result, and a mistaken string appears as a bundle. Therefore, in order to improve the reliability of speech recognition, it is very important to detect the speech recognition type cluster section.

The technical idea of the present invention provides a method for efficiently predicting a section in which a false-positive bundle section occurs in natural language speech recognition.

The method for detecting a false-positive cluster in natural language speech recognition according to an embodiment of the present invention includes the steps of extracting a feature vector from an externally-input voice, using an acoustic model and a language model for the feature vector, Performing non-decoding on the feature vector, performing a second viterbi decoding using the acoustic model for the feature vector, performing a second viterbi decoding on the first character and first time information obtained in accordance with the first viterbi decoding, And comparing the second string and second time information obtained in accordance with the Viterbi decoding, wherein the weight for the language model at the time of performing the first Viterbi decoding is not 0, and the second Viterbi decoding The weight for the language model of the poem may be zero.

According to the embodiment of the present invention, it is possible to provide a method for efficiently predicting a section in which a false-positive bundle section occurs in natural language speech recognition.

1 is a block diagram illustrating a speech recognition system according to an embodiment of the present invention.
FIG. 2 is a diagram illustrating a schematic operation of the false-positive bundle interval detection of the speech recognition system according to the embodiment of the present invention.
FIG. 3 is a table showing an example of a false-positive bundle interval detection according to an embodiment of the present invention.
4 is a flowchart illustrating a speech recognition operation according to an embodiment of the present invention.
5 is a block diagram illustrating a mobile device to which a speech recognition system according to an embodiment of the present invention is applied.

It is to be understood that both the foregoing general description and the following detailed description are exemplary and should provide a further description of the claimed invention. Reference numerals are shown in detail in the preferred embodiments of the present invention, examples of which are shown in the drawings. Wherever possible, the same reference numbers are used in the description and drawings to refer to the same or like parts.

The terminology described herein is used for the purpose of describing a specific embodiment only, and is not limited thereto. Terms such as "one" should be understood to include plural forms unless explicitly referred to as " one ". The terms "comprising" or "comprising" are used to specify the presence of stated features, steps, operations, components, and / or components and may include additional features, steps, operations, components, And / or does not exclude the presence of their group.

1 is a block diagram illustrating a speech recognition system 100 according to an embodiment of the present invention. Referring to FIG. 1, the speech recognition system 100 may include a feature extraction unit 110, a search unit 120, a recognition result output unit 130, and a database 140.

For example, the speech recognition system 100 may be a smart phone, a mobile device such as a tablet PC, or a type of module provided in a computer and operated in the form of software or firmware. Or the speech recognition system 100 may be a device such as a semiconductor chip fabricated as a hardware for a specific purpose.

Referring to FIG. 1, the feature extraction unit 110 may extract feature vectors from a voice input from the outside. For example, before the feature information is extracted, a preprocessing process such as removing echoes or removing noise from an externally input voice signal may be performed. For example, the feature information may be information that effectively represents a digitally processed voice signal.

The search unit 120 can search for a word string having the highest degree of similarity to the feature vector generated by the feature extraction unit 110. For example, both an acoustical model and a linguistic model may be needed to find the word sequence most similar to the feature vector. The required models may be the acoustic model 140 and the language model 150, respectively. For example, the search unit 120 may search for the most probable word sequence using a viterbi algorithm.

The recognition result output unit 130 may output the word sequence searched by the search unit 120 in the form of text. For example, the word sequence searched by the search unit 120 may be a word sequence that can be recognized by the user on the display of the mobile device or the computer equipped with the speech recognition system 100 according to the embodiment of the present invention. . ≪ / RTI >

The acoustic model 140 is a kind of database required for words to be recognized by the search unit 120. For example, the acoustic model 140 may include a phoneme unit (e.g., 'a', 'b', 'c', ..., 'a', ' ',' ㅓ ', ...) can be learned by deep learning. Particularly, in the case of such an acoustic model, since each phoneme to be pronounced is influenced by surrounding phonemes, a context-dependent phone model as well as a simple phoneme like unit model can be used. In particular, a training method can be used to estimate the parameters of each acoustic model. This learning method is applied to set an acoustic model that is insensitive to the characteristics of a speaker or environmental noise using speech learning data collected in a large variety of environments.

The language model 150 may be statistical data that can be used to find relationships between words within a given sentence and reflect them in speech recognition. The language model 150 may include knowledge of learned vocabulary or vocabulary usage. For example, in the language model 150, a statistical model may be applied to indicate a degree of probability that a word such as 'a', 'a', 'a', and the like may appear after 'father'. This means that if the words are given in order, then the next word is related to the previous word.

In addition, a pronunciation rule database (not shown) may be used to reflect general phonological phenomena such as consonantation, palatalization, etc. for word string search, and a vocabulary dictionary (not shown) may be used to register the recognition vocabulary itself have.

According to the embodiment of the present invention, the weights of the acoustic model 140 and the language model 150 are differentiated in order to detect a false-positive bundle in speech recognition. For example, the search unit 120 may perform viterbi decoding using both the acoustic model 140 and the language model 150 with respect to the input voice. In this case, the language model weights are used to adjust the range of the acoustic model probability and the range of the language model probability, and the optimum value can be predetermined in the course of learning or development. After that, the search unit 120 may perform Viterbi decoding using the acoustic model 140. That is, by setting the weight for the language model to '0', speech recognition can be performed in a state in which the language model probability is ignored. That is, it is possible to acquire the speech recognition result character string using only the probability of the acoustic model, using the search space used in the large-capacity speech recognition.

Thereafter, the decoded result using both the acoustic model 140 and the language model 150 (i.e., when the weight of the language model is not 0) and the decoded result using the acoustic model 140 And the weight of the model is 0), the erroneous-type bundle section can be detected. Such a false-cell bundle interval detection operation may be performed directly by the recognition result output unit 130 or may be performed by a separate comparison unit (not shown). The decoding result when the weight of the language model is 0 may be larger than the decoding result when the weight of the language model is not zero. Therefore, a section in which the difference in the number of syllables is large in the aligned string section can be recognized as a false recognition section section. The outline of the operation of detecting the false-positive bundle interval is shown in FIG. 2 and FIG.

FIG. 2 is a diagram illustrating a schematic operation of the false-positive bundle interval detection of the speech recognition system according to the embodiment of the present invention. FIG. 3 is a table showing an example of a false-positive bundle interval detection according to an embodiment of the present invention. Hereinafter, Fig. 2 and Fig. 3 will be described together.

First, in step S110, a voice may be received from an external (e.g., user). By way of example, it is assumed that the user inputs a voice saying "Yes, I feel a little burdened a little." Although not shown in the figure, feature vectors may be extracted from the received speech after a preprocessing operation such as echo cancellation or noise cancellation is performed on the received speech.

In step 120, a first Viterbi decoding operation may be performed. The first viterbi decoding operation is performed by weighting both the acoustic model and the language model. That is, the weight for the language model is not zero. Illustratively, according to the first Viterbi decoding result, the sentence "Yes, I think so" can be recognized. Here, it can be seen that "something that is a little burdensome" is mistaken as "something like that". This can be interpreted as the fact that the probability of the speech model is strongly applied because the probability of the acoustic model is low due to the unclear speech of this section or the mixing of noise.

In step S130, a second viterbi decoding operation may be performed. The second Viterbi decoding operation is performed by weighting the acoustic model. That is, the weight for the language model is zero. As a result of performing speech recognition in the state of ignoring the probability of language model, the sentence such as "Negev Gold Puy fasting < / RTI > In this process, since the probability of the language model is excluded, the recognized string may differ from the actual speech string, but the number of uttered syllables can be similarly obtained.

In step 140, the first viterbi decoding result and the second viterbi decoding result may be compared. That is, the erroneous-type bundle section is detected by comparing the string information and the time information obtained in step S120 with the string information and the time information obtained in step S130. The whole sentence can be divided into a specific section by dividing the whole sentences according to time, or comparing the time when the corresponding strings are arranged using time information. For example, the first Viterbi decoding result may be divided into three sections such as "Yes // I think so," and the second Viterbi decoding result may be " // Walking in the sky ".

When the first Viterbi decoding result is compared with the second Viterbi decoding result, it can be seen that the number of syllables is significantly different in the second section of both results. That is, the second section in which the number of syllables is significantly different can be detected as a false-positive bundle section. Although the first Viterbi decoding operation is illustratively performed before the second Viterbi decoding operation in the present embodiment, the second Viterbi decoding operation may be performed before the first Viterbi decoding operation. Alternatively, the first viterbi decoding operation and the second viterbi decoding operation may be performed simultaneously.

According to such a search method, it is possible to easily retrieve the erroneous speech bundle section of the input speech by performing the Viterbi decoding only by assigning a weight to the language model, without performing a complicated algorithm or operation.

4 is a flowchart illustrating a speech recognition operation according to an embodiment of the present invention.

In step S210, feature vectors may be generated from the feature information extracted from the input speech. The feature information may be information for effectively representing a digitally processed voice signal.

In step S220, a first Viterbi decoding operation may be performed. For example, the first viterbi decoding operation may be an operation of performing decoding using the acoustic model and the language model for the feature vectors generated in step S210. At this time, the weight for the language model may not be zero. For example, this step may be performed by the search unit 120 shown in FIG. 1 and may be performed using the acoustic model 140 and the language model 150.

In step S230, a second Viterbi decoding operation may be performed. For example, the second viterbi decoding operation may be an operation of performing decoding using the acoustic model for the feature vectors generated in step S210. At this time, the weight for the language model may be zero. That is, speech recognition is performed in a state in which the language model probability is ignored. For example, this step may be performed by the search unit 120 shown in FIG. 1 and may be performed using the acoustic model 140.

In step S240, the first Viterbi decoding result and the second Viterbi decoding result may be compared. For example, this step may be performed by the recognition result output unit 130 shown in FIG. 1, or may be performed by a separately provided comparison unit (not shown). Since the language model is not considered in the second viterbi decoding operation, the first viterbi decoding result and the second viterbi decoding result may differ greatly in the number of syllables. Therefore, a comparison operation using the string information and time information obtained in the first viterbi decoding operation and the string information and the time information obtained in the second viterbi decoding operation can be performed.

In step S250, the false-positive cluster section detection operation can be executed. This step can be executed based on the comparison result in step S240. For example, as shown in the table shown in FIG. 3, since the first Viterbi decoding operation result and the second Viterbi decoding operation result have a large difference in the number of syllables, the erroneous- It can be easily detected. If the erroneous-type bundle section is detected according to the determination result in this step, a guide to re-enter the voice may be provided to the user. For example, such guidance may be provided to the user through the speaker or display of the mobile device.

FIG. 5 is a block diagram illustrating a mobile device 1000 to which a speech recognition system (shown as a speech recognition module 1610) according to an embodiment of the present invention is applied. For example, the mobile device 1000 may be a variety of electronic devices such as smart phones, tablet PCs, digital cameras, and the like. Referring to FIG. 5, the mobile device 1000 may be configured to support a mobile industry processor interface (MIPI) standard or an eDP (Embedded DisplayPort) standard. The mobile device 1000 includes an application processor 1100, a display unit 1200, an image processing unit 1300, a data storage 1400, a wireless transceiver unit 1500, a DRAM 1600, an Attitude Heading Reference System (AHRS) 1700, and an infrared sensor 1800.

The application processor 1100 may control the overall operation of the mobile device 1000. The application processor 1100 may include a DSI host that performs interfacing with the display unit 1200 and a CSI host that performs interfacing with the image processing unit 1300.

The display unit 1200 may include a display panel 1210 and a display serial interface (DSI) peripheral circuit 1220. The display panel 1210 can display image data. The DSI host embedded in the application processor 1100 can perform serial communication with the display panel 1210 through the DSI. The DSI peripheral circuit 1220 may include a timing controller, a source driver, and the like necessary for driving the display panel 1210.

The image processing unit 1300 may include a camera module 1310 and a camera serial interface (CSI) peripheral circuit 1320. The camera module 1310 and the CSI peripheral circuit 1320 may include lenses, image sensors, and the like. The image data generated in the camera module 1310 may be processed in the image processor, and the processed image may be transferred to the application processor 1100 via the CSI.

Data storage 1400 may include embedded storage and removable storage. The embedded storage and removable storage can communicate with the application processor 1100 via the M-PHY layer. For example, application processor 1100 and removable storage may communicate by various card protocols (e.g., UFDs, MMC, eMMC secure digital (SD), mini SD, Micro SD, etc.). The embedded storage and removable storage may be configured as non-volatile memory devices such as flash memory.

The wireless transceiver unit 1500 may include an antenna 1510, an RF unit 1520, and a modem 1530. The modem 1530 is shown communicating with the application processor 1100 via the M-PHY layer. However, depending on the embodiment, the modem 1530 may be embedded in the application processor 1100.

The DRAM 1600 may be loaded with various applications, firmware, software, etc., which are driven by the application processor 1100. For example, the DRAM 1600 may be loaded with a speech recognition module 1610 according to an embodiment of the present invention. The speech recognition module 1610 loaded into the DRAM 1600 may be driven by the application processor 1100. Although illustratively DRAM is shown as being used, it is not so limited, and various memory devices that can be loaded by the speech recognition module 1610 and driven by the application processor 1100 may be used.

The speaker 1700 can receive the voice to be recognized. The voice input from the user through the speaker 1700 will be processed in the voice recognition module 1610 after a preprocessing such as echo cancellation or noise cancellation.

It will be apparent to those skilled in the art that the structure of the present invention can be variously modified or changed without departing from the scope or spirit of the present invention. In view of the foregoing, it is intended that the present invention cover the modifications and variations of this invention provided they fall within the scope of the following claims and equivalents.

100: Speech Recognition System
110: Feature extraction unit
120:
130: recognition result output unit
140: Acoustic model
150: Language model

Claims (1)

1. A method for detecting a false recognition bundle interval in speech recognition using a device having a speech recognition system, the method comprising:
Extracting a feature vector from an externally input voice;
Performing a first Viterbi decoding using an acoustic model and a language model for the feature vector;
Performing a second viterbi decoding using the acoustic model for the feature vector;
Comparing the first string and the first time information obtained according to the first Viterbi decoding with the second string and the second time information obtained according to the second Viterbi decoding,
Wherein the weight for the language model at the time of performing the first Viterbi decoding is not zero and the weight for the language model at the time of performing the second Viterbi decoding is zero.

Images