KR101261754B1 - Apparatus and method for generating corpus for training phrase break prediction - Google Patents

Abstract

Disclosed are a device and a method for generating a hangout training corpus. The punctuation generating device for cutting-reading training is provided with a boundary extractor for extracting word boundaries from each of the plurality of transcription data by receiving a plurality of transcription data which are labeled with word boundaries from each of a plurality of transcriptions and transcribed the extracted words. And a boundary determination unit for determining a word boundary that satisfies a predetermined criterion by cutting each boundary, and determining a break reading of the text, and a break reading corpus generation unit for generating a break reading training corpus by labeling the determined break reading. Therefore, since the cut-out training corpus is formed using a plurality of transcripts transferred from a plurality of transcripts for the same text, the characteristics of the individual transcripts can be minimized to reflect the cut-out characteristics of the universal transcript as much as possible. It can show high performance on average.

Description

APAPATUS AND METHOD FOR GENERATING CORPUS FOR TRAINING PHRASE BREAK PREDICTION}

The present invention relates to a speech synthesis system, and more particularly, by generating a reading-ready training corpus using a plurality of transcription data transferred from a plurality of transcriptions for the same text, thereby minimizing reflection of characteristics of individual transcriptions. The present invention relates to an apparatus and method for generating a corpus for sever reading, which can fully reflect the sperm reading characteristics of the phosphorus transfer.

In order to generate natural synthesized sound in speech synthesis system, it is very important to correctly estimate the position of reading in a given input text. In the related art, a training corpus generated by displaying the extent to which the translator hears and cuts the recorded sentence for use in speech synthesis was used. The conventional training corpus formed in this way reflects the characteristics of the talker in the recorded sentences. For example, a fast-talker with a low speech frequency reads less frequently, whereas a slow-talker with a slower speech rate frequently reads more frequently, so that the same sentence can be read differently depending on the characteristics of the talker. do. In addition, the conventional reading-reading corpus is disclosed in the "automatic labeling method of reading by speech synthesis" (Korean Patent Application No. 10-1999-0053451). The automatic labeling method of the broken reading is disclosed in a method of labeling the broken reading in consideration of whether the idle section of the voice database is after the closing sound of the closed sound, but in this case, the characteristics of the voice expert are reflected in the labeling result of the reading. There is a problem that it is difficult to create a universal training corpus.

An object of the present invention for solving the above problems, by using a plurality of transcription data transferred from a plurality of transcriptions for the same text generation of a reading reading corpus that can fully reflect the reading characteristics of the universal transcription To provide a device.

Another object of the present invention for solving the above problems, using a plurality of transcription data transferred from a plurality of transcriptions for the same text using a plurality of transcriptions for reading the cut-off training corpus can be reflected as much as possible To provide a method of generation.

Technical problems of the present invention are not limited to the technical problems mentioned above, and other technical problems not mentioned will be clearly understood by those skilled in the art from the following description.

In accordance with an aspect of the present invention for achieving the above object of the present invention, the apparatus for generating a reading-reading corpus is provided with a plurality of transcription data, each of which is transcribed with word boundaries labeled for the same text from each of the plurality of transcription characters. A boundary extracting unit for extracting word boundaries from each of a plurality of transcription materials, a boundary determining unit for comparing word extracted boundaries to determine word boundaries satisfying a predetermined criterion, and determining to read the text; It includes a reading block generation unit for generating a reading training corpus by labeling the reading.

In accordance with an aspect of the present invention for achieving the above object of the present invention, a method for generating a reading-reading corpus is provided with a plurality of transcription materials, each of which is transcribed with a word boundary for the same text from a plurality of transcriptions. Extracting word boundaries for each of the provided plurality of transcripts, comparing each extracted word boundary with each other, and determining to cut out the word boundary satisfying a predetermined criterion; And labeling the determined cutoff to generate a cutout training corpus.

The apparatus and method for generating a reading-reading corpus according to the present invention as described above is formed by using a plurality of transcription data transferred from a plurality of transcriptions for the same text, so that the reflection of the characteristics of the individual-transcriber is minimized. As a result, it is possible to fully reflect the reading characteristics of the universal transcript, which has an advantage of showing high performance on average.

1 is a block diagram showing the configuration of a device for generating a scavenger training for reading according to an embodiment of the present invention.
2 is a conceptual diagram illustrating a process of generating a cut-out instruction corpus from a plurality of transcription data according to an embodiment of the present invention.
3 is an exemplary diagram showing a mismatch in reading between a plurality of transcription materials transferred from a plurality of transcriptions for the same text.
Figure 4 is a flow chart showing a method for generating a cut off training corpus according to an embodiment of the present invention.

While the invention is susceptible to various modifications and alternative forms, specific embodiments thereof are shown by way of example in the drawings and will herein be described in detail. It should be understood, however, that the invention is not intended to be limited to the particular embodiments, but includes all modifications, equivalents, and alternatives falling within the spirit and scope of the invention. Like reference numerals are used for like elements in describing each drawing.

The terms first, second, A, B, etc. may be used to describe various elements, but the elements should not be limited by the terms. The terms are used only for the purpose of distinguishing one component from another. For example, without departing from the scope of the present invention, the first component may be referred to as a second component, and similarly, the second component may also be referred to as a first component. And / or < / RTI > includes any combination of a plurality of related listed items or any of a plurality of related listed items.

When a component is referred to as being "connected" or "connected" to another component, it may be directly connected to or connected to that other component, but it may be understood that other components may be present in between. Should be. On the other hand, when an element is referred to as being "directly connected" or "directly connected" to another element, it should be understood that there are no other elements in between.

The terminology used herein is for the purpose of describing particular example embodiments only and is not intended to be limiting of the present invention. Singular expressions include plural expressions unless the context clearly indicates otherwise. In this application, the terms "comprise" or "have" are intended to indicate that there is a feature, number, step, operation, component, part, or combination thereof described in the specification, and one or more other features. It is to be understood that the present invention does not exclude the possibility of the presence or the addition of numbers, steps, operations, components, components, or a combination thereof.

Unless defined otherwise, all terms used herein, including technical or scientific terms, have the same meaning as commonly understood by one of ordinary skill in the art. Terms such as those defined in the commonly used dictionaries should be construed as having meanings consistent with the meanings in the context of the related art and shall not be construed in ideal or excessively formal meanings unless expressly defined in this application. Do not.

Hereinafter, preferred embodiments according to the present invention will be described in detail with reference to the accompanying drawings.

1 is a block diagram showing the configuration of a device for generating a scavenger training for reading according to an embodiment of the present invention.

Referring to FIG. 1, the apparatus for generating a thorough reading training corpus is provided with a plurality of transcription data, each of which is labeled by word boundaries from each of the plurality of transcription characters for the same text, and is transferred to each word. By extracting a boundary, comparing the extracted word boundary to determine the word boundary that satisfies a predetermined criterion as phrase break prediction of the text, and labeling the determined break reading to correlate the break reading training corpus (corpus) By minimizing the reading characteristics of each individual transcript and creating a reading training corpus that fully reflects the universal cutting characteristics.

The apparatus for generating a disconnect reading training corpus 100 includes a boundary extractor 110, a boundary determining unit 120, and a disconnect reading corpus generation unit 130.

The boundary extracting unit 110 may receive a plurality of transcription data that is transferred by labeling word boundaries from each of the plurality of transcription characters for the same text, and extract the word boundaries from each of the plurality of transcription data.

The boundary extractor 110 uses a neural network commonly used to extract word boundaries, a method using a classification and regression tress (CART), and a hidden markov model (HMM). The method, the method by a rule, etc. can be used.

The boundary determination unit 120 may compare each of the extracted word boundaries to determine the word boundary that satisfies a preset criterion by cutting the text, and the boundary determination unit 120 determines the first determination module 121. ) And the second determination module 122.

The first determination module 121 extracts a word boundary that satisfies a preset first criterion among the word boundaries extracted from each of the plurality of transcription data, and determines the first word boundary. The preset first criterion may mean that all of the plurality of transcription data are matched among the word boundaries extracted from each of the plurality of transcription data. In other words, a word boundary that matches all of the plurality of transcription data among the word boundaries extracted from each of the plurality of transcription data may be extracted and determined as the first word boundary.

The second determination module 122 may extract a word boundary that satisfies a preset second criterion among word boundaries not included in the extracted first word boundary, and determine the second word boundary. The preset second criterion may mean that a match is made in a majority of the plurality of transcription data among word boundaries not included in the extracted first word boundary. That is, the word boundary corresponding to a majority of the plurality of transcription data is extracted from the word boundary not included in the extracted first word boundary among the word boundaries extracted from each of the plurality of transcription data, and is determined as the second word boundary. Can be.

For example, "This flower is very pretty." Warrior 1, which is labeled with the boundary of the word transferred from a plurality of warriors, reads "This / flower / true / pretty." Transcript 2 is "This / flower / pretty." And transcription 3 is "This flower is beautiful / beautiful." In the case of receiving the word, the first word boundary, which is a word boundary that matches all the transcription materials 1 to 3, "This flower // is very pretty." Is determined in the first determination module and is matched in a majority of word boundaries not included in the first word boundary, "this / flower is true / pretty." The second word boundary can be determined. '//' means a first word boundary, and '/' means a second word boundary.

In addition, the second determination module 122 extracts a word boundary in consideration of qualities that affect reading when a half of the plurality of transcription data are matched among word boundaries not included in the extracted first word boundary. May be determined as the second word boundary.

Qualities affecting the reading may mean a utterance habit, ignition rate, ignition intensity of the transcript.

For example, "This flower is very pretty." Warrior 1, which is labeled with the boundary of the word transferred from a plurality of warriors, reads "This / flower / true / pretty." Transcript 2 is "This / flower / pretty." The transcription data 3 is "this flower is / pretty / pretty", and the transcription data 4 is a case where a plurality of transcription data "this flower / is pretty" is provided, all of the transcription data 1 to 4 The first word boundary "this flower // is very pretty", which is a matching word boundary, is determined by the first decision module, and the word boundary "yi / flower" and "this flower" are not included in the first word boundary. "," That's pretty "and" True / pretty. " '//' means a first word boundary, and '/' means a second word boundary. In this case, considering the utterance habits, ignition rate, ignition intensity, etc. of the transcripts, which affect the cut-offs of the transcription materials 1 to 4, the weighted transcripts transferred to the universal cut-outs are weighted to match the half number. One of the two word boundaries may be determined.

The cutout corpus generation unit 130 may generate the cutout training corpus by labeling the determined cutout.

2 is a conceptual diagram illustrating a process of generating a cut-out instruction corpus from a plurality of transcription data according to an embodiment of the present invention.

Referring to FIG. 2, transcription data 1 to transcription material 7, which are a plurality of transcription data that are labeled and transcribed word boundaries from each of a plurality of transcriptions for the same text, are illustrated. A word boundary is extracted from the plurality of transcripts, and a word boundary satisfying a predetermined criterion may be determined as a first word boundary by comparing the extracted word boundary with each other, and the word not included in the first word boundary. A second word boundary (the part shown by a dotted line in FIG. 2) is determined based on a preset criterion among the boundaries to determine the first and second word boundaries as phrase break prediction of the text. It represents the concept of creating a corpus for hung training by labeling hung.

3 is an exemplary diagram showing a mismatch in reading between a plurality of transcription materials transferred from a plurality of transcriptions for the same text.

Referring to FIG. 3, it shows how differently the plurality of transcriptions are transferred to the same text, and transfers the reading position and the intensity when the speech is naturally uttered so that there is no emphasis portion in the text at a normal speaking speed. The result obtained by doing so is shown.

The cut-out ratio divided into long cut (major shown in FIG. 3), short cut (minor shown in FIG. 3), and not cut (Non shown in FIG. 3) has different distributions depending on the transcription. As described above, the different distributions are displayed according to the transcripts because the same text is transferred by reflecting the individual characteristics of the transcripts.

For example, because the transcriber LSH represents the least percentage of non-reading, we can see that the normal speaking speed of the transcriber LSH is relatively faster than the normal translator, while the transcriber YUI has the highest percentage of long reads. As the ratio is expressed, it can be seen that the normal speaking speed of the warrior YUI is relatively slow compared to that of the general warrior.

That is, it can be seen that FIG. 3 reflects the characteristics of individual translators influencing cutting off even the transcription data obtained by copying the same text.

Table 1 below is a cut-out training formed by using a system for training the cut-out training corpus formed by the apparatus and the method for generating a cut-out training corpus according to an embodiment of the present invention, and the respective transcription data transferred from the individual translators. Shows a comparison table comparing systems trained with dragon corpus.

Referring to Table 1, the system trained with the cut-out training corpus formed by using only the transcription data transferred from the individual transcripts shows a high cut-out match rate for the corresponding transcript. However, the system trained with the cut-out training corpus formed using only the transcription data transferred from the individual transcribers reflects the cut-out characteristics of the transcribed and shows low cut-out matching rate for the translators other than the transcribed.

On the other hand, the system is trained with a cut reading corpus formed by the apparatus and method for generating a cut reading training corpus according to an embodiment of the present invention, a plurality of transcriptions in which the cut reading training corpus is transferred from a plurality of transcriptions for the same text. Since it is formed by using the data, it can be seen that the performance of the average transcription can be reflected as much as possible by minimizing the reflection of individual transcription characteristics.

Figure 4 is a flow chart illustrating a method for generating a corpus for breaking reads according to an embodiment of the present invention.

Referring to FIG. 4, the method for generating a thorough reading training corpus is provided with a plurality of transcription data that is transcribed with word boundaries labeled from each of the plurality of transcription characters for the same text (S 110), and in each of the transcription data. The word boundary may be extracted (S 120).

The word boundary extraction may be performed by using a neural network which is generally used to extract word boundaries, a method using a classification and regression tress (CART), a method using a hidden markov model (HMM), Rule-based methods and the like can be used.

By comparing each of the extracted word boundaries, word boundaries satisfying a predetermined criterion may be determined as phrase break prediction (S 130 and S 140).

In detail, a word boundary that satisfies a preset first criterion among the word boundaries extracted from each of the plurality of transcription data may be extracted and determined as the first word boundary (S 130). It may mean that all of the plurality of transcription data among the word boundary extracted from each of the transcription data. In other words, a word boundary that matches all of the plurality of transcription data among the word boundaries extracted from each of the plurality of transcription data may be extracted and determined as the first word boundary.

In addition, a word boundary that satisfies a second preset criterion among word boundaries not included in the extracted first word boundary may be extracted and determined as the second word boundary (S 140).

The preset second criterion may mean coincidence in a majority of the plurality of transcription data among the word boundaries not included in the extracted first word boundary among the word boundaries extracted from each of the plurality of transcription data. That is, a word boundary matching a majority of the plurality of transcription data among the word boundaries not included in the extracted first word boundary may be extracted and determined as the second word boundary.

For example, "This flower is very pretty." Warrior 1, which is labeled with the boundary of the word transferred from a plurality of warriors, reads "This / flower / true / pretty." Transcript 2 is "This / flower / pretty." And transcription 3 is "This flower is beautiful / beautiful." In the case of receiving the word, the first word boundary, which is a word boundary that matches all the transcription materials 1 to 3, "This flower // is very pretty." Is determined in the first determination module and is matched in a majority of word boundaries not included in the first word boundary, "this / flower is true / pretty." The second word boundary can be determined. '//' means a first word boundary, and '/' means a second word boundary.

The preset second criterion may be cut off when a match is made in half of the plurality of transcription data among the word boundaries not included in the extracted first word boundary among the word boundaries extracted from each of the plurality of transcription materials. The word boundary may be extracted and determined as the second word boundary in consideration of the qualities that are affected.

Qualities affecting the reading may mean a utterance habit, ignition rate, ignition intensity of the transcript.

For example, "This flower is very pretty." Warrior 1, which is labeled with the boundary of the word transferred from a plurality of warriors, reads "This / flower / true / pretty." Transcript 2 is "This / flower / pretty." The transcription data 3 is "this flower is / pretty / pretty", and the transcription data 4 is a case where a plurality of transcription data "this flower / is pretty" is provided, all of the transcription data 1 to 4 The first word boundary "this flower // is very pretty", which is a matching word boundary, is determined by the first decision module, and the word boundary "yi / flower" and "this flower" are not included in the first word boundary. "," That's pretty "and" True / pretty. " '//' means a first word boundary, and '/' means a second word boundary. In this case, considering the utterance habits, ignition rate, ignition intensity, etc. of the transcripts, which affect the cut-offs of the transcription materials 1 to 4, the weighted transcripts transferred to the universal cut-outs are weighted to match the half number. One of the two word boundaries may be determined.

A cut reading training corpus labeled cut off may be generated based on the first and second word boundaries (S 150).

It will be apparent to those skilled in the art that various modifications and variations can be made in the present invention without departing from the spirit or scope of the present invention as defined by the following claims It can be understood that

100: Breakdown Training Coral Generator
110: boundary extracting unit 120: boundary determining unit
121: first determination module 122: second determination module
130: Cut-read corpus generator
1000: Chop Training Coral Database

Claims (12)

A boundary extracting unit configured to receive a plurality of transcription data obtained by labeling word boundaries of the same text from each of the plurality of transcription characters, and extract a word boundary from each of the plurality of transcription materials;
A boundary determination unit which compares each of the extracted word boundaries and decides to cut the word boundary corresponding to a majority among the plurality of transcription materials by reading the text; And
A cut-read training corpus generating device comprising a cut-read corpus generating unit for labeling the determined cut-off reading to generate a cut-read training corpus. The method of claim 1,
The boundary determination unit,
A first determination module for comparing each extracted word boundary to determine a word boundary that matches in all of the plurality of transcription data as a first word boundary; And
And a second determination module for determining a word boundary matched in a majority of the plurality of transcription data among the word boundaries not included in the determined first word boundary as a second word boundary. Device. delete delete delete delete Receiving a plurality of transcripts from which a word boundary for the same text is labeled and transferred from each of the plurality of transcripts;
Extracting word boundaries for each of the provided plurality of transcription data;
Comparing each of the extracted word boundaries to determine a word boundary that is matched in a majority of the plurality of transcripts by cutting off the text; And
And labeling the determined cutoff reading to generate a cutout training corpus. The method of claim 7, wherein
Comparing each of the extracted word boundaries and determining the word boundaries that are matched in a majority of the plurality of transcripts to cut the text,
Comparing each of the extracted word boundaries to determine a word boundary corresponding to all of the plurality of transcription data as a first word boundary; And
And determining a word boundary that matches a majority among the plurality of transcription data among word boundaries not included in the determined first word boundary as a second word boundary.

delete delete delete delete

Images