JP2014219635A - Pause insertion device and method and program thereof - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a pause insertion device for inserting a pause into an appropriate position in speech-synthesizing text information according to a reading speed.SOLUTION: A feature amount extracting part 110 inputs a piece of text information and extracts a feature amount vector which represents a feature amount of the text information. A reading speed model makes an association between the feature amount vector and a piece of pause information which represents insertion/not insertion of a pause. A model selection section 130 inputs a specified reading speed and a speed model and selects a reading speed model according to the specified reading speed to output the same. A pause position prediction section 140 inputs the feature amount vector and makes an association of the feature amount vector with the selected reading speed model to output a pause insertion position.

Description

  The present invention relates to a pose imparting apparatus for determining a pose position from text information for speech synthesis, a method thereof, and a program.

  In conventional speech synthesis, text information to be synthesized is related to information guidance, news, etc., and the speech synthesis read-out speed is assumed to be constant. Therefore, a pose assignment device, method and program for collecting a target voice and reproducing the pose assignment with the sound are constructed on the premise of generating synthesized speech at a constant speed even in pose assignment. It was. As a representative technique, for example, a pose imparting technique described in Non-Patent Document 1 is known.

  Should the pose assignment technique disclosed in Non-Patent Document 1 statistically model the appearance of words, parts of speech, and readings (syllables) as features, and place the pose position after or before each word in the text? This is a judgment. Also, when creating rules for predicting pose positions manually, rules are created to reproduce the pose positions of those speeches, with the goal of using speech collected at a constant speech synthesis reading speed. I came.

V. Keri, etal., “Pause Prediction from Lexical and Syntax Information”, International Conference on Natural Language Processing (ICON-2007), 2007.

  With the advancement of speech synthesis technology, the required speech synthesis sounds are also diversifying. Text information for speech synthesis is read at a constant speed, such as conventional information guidance and news, etc., as well as human emotions and other situations, such as text information in dialogue scenes and text information for advertising Text information that needs to be read out at various speeds is increasing. However, as described above, the addition of a pose to text information for speech synthesis has been limited to a technique based on a rule (model) that presupposes a constant reading speed or does not consider a difference in reading speed. The conventional rules are designed to reproduce the average pause position in the collected audio data, so text such as dialogue scenes where text information needs to be read out faster or slower than average. When combining, a pose cannot be given at an appropriate position. In addition, since there is no idea to absorb the difference in reading speed, the rules for giving poses are also averaged, and there is a problem in the accuracy of the rules (model). That is, the conventional pose imparting device has a problem that it cannot cope with a plurality of reading speeds.

  The present invention has been made in view of these problems, and provides a pose imparting apparatus capable of imparting accurate pose information corresponding to a reading speed to text information, and a method and program thereof. Objective.

  The pose imparting apparatus of the present invention includes a feature amount extraction unit, a model for each reading speed, a model selection unit, and a pose position prediction unit. The feature quantity extraction unit receives text information as an input and extracts a feature quantity vector that is a feature quantity necessary for pose position prediction. The model according to the reading speed associates the feature amount vector with pose information meaning that a pose is placed or not placed. The model selection unit receives the designated reading speed and the model for each reading speed, and selects and outputs a model for each reading speed corresponding to the designated reading speed. The pose position prediction unit predicts the pose provision position of the text information by associating the feature amount vector with the model according to the reading speed selected by the model selection unit.

  According to the pose imparting apparatus of the present invention, since a pose is imparted to text information using a model for each reading speed, an accurate pose imparting position corresponding to the reading speed can be predicted. Moreover, since the rule (model) is created on the premise that the reading speed changes, it is possible to improve the accuracy of the pose assignment.

The figure which shows the function structural example of the pose provision apparatus 100 of this invention. The figure which shows the operation | movement flow of the pause provision apparatus. The figure which shows the example of a part of model according to reading speed. The figure which shows the function structural example of the pose provision apparatus 200 of this invention. The figure which shows the operation | movement flow of the pause provision apparatus. 3 is a diagram conceptually showing a full speed model 220. FIG.

Embodiments of the present invention will be described below with reference to the drawings. The same reference numerals are given to the same components in a plurality of drawings, and the description will not be repeated.
[Invention]
Prior to the description of the embodiments, a new concept of the present invention will be described. When speaking a sentence longer than a certain length, the pose when speaking slowly is usually different from the pose when speaking quickly. For example, when reading the sentence “Thank you very much for a very valuable souvenir.”, There is a high possibility that the _first pose P1 will be placed after “Te”. On the other hand, when reading late, pose P ₁ after “precious”, pose P ₂ after “souvenir”, and pose P ₃ after “te” are natural. Thus, the position where the pose is given naturally varies depending on the speed at which the text information is read out.

  If, for example, a part-of-speech string is used as a feature quantity for giving a pose, and a match is seen in partial part-of-speech strings for several words, there is a high possibility that a partial part-of-speech string with the same part-of-speech string appears. By extracting a characteristic partial part-of-speech sequence to which a pose is given from the partial part-of-speech sequence and constructing a pose assignment model, it is possible to assign an accurate and natural pose, and natural and easy-to-hear speech synthesis. Sound can be generated.

  The present invention prepares a conventional model for determining the position of a pose by inputting, for example, a feature amount represented by a partial part-of-speech string, corresponding to the reading speed, and a model and text corresponding to the specified reading speed. The pose provision position is predicted from the correspondence with the feature amount extracted from the information. According to the pose imparting apparatus of the present invention, since a pose imparting model is prepared for each reading speed, accurate (natural) pose position information corresponding to the reading speed can be predicted.

  FIG. 1 shows an example of the functional configuration of a dose applying apparatus 100 according to the present invention. The operation flow is shown in FIG. The pose imparting apparatus 100 includes a feature amount extraction unit 110, a model 120 for each reading speed, a model selection unit 130, a pose position prediction unit 140, and a control unit 150. The pause giving device 100 is realized by a predetermined program being read into a computer composed of, for example, a ROM, a RAM, and a CPU, and the CPU executing the program. The same applies to other embodiments described below.

  The feature quantity extraction unit 110 receives text information as an input and extracts a feature quantity vector that is a feature quantity of the text information (step S110). The feature amount is, for example, a syllable or a part of speech string that constitutes a word of text information. Extraction of these feature quantities can be obtained by, for example, the method of Reference Document 1 (Yuji Matsumoto, morphological analysis system “tea bowl”, information processing, 41 (11), pp. 1208-1214, 2000). Instead of syllables, appearance forms that are word faces can be used.

  Table 1 shows examples of feature vectors and examples of word appearances and part-of-speech sequences. “Thank you very much for receiving a very valuable souvenir.” This is shown in the example.

  A feature quantity vector is the appearance form or part of speech, or the appearance form and part of speech arranged in the time axis direction. The feature quantity vector that can be represented by the appearance form W and the part of speech P can be represented by Expression (1).

  The feature quantity vector may be only the appearance form W or only the part of speech P. This feature quantity vector is created by extracting the above-mentioned appearance form and part of speech from a plurality of words on the sentence head side from the word to be judged and a plurality of words on the sentence end side from the word to be judged. n is the maximum value of the number counted from the word at the beginning of the sentence toward the word at the end of the sentence, and can be determined in advance by experiment.

  The model 120 for each reading speed is a statistical model corresponding to each reading speed created from the recorded data of the feature amount extracted from the text information and the pose imparting position for each speed. The pose information indicating that the pose is placed or not placed is provided by associating with the quantity vector. The statistical model is, for example, a classification tree that classifies the presence or absence of a pose. The classification tree is a Classification and Regression Tree described in Non-Patent Document 1, and is well known. The classification tree may be replaced with a probability model (CRF: Conditional Random Field) that can be expressed by Equation (2). The probabilistic model P (y | x) is also well known as described in Reference 2 (supervised by Okumura, written by Takamura, “Introduction to Machine Learning for Language Processing”, Corona, p. 153). .

Here, x is a feature vector represented by the appearance form W or the part of speech P, and y is a symbol representing a variable indicating whether or not a pose is given. φ is a feature vector that takes x and y as arguments, and for each of the various types of x described above, a function that returns 1 if y indicates that a pose is placed and returns 0 if it is not placed is a vector It is. w is a weight vector of each feature function φ. Z _{x, w} are normalization terms for making the sum of P (y | x) equal to 1. “·” Indicates an inner product between vectors. These w, φ, Z _{x, and w} constitute a reading speed model 120. These classification trees and probability models can be created by a well-known general method. Then, the pose grant determination can be realized by using the one with the larger probability value when the grant is performed and when the grant is not performed.

The model 120 according to the reading speed may be a classification tree or a probability model. Or it may be an n-gram model. Such a reading speed model 120 is prepared for each reading speed, such as a low speed model 120 ₁ , a medium speed model 120 ₂ , and a high speed model 120 ₃ .

Speed for the model 120 ₂ medium is, for example, about 400 characters per minute is a standard reading speed of the announcer (Reference 3: Miki Tomo乃addition, "NHK Science and Technical Research Laboratories, NHK Engineering Services, Victor Company of Japan, speaking speed Development of radio / TV with conversion technology ”Okochi Prize Case Research Project, IIR Case Study CASE # 10-03, Hitotsubashi University Innovation Research Center, April 2010) is there. Low velocity model 120 ₁ assuming a reading speed of, for example, about 80% of medium-speed model 120 ₂ model, speed model 120 ₃ assuming a reading speed of, for example, about 120% of the medium speed model 120 ₂ It can also be a model. As described above, the model 120 for each reading speed learned by using the specific reading speed as an index can make the accuracy of the pose assignment more accurate than the conventional model that has no idea of corresponding to the reading speed in the first place.

The model selection unit 130 receives the designated reading speed and the model for each speed, and selects and outputs a model for each reading speed corresponding to the designated reading speed (step S130). If the specified readout speed is inputted from the outside of the "slow", the model selection unit 130 outputs the pause position prediction unit 140 selects the low-speed model 120 _1. When the designated reading speed is “medium speed”, the medium speed model 120 ₂ is selected, and when the designated reading speed is “high speed”, the high speed model 120 ₃ is selected.

  The pose position prediction unit 140 receives the feature amount vector output from the feature amount extraction unit 110, and associates the feature amount vector with the model according to the reading speed selected by the model selection unit 130, thereby giving a pose of text information. The position is predicted (step S140). Steps S110 and S140 are repeated until the processing for all text information is completed (No in step S150). This repetitive operation process is performed by the control unit 150. The control unit 150 is a general unit that controls the time-series operation of each unit of the pose imparting apparatus 100 and does not perform any special processing.

The prediction of the pose assignment position will be described with reference to FIG. FIG. 3 is a part of an example of a model for each reading speed represented by a classification tree. FIG. 3 is a classification tree corresponding to the feature vector represented by the part-of-speech string and the appearance form, and shows a part corresponding to the processing of the “very valuable” part of the above sentence. As shown in Table 1, “very” is W ₁ , “precious” is W ₂ , and “na” is W ₃ . In the case of determining whether placing a pause immediately after the W ₃ will be described. In feature vector of "very valuable", and follow the classification tree, (Yes of step S1201) part of speech _{P 1} of appearance form _{W 1} is a noun, appearance form part of speech _{P 2} of the _{W 2} is a noun, adjective stem (step Yes of S1202), part of speech _{P 3} is an auxiliary verb of appearance form _{W 3} (Yes of step S1203), follow a branch of, Yes occurred form _{W 3} = so "Do" (step S1204), pose after the appearance form _{W 3} Is assigned (step S1205). In this case, the pose position prediction unit 140 outputs, for example, “pause grant”. Pose grant is meant to grant a pause after the appearance form W _3. If the appearance form W ₃ is other than “NA”, no pose is given after the appearance form W ₃ (step S1206).

As described above, the correspondence between the feature amount vector and the model according to the reading speed is to trace the model 120 according to the reading speed constituted by the feature amount vector, for example, by a classification tree. In this way, the position to which the pose is given can be predicted by tracing the model 120 according to the reading speed constituted by the classification tree with the feature amount vector. Here it was subjected to only the use judgment as to whether or not to grant a pause word of the beginning of a sentence side of the W _3, as well as a tree even if the end of the sentence side word than W ₃ is at the branch point of the classification tree By following the above, it is possible to determine whether or not to give a pose.

The reading speed by model consists of classification tree shown in FIG. 3, for example in the case of a low speed for model 120 _1, for example, the text information as a "Thank you very much I received a very valuable souvenir", a comma pause position If it is expressed by “,”, the pose information is given to three locations “Thank you for receiving a very valuable souvenir”.

  The pose information in the case of the high-speed model 1203 is given in one place, for example, “Thank you for receiving a very valuable souvenir”. By preparing a model for each reading speed in this way, it is possible to give an accurate and natural pose corresponding to the specified reading speed to the same text information.

  A method for assigning a pose using the probability model expressed by Equation (2) will be described. In the case of the probabilistic model of the formula (2), weights are given to φ, which are various combinations of a vector composed of a subset of elements of the feature vector x and y indicating whether or not to place a pose at the time of model construction. w is calculated. The subset of x referred to by φ can be selected and set in various ways by the model designer based on knowledge regarding pose assignment.

For example, three φs can be formed from W ₁ alone and y, W ₂ and W ₃ and y, and W ₁ and W ₂ , W ₃ and y. Further, the weight w can be determined using data at the time of model construction. For example, when determining whether to give a pose after “na” in the “very precious” portion described above, calculate whether these three φs are 1 or 0, and multiply each by adding a weight. (In other words, inner product calculation of w and φ), a value obtained through an exponential function is a probability value. This probability value is calculated for both cases with and without a pause, and the larger probability value is adopted as a result.

  FIG. 4 shows an example of the functional configuration of the position imparting device 200 of the present invention. The operation flow is shown in FIG. The pose imparting apparatus 200 includes a feature amount extraction unit 110, a full speed model 220, and a pose position prediction unit 240.

  The feature quantity extraction unit 110 receives text information as an input and extracts a feature quantity vector that is a feature quantity of the text information (step S110). The feature extraction unit 110 is the same as the pose imparting device 100 as is clear from the reference numerals, and the feature amount to be extracted is also the same.

  The total speed model 220 is a statistical model corresponding to a plurality of reading speeds created from feature amounts extracted from text information, and a pose is placed or placed by associating the statistical models with the feature amount vectors. It gives pose information that means no.

The full speed model 220 includes a plurality of statistical models of the reading speed such as the low speed model 120 ₁ , the medium speed model 120 ₂ , and the high speed model 120 ₃ . FIG. 6 conceptually shows the full speed model 220.

FIG. 6A shows an example in which velocity feature quantities are arranged at the branch points at the roots of the classification tree constituting the entire velocity model 220. And by speed model 120 low speed model 120 ₁ model for medium speed ₁₂₀₂ and speed model 120 ₃ shown in (FIG. 1) reading constitute the total velocity model 220 as one classification tree .

  FIG. 6B shows an example of the entire speed model 220 ′ in which the speed feature amount is arranged at various branch points of the classification tree. As described above, the branching based on the speed feature amount may exist in various parts of the classification tree. “¬V1”, “¬V2”, and “¬V3” shown in FIG. 6 mean branches based on different feature amounts that are not speeds. The full speed models 220 and 220 ′ having the structure shown in FIG. 6 are general ones that can be created by a conventional method in the same manner as the above-described model for each reading speed.

  The pose position prediction unit 240 creates a feature vector composed of a designated reading speed inputted from the outside and a feature amount output from the feature extraction unit 110, and associates the feature vector with the entire speed model 220. To predict the pose assignment position (step S240). Steps S110 and S240 are repeated until the processing for all text information is completed (No in step S250). This repetitive operation process is performed by the control unit 250.

  Even with the configuration of the pose imparting apparatus 200 described above, it is possible to impart a pose to an accurate and natural position corresponding to the reading speed specified for the text information for speech synthesis.

  When the processing means in the above apparatus is realized by a computer, the processing contents of the functions that each apparatus should have are described by a program. Then, by executing this program on the computer, the processing means in each apparatus is realized on the computer.

  The program is distributed by selling, transferring, or lending a portable recording medium such as a DVD or CD-ROM in which the program is recorded. Further, the program may be distributed by storing the program in a recording device of a server computer and transferring the program from the server computer to another computer via a network.

  Each means may be configured by executing a predetermined program on a computer, or at least a part of these processing contents may be realized by hardware.

Claims (6)

A feature quantity extraction unit that extracts text vector, which is a feature quantity necessary for pose position prediction, using text information as input;
A model according to the reading speed that associates the feature vector with pose information that indicates whether or not to place a pose;
A model selection unit for selecting and outputting a model according to the reading speed corresponding to the specified reading speed, with the specified reading speed and the reading speed model as inputs,
A pose position prediction unit that predicts a pose imparting position of the text information by associating the feature amount vector with the model according to the reading speed selected by the model selection unit,
A pose imparting device comprising: A feature quantity extraction unit that extracts text vector, which is a feature quantity necessary for pose position prediction, using text information as input;
A total speed model that associates the feature vector with the input designated reading speed and pose information that indicates whether or not to place a pose;
A pose position prediction unit that predicts a pose imparting position of the text information by associating the designated reading speed and the feature vector with the specified reading speed, the feature vector, and the entire speed model;
A pose imparting device comprising: In the pose grant apparatus according to claim 1 or 2,
The feature vector is expressed by the following equation:

A pose imparting device, which is a vector of word appearance forms W and parts of speech P. A feature quantity extraction process for extracting a feature quantity vector, which is a feature quantity necessary for pose position prediction, using text information as input,
A model for each reading speed corresponding to the above-mentioned designated reading speed is selected and output by inputting a model for each reading speed associating the designated reading speed with the feature vector and the pose information indicating that the pose is placed or not. The model selection process,
A pose position prediction process for predicting a pose imparting position of the text information by associating the feature quantity vector with the model according to the reading speed selected by the model selection unit, using the feature quantity vector as an input;
A pose granting method comprising: A feature quantity extraction process for extracting a feature quantity vector, which is a feature quantity necessary for pose position prediction, using text information as input,
By inputting a total speed model that associates the designated reading speed with the feature vector and pose information that indicates whether or not to place a pose, the correspondence between the designated reading speed, the feature vector, and the whole speed model A pause position prediction process for predicting a pause position of text information;
A pose granting method comprising:   A program for operating a computer as the pose imparting device according to claim 1.

Images