JP2002366177A - Node extracting device for natural voice - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a node extracting device for a natural voice which stably and efficiently extract nodes needed to generate a pitch pattern by spline approximation. SOLUTION: A pattern extraction part 1 extracts a basic frequency pattern according to an inputted natural voice 102. An input part 2 inputs language information 101. A pattern sectioning part 3 sections the basic frequency pattern into access phrases according to the language information 101. A no-voice control part 4 corrects a voiceless section of an access phrase pattern curve into a smooth curve by interpolation. A differential arithmetic part 5 finds the primary differential curve and secondary differential curve of the access phrase pattern curve. A node extraction part 6 extracts nodes from the access phrase pattern curve according to the primary differential curve, secondary differential curve, and language information 101. A node information output part 7 outputs a node 103 as a final node to the outside.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a node extracting device for a speech synthesizer, and more particularly, to a natural speech node extracting device for extracting nodes necessary for generating a pitch pattern by spline approximation. It is about.

[0002]

2. Description of the Related Art Recent speech synthesizers synthesize speech according to a rule synthesis method. In the rule synthesis method, a pitch pattern indicating a temporal change pattern of a fundamental frequency of a voice is generated by giving a node as a parameter to the rule synthesis engine, and is used as one of rules for synthesizing the voice.

[0003] A speech synthesizer previously stores nodes extracted from natural speech by a node extractor, performs spline approximation based on the nodes, and generates a pitch pattern. The spline approximation is a process of connecting discrete points called nodes in order and using a spline function to approximate a smooth curve as a whole.

[0004] In a speech synthesis apparatus of the rule synthesis system, as one of the rules of speech synthesis, a continuous speech of an arbitrary vocabulary is directly synthesized from phonetic symbols or characters which are separately input, using a generated pitch pattern. . It is important for the node extraction device to be able to extract nodes without being influenced by conditions such as the gender of the speaker and the utterance speed, and some proposals have been made.

[0005] The IEICE SP2000-29 describes a node extraction method used in a natural speech node extraction apparatus (IEICE IEICE Technical Report, July 2000, page 20, author: Hiroyoshi Morikawa Naohiro Tsuboi Yuichiro Yanagi, Title: "Modeling and Analysis of Speech Pitch Pattern Using Smoothing Spline Function"). In the node extraction method (node selection method) performed by the node extraction device, a fundamental frequency is extracted from natural speech, and the extracted fundamental frequency is obtained as a plurality of data points plotted for each time. A node is selected using the two methods shown in FIG.

FIG. 6 is a flowchart of the first node selection method. The start point and the end point of the plurality of data points are set as two nodes, the interval from the start point to the end point is equally divided at a time interval dt, and the closest data point is extracted for each division point to be a node candidate (step S81). The inclination between adjacent node candidates is obtained, and if the magnitude of the inclination is smaller than the threshold value TH2, the candidate is deleted from the node candidates (step S82).

[0007] Based on the nodes and the node candidates, a smooth spline function is obtained to calculate an error from the fundamental frequency pattern curve (step S83). Here, the fundamental frequency pattern curve handles a group of a plurality of data points as a curve. The error in step S83 is equal to the threshold value TH3.
If it is smaller, the node candidate with the smallest inclination is deleted, and the processing is executed from step S83 (step S83).
86). If the error in step S83 is larger than the threshold value TH3, the finally remaining node candidate is determined as a node (step S85).

FIG. 7 is a flowchart of the second node selection method. The start point and the end point of the plurality of data points are two nodes, and both nodes are connected by a straight line, and the data point farthest from the straight line is set as a node candidate (step S91). Based on the nodes and the node candidates, a smooth spline function is obtained to calculate an error from the fundamental frequency pattern curve (step S9).
2).

If the error in step S92 is larger than the threshold value TH3, the data point furthest from the spline function is added as a new node candidate, and the process is executed from step S92 (step S95). If the error in step S92 is smaller than the threshold value TH3, the finally remaining node candidate is determined as a node (step S94).

[0010]

In the above-described conventional natural speech node extraction device, the first node selection method empirically determines the time interval dt because the time interval dt to be divided depends on the speech speed. There must be. Further, it is necessary to change the threshold value TH2 or TH3 to be compared with the error according to the utterance speed and to determine the threshold value empirically, so that a node cannot be obtained stably. Similarly to the first node selection method, the second node selection method uses the threshold value TH2 or TH3, and thus needs to be determined empirically depending on the speech speed.

In general, if a node is extracted from a fundamental frequency pattern curve of natural speech without considering its shape,
The pitch pattern generated by the spline approximation based on the node has an effect such as a waving phenomenon, and a shape or pattern different from the fundamental frequency pattern of the natural voice may be generated.

In the above-described conventional natural speech node extraction device,
Since a node is selected by comparing whether or not the error based on the node and the node candidate is within a threshold value, a pitch pattern different from the fundamental frequency pattern of natural speech may be generated, and the shape of the fundamental frequency pattern curve may be generated. Is not considered enough.

SUMMARY OF THE INVENTION The present invention has been made to solve the above-mentioned problems of the prior art, and stably and efficiently extracts necessary nodes in order to generate a pitch pattern by spline approximation. The present invention provides a natural speech node extraction device.

[0014]

To achieve the above object, a natural sound node extraction device according to the present invention comprises a pattern extraction unit for extracting a fundamental frequency pattern of a naturally occurring sound, and a method for extracting the fundamental frequency pattern for each accent phrase. And a differential operation unit for obtaining a first derivative curve and a second derivative curve of the divided fundamental frequency pattern; and a differentiator for the divided fundamental frequency pattern, the first derivative curve and the second derivative curve. And a node extracting unit for extracting a node of the fundamental frequency pattern based on the information.

The natural speech node extraction device of the present invention extracts nodes based on a fundamental frequency pattern curve obtained by extracting a fundamental frequency by dividing a natural speech by an accent phrase, and its first and second derivative curves. As a result, a change point or the like characterizing the shape of the fundamental frequency pattern curve is extracted as a node, and this node is extracted independently of the speech speed, so that stable and efficient node extraction can be performed.

In the natural sound node extraction device according to the present invention, the node extraction unit extracts a zero point of the primary differential curve as a node, and extracts a maximum point and a minimum point of the secondary differential curve, respectively. It is preferable to extract it as a node. In this case, node extraction can be reliably performed at a change point or the like that characterizes the shape of the fundamental frequency pattern curve.

Further, the natural speech node extraction device of the present invention may further comprise an input unit for designating whether the natural speech is a question sentence and whether or not the natural speech includes an accent position. preferable. In this case, the node extraction unit includes:
When natural speech is a question sentence, the lowest frequency point immediately before the end point of the divided fundamental frequency pattern is extracted as a node. When natural speech includes an accent position, the first derivative curve of the primary differential curve is extracted. The highest point and the lowest point of the second derivative curve after the zero point can be extracted as nodes. In this case, using language information such as accent positions,
Since the shape of the fundamental frequency pattern curve can be sufficiently considered, node extraction can be performed more reliably at a change point or the like that characterizes the shape of the fundamental frequency pattern curve.

It is also a preferred embodiment of the present invention that the node extracting unit newly extracts an intermediate point between two adjacent nodes obtained earlier as a node. In this case, it is ensured that the pitch pattern is generated based on the nodes.

[0019]

BRIEF DESCRIPTION OF THE DRAWINGS FIG. 1 is a block diagram showing a natural speech node extracting apparatus according to an embodiment of the present invention. FIG. 1 is a block diagram of a natural speech node extraction device according to an embodiment of the present invention. The natural speech node extraction device includes a pattern extraction unit 1, a pattern classification unit 3, an input unit 2, an unvoiced control unit 4, a differential operation unit 5, a node extraction unit 6, and a node information output unit 7.

The pattern extracting unit 1 extracts a fundamental frequency pattern based on the input natural speech 102 and inputs the pattern to the pattern dividing unit 3. The fundamental frequency pattern is a plurality of data points from which the fundamental frequency has been extracted at the point in time when the short time interval is extracted. A data point is composed of an extraction time and a fundamental frequency.

The input unit 2 inputs the input linguistic information 101 to the pattern classification unit 3. The language information 101 includes a start time and an end time of an accent phrase, an accent position time,
This is information including a start time and an end time of a consonant vowel included in the accent phrase, and a sentence type indicating whether the sentence is a question sentence or a declarative sentence. The pattern classification unit 3 divides the fundamental frequency pattern into accent phrases based on the linguistic information 101 and inputs the same to the unvoiced control unit 4.

FIG. 2 shows information about "Is it OK?" The natural sound is plotted as points at frequencies generated at each time. The black shaded portions in the figure indicate the frequency characteristics (spectral display) of natural speech.
As shown in FIG. 2A, the fundamental frequency of natural speech is 2
It is shown as a white line * in the black shaded region from 00 Hz to 400 Hz.

The unvoiced control section 4 examines unvoiced sections included in the fundamental frequency pattern curve based on the linguistic information 101. If there is an unvoiced section, the fundamental frequency pattern curve is likely to cause an error when extracting nodes required for spline approximation. Therefore, the fundamental frequency pattern curve is interpolated and corrected as a smooth accent phrase pattern curve.

As shown in FIG. 2B, when there is a unvoiced section (“sh”) including a consonant, it is stretched from a nearby voiced section (“o” or “i−”) and is drawn by a straight line or a curve. Interpolate. If the start point or the end point of the accent phrase is unvoiced, a value smaller by several Hz than the value of the nearby voiced section is interpolated as the start point or the end point. The silent control unit 4
Each accent phrase pattern curve is represented by a node (B
1, P1, E1, E2), and is input to the node extractor 6 after being smoothed continuously.

FIG. 3 is a flowchart of a node extraction method performed by the natural speech node extraction apparatus of FIG. The differential operation unit 5 includes a first-order differential curve of the accent phrase pattern curve and 2
The next differential curve is obtained and input to the node extraction unit 6. The node extraction unit 6 extracts nodes based on the primary differential curve, the secondary differential curve, and the language information 101.

FIGS. 4A, 4B and 4C show an accent phrase pattern curve of a declarative sentence, its first derivative curve,
And second derivative curves are shown. The accent phrase pattern curve has a start point B1 of the accent phrase pattern and an end point E1 of the accent phrase pattern. The first derivative curve has a zero crossing point P1 whose sign changes from positive to negative. The second derivative curve includes the highest point A1 before the zero crossing point P1, the highest point C2 after the zero crossing point P1, and the zero crossing point P1.
It has the previous lowest point A2 and the lowest point C1 after the zero crossing point P1.

The data point B1 which is the start point of the accent phrase pattern curve is extracted as the node B1, and the data point E1 which is the end point of the accent phrase pattern curve is extracted as the node E1 (step S11). First-order differentiation of the accent phrase pattern curve is performed to obtain a first-order differential curve (step S1)
2).

A zero-crossing point P1 at which the sign of the primary differential curve changes from positive to negative is obtained, and a node P1 which is a data point P1 on the accent phrase pattern curve corresponding to the zero-crossing point P1 is extracted. If there are a plurality of zero-crossing points, the crossing point closest to the highest frequency point of the accent phrase pattern curve is set as the zero-crossing point P1 (step S13). If the sentence type of the linguistic information 101 is not a question sentence (step S14), the process proceeds to step S16 to execute the next process.

FIGS. 5A, 5B, and 5C show an accent phrase pattern curve of a question sentence, its first derivative curve,
And second derivative curves are shown. The accent phrase pattern curve has a start point B1 of the accent phrase pattern, a lowest frequency point E1, and an end point E2 of the accent phrase pattern. The first derivative curve has a zero crossing point P1. The second derivative curve has the highest point A1 before the zero crossing point P1, the highest point C2 after the zero crossing point P1, and the lowest point C1 after the zero crossing point P1.

If it is a question sentence in step S14, the data point E1 which is the lowest frequency point of the accent phrase pattern curve
Is extracted as the node E1, and the data point E2, which is the end point of the accent phrase pattern curve, is extracted as the node E2 (step S15). Further, the zero-crossing point E1 at which the sign of the primary differential curve changes from negative to positive may be examined, and the data point E1 on the accent phrase pattern curve corresponding to the zero-crossing point E1 of the primary differential curve may be set as the lowest frequency point E1.

As shown in FIG. 4C, the accent phrase pattern curve is second-order differentiated to obtain a second derivative curve (step S16). The vertex of the second derivative curve in the section from the node B1 to the node P1 of the accent phrase pattern curve is examined, and the node A1 which is the data point A1 on the accent phrase pattern curve corresponding to the highest point A1 before the zero crossing point P1 is extracted. Then, in the section from the node A1 to the node P1, the zero-crossing point P1
The node A2, which is the data point A2 on the accent phrase pattern curve corresponding to the previous lowest point A2, is extracted (step S2).
17).

Next, the accent position time of the language information 101 is checked, and if the accent position is not included (step S
18) Go to step S20 to execute the next process. The accent position indicates an accent position. For example, since "questionnaire" goes down with the sound following "a", "a" has an accent, and the end position of the sound of "a" is the accent position.

If the accent position is included, 2 in the section from the node P1 to the node E1 of the accent phrase pattern curve
The vertex of the next differential curve is examined, and a node C2, which is a data point C2 on the accent phrase pattern curve corresponding to the highest point C2 after the zero crossing point P1, is extracted, and in a section from the node P1 to the node C2, the node after the zero crossing point P1 Is the data point C1 on the accent phrase pattern curve corresponding to the lowest point C1
Is extracted (step S19).

However, if there is no highest or lowest point of the second derivative curve in the designated section in step S17 or S19, no node on the accent phrase pattern curve is extracted. FIG. 5C shows an example in which there is no lowest point A2 before the zero crossing point P1 of the second derivative curve, and no node A2 of the accent phrase pattern curve is extracted.

Further, when obtaining the apex of the second derivative curve for the designated section, the third derivative curve is further obtained, and the zero crossing point where the third derivative curve changes sign to positive or negative is examined.
The vertex of the second derivative curve corresponding to the zero crossing point of the third derivative curve may be obtained.

If the nodes B1, A1, A2, P1, C1, C2, and E1 extracted on the accent phrase pattern curve alone are not enough to generate a pitch pattern, An intermediate point between two adjacent nodes obtained earlier may be newly extracted as a node.
The final node is input to the node information output unit 7, and the process ends (step S20).

The node information output unit 7 outputs the final node from the node extraction unit 6 as a node 103 to the outside. In addition, the node information output unit 7 may execute the process of adding the intermediate point corresponding to the step S20 and obtaining the final node.

According to the above embodiment, nodes are extracted based on a fundamental frequency pattern curve obtained by extracting a fundamental frequency by dividing a natural voice by an accent phrase and its primary differential curve and secondary differential curve. A change point or the like characterizing the shape of the frequency pattern curve is extracted as a node, and this node is extracted regardless of the speech speed, so that stable and efficient node extraction can be performed.

The speech synthesizer generates a pitch pattern on the basis of the nodal point, and determines one of the rules of speech synthesis (rule synthesis method).
And synthesizes speech based on phonetic symbols or character strings that are input separately. Pitch patterns are most closely related to accents and intonations and give not only natural and audible tones, but also indicate units of words and phrases and are easy to understand as sentences. The speech synthesizer is
If the generated pitch pattern faithfully reproduces the actual fundamental frequency pattern, a natural and easy-to-hear voice can be synthesized.

In the node extraction device of the present invention, a change point or the like characterizing the shape of the fundamental frequency pattern curve is extracted as a node, so that the generated pitch pattern can faithfully reproduce the actual fundamental frequency pattern. The present invention is suitably used not only for the synthesizing apparatus but also for an apparatus employing the rule synthesizing method.

Although the present invention has been described based on the preferred embodiment, the node extraction method of the present invention is not limited to the configuration of the above-described embodiment, but may be the same as that of the above-described embodiment. The node extraction device for natural speech modified and modified from the above is also included in the scope of the present invention.

[0042]

As described above, in the natural speech node extraction apparatus of the present invention, a fundamental frequency pattern curve obtained by extracting a fundamental frequency by dividing a natural speech by an accent phrase, its first derivative curve and its second derivative curve. By extracting nodes based on, the change points that characterize the shape of the fundamental frequency pattern curve are extracted as nodes, and since these nodes are extracted independently of the speech speed, stable and efficient node extraction It can be performed.

[Brief description of the drawings]

FIG. 1 is a block diagram of a natural sound node extraction device according to an embodiment of the present invention.

FIG. 2 shows information related to "Are you sure?"

FIG. 3 is a flowchart of a node extracting method performed by the natural speech node extracting apparatus of FIG. 1;

FIGS. 4A, 4B and 4C show an accent phrase pattern curve of a declarative sentence, its first derivative curve and its second derivative curve, respectively.

FIGS. 5 (a), (b), and (c) show an accent phrase pattern curve of a question sentence, its first derivative curve, and its second derivative curve, respectively.

FIG. 6 is a flowchart of a first node selection method.

FIG. 7 is a flowchart of a second node selection method.

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 1 Pattern extraction part 2 Input part 3 Pattern division part 4 Unvoiced control part 5 Differential calculation part 6 Node extraction part 7 Node information output part 101 Language information 102 Natural speech 103 Node

Claims (7)

[Claims] A pattern extracting unit for extracting a fundamental frequency pattern of a naturally occurring sound; a pattern dividing unit for dividing the fundamental frequency pattern for each accent phrase; a first derivative curve of the divided fundamental frequency pattern; A differential operation unit for obtaining a second derivative curve, and a node extraction unit for extracting a node of the fundamental frequency pattern based on the divided fundamental frequency pattern, the first derivative curve and the second derivative curve. Do
Node extraction device for natural speech. 2. The natural speech node extraction device according to claim 1, wherein the node extraction unit extracts a zero point of the first derivative curve as a node. 3. The node extractor extracts a highest point and a lowest point of the quadratic differential curve as nodes, respectively.
Or the natural speech node extraction device according to 2. 4. Whether or not the natural voice is a question sentence
The natural speech node extraction device according to any one of claims 1 to 3, further comprising an input unit for specifying whether or not to include an accent position. 5. The natural node according to claim 4, wherein when the natural speech is a question sentence, the node extracting unit extracts a lowest frequency point immediately before an end point of the divided fundamental frequency pattern as a node. Speech node extraction device. 6. When the natural speech includes an accent position, the node extraction unit extracts the highest point and the lowest point of the secondary differential curve after the zero point of the primary differential curve as nodes, respectively. Item 4. The natural speech node extraction device according to item 4 or 5. 7. The node extracting unit newly extracts an intermediate point between two adjacent nodes obtained earlier as a new node.
The node extraction device for natural speech according to any one of claims 3 to 6.