JP2785628B2 - Pitch pattern generator - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a pitch pattern generator in a speech synthesizer or the like for outputting text written in text.

[0002]

2. Description of the Related Art In a speech synthesizer or the like for outputting an arbitrary sentence as speech, it is extremely important to generate a natural pitch pattern in order to improve the quality of synthesized speech.

Conventionally, pitch patterns are generated by superimposing a component called a tone component, which gradually decreases throughout the entire utterance, based on a word called an accent component or through a component accent type for each phrase. Have been. For example, a method is known in which a speech component is approximated with a straight line or a “H” -shaped pattern that monotonically decreases, and an accent component is approximated with a polygonal line. Such a conventional technique is described in, for example, the material of the Society of Sound Engineers of Japan, S7.
8-07 (1978-4), "Examination of tone rules for sentence speech" (Literature 1).

FIG. 5 is a diagram for explaining a method of generating a pitch pattern according to the prior art. Here's an example that shows how to generate a pitch pattern for "He bought a white flower." This phrase, "He bought a white flower,""Bought a flower,""Is a unit having four accent nuclei. Therefore, the accent component 21
Is approximated by a polygonal line with four peaks. The shape of each mountain is the accent type of each accent phrase,
It is determined based on the number of mora and the like. This is superimposed on the speech component 22 represented by a straight line descending to the right to generate a pitch pattern 23 of the sentence. L1-L4 in FIG.
Is called a stress bell. The relationship between the relative levels of stress levels of adjacent accent phrases reflects the structure of the sentence and is important to the naturalness of the petit. That is, if the connection between the front and rear accent phrases is weak, the stress level of the subsequent accent phrase becomes higher than the stress level of the preceding accent phrase. Conversely, if the connection is strong in meaning, the magnitude of the stress level of the subsequent accent phrase becomes small.

[0005] In the conventional pitch pattern generation method such as Reference 1, the number of accent phrases from the preceding accent phrase to the received accent phrase is used as a scale indicating the strength of connection between adjacent accent phrases. Is called the degree of separation. The degree of separation is obtained from the dependency structure of the sentence. FIG. 6 is a diagram for explaining the degree of separation.
Greater separation at an accent phrase boundary means that the preceding accent phrase across that boundary is semantically connected to the more distant accent phrase and weaker to the next accent phrase. It represents that. “He” in FIG. 6 is related to “bought”, and the connection with “white” immediately after is weak. When the preceding accent phrase is directly related to the succeeding accent phrase, the degree of separation has a minimum value of 1. In the example of FIG. 6, the relationship between “white” and “flower” and the relationship between “flower” and “bought” are such. The stress level of the subsequent accent phrase is set to be larger than the stress level of the preceding accent phrase at the accent phrase boundary where the degree of separation is large, and smaller than the stress level of the subsequent accent phrase at the boundary of the accent phrase where the degree of separation is small. To be.

As described above, in the conventional pitch pattern generation method, the strength of the connection between adjacent accent phrases is examined using the sentence dependency structure, and the stress level of each accent phrase is determined according to the strength. . The pitch component of the entire sentence is generated by superimposing the thus generated accent component on the speech tone component.

[0007]

In the conventional pitch pattern generation method, a dependency structure of an accent phrase constituting a sentence is used, and it is premised that this is correctly obtained. However, it is generally difficult to always accurately analyze the dependency structure of a sentence. For this reason, there has been a problem that a generated pitch pattern becomes unnatural due to an analysis error of the dependency structure.

An object of the present invention is to provide a pitch pattern generation device capable of generating a natural pitch pattern without using a sentence dependency structure.

[0009]

In order to solve the above-mentioned problems, a pitch pattern generating apparatus according to the present invention is a pitch pattern generating apparatus in a voice synthesizing apparatus which reads an arbitrary sentence by voice, and the size of an adjacent accent phrase is large. A stress-bell-ratio storage unit that stores the ratio of each word for each part-of-speech combination, a morphological analysis unit that divides an input sentence into words that make up the word, and determines the part-of-speech and accent phrase boundaries of each word, An accent component generating unit that reads out the ratio of the size of the accent phrase from the stress level ratio storage unit according to the combination of the parts of speech of the words before and after the phrase boundary; And a pitch pattern generation unit that generates a pitch pattern.

[0010]

According to the present invention, it is possible to generate a more natural pitch pattern than the conventional method by generating a pitch pattern based on the part of speech of a word that can be more accurately obtained than the dependency structure of a sentence. And

[0011]

Next, embodiments of the present invention will be described with reference to the drawings. As described above, in the conventional method, it is difficult to accurately analyze the dependency structure,
An unnatural pitch pattern may be generated.
On the other hand, in the present invention, a pitch pattern is generated without using a dependency structure. In the present invention, a pitch pattern is generated based on the part of speech of a word that can be obtained more accurately than the dependency structure of a sentence. The combination of the parts of speech of the words before and after each accent phrase boundary is considered to reflect the strength of the connection in the meaning of the preceding and following accent phrases. FIG. 3 is a diagram for explaining a pitch pattern generation method according to the present invention. This sentence consists of four accent phrases, "he", "white", "flower", and "bought". For example, the combination of parts of speech at the boundary between “white” and “flower” is “adjective / adjective + noun”. From this combination, it is easy to infer that the preceding adjective directly modifies the next noun.

Therefore, for each combination of the parts of speech of the words before and after the accent phrase boundary, the ratio of the stress level of the accent phrase before and after the boundary (the ratio of the stress level of the succeeding accent phrase to the stress level of the preceding accent phrase,
Or its reciprocal) is set in advance. FIG. 4 is a diagram illustrating an example of a stress level ratio for each part of speech combination. It is conceivable that these values are determined based on, for example, a sentence voice uttered by a human.

When generating a pitch pattern, first, a sentence to be read is divided into words by morphological analysis.
Determine the part of speech. Next, at each accent phrase boundary, the stress level ratio of the accent phrase before and after the boundary is determined based on the part of speech of the word before and after the boundary. In FIG. 3, for example, the stress level of “flower” is 0.7 times the preceding “white” stress level. This is a value determined from the fact that these two accent phrases are a chain of “adjective / adnominal + noun”. After the stress level ratio is determined at each accent phrase boundary, the ratio of the stress level of all accent phrases to the stress level of the accent phrase at the beginning of the sentence is determined. For example, the ratio of “flower” to “he” is obtained as 1.2 × 0.7 = 0.84. Thus, given the stress level value of the accent phrase at the beginning of the sentence (for example, 80 Hz), the stress level values of all accent phrases in the sentence can be calculated. A pitch pattern of a sentence is generated by superimposing the thus obtained accent component on the speech tone component.

Although the description has been made based on the method of superimposing the accent component on the speech component, the present invention can be easily applied to a system that does not use the speech component. For example, instead of generating an accent component and a speech component using different models and superimposing them, at least one pitch frequency (for example, at the time when the pitch frequency reaches a peak) is given to each accent phrase, and based on the value, A device that determines the shape of the accent phrase and directly generates a pitch pattern of the entire sentence is conceivable. In this device, for example, if the peak value of the pitch frequency of each accent phrase is given, instead of the stress level ratio, the ratio of the peak value of the pitch frequency of the adjacent accent phrase is prepared for each part of speech combination. You should leave it.

In general, the part of speech is specified by morphological analysis,
It can be performed more accurately than the estimation of the dependency structure by the dependency analysis. According to the present invention, since the pitch pattern is generated using the information of the part of speech, errors are less likely to occur than in the conventional pitch pattern generation method, and a natural pitch pattern can be generated.

FIG. 1 is a block diagram showing an embodiment for realizing a pitch pattern generation device according to the present invention.
First, a character string representing a text to be read aloud is input from the character string input terminal 11. The input character string is sent to the morphological analyzer 12. The morphological analysis unit 12
The sentence represented by the input character string is decomposed into words, and the parts of speech and accent phrase boundaries of each word are determined. This result is sent to the accent component generator 13 and the speech component generator 15.

The stress level ratio storage unit 14 stores the ratio of the stress level of the accent phrase before and after the boundary for each combination of the parts of speech of the words before and after the boundary of the accent phrase. The accent component generation unit 13 reads the stress level ratio from the stress level ratio storage unit 14 based on the combination of the parts of speech of the words before and after each accent phrase boundary. Using the read stress level ratio, the stress level values of all the accent phrases in the text are determined by the method described in the section of the action, and an accent component is generated. The speech tone component generation unit 15 divides the input sentence into a plurality of speech tone components as necessary based on the analysis result of the morphological analysis unit 12, and then approximates the pitch immediately after the pitch frequency decreases as time elapses. The generated speech-tone component is generated. The pitch pattern generation unit 16 generates a pitch pattern of the entire sentence by adding the accent component and the speech component generated by the accent component generation unit 13 and the speech component generation unit 15, respectively, and outputs a pitch pattern output terminal. 17 to output.

FIG. 2 is a more detailed block diagram of the apparatus shown in FIG. In the figure, the elements denoted by the same reference numerals as those in FIG. 1 have the same functions. First, a character string representing a sentence to be read aloud is input from the character string input terminal 11. The input character string is a morphological analysis unit 121
Sent to The morphological analysis unit 121 divides the sentence represented by the input character string into words using the word dictionary 122, and determines the reading of each word, the part of speech, the accent type, and the accent phrase boundary position. Such a morphological analyzer is described in detail in U.S. Patent Application No. 07 / 487,044, and a description thereof will be omitted.

The word reading, part of speech, accent type, and accent phrase boundary position generated by the morphological analysis unit 121 are stored in an accent component model reading unit 131, a stress level ratio reading unit 133, and a phoneme duration storage unit 152.
Respectively. Accent component model storage unit 13
2 stores the outline of the pitch pattern for each accent type of the word. Accent component model reading unit 1
31 reads out the outline of the pitch pattern of the accent phrase stored in the accent component model storage unit 132 according to the accent type of each word sent from the morphological analysis unit 121, and reads the accent component model editing unit 134
Send to

As shown in FIG. 4, the stress level ratio storage unit 14 stores, for each part of speech combination before and after the accent phrase boundary, the value of the stress level ratio of the accent phrase before and after the boundary. ing. The stress level ratio reading unit 133 reads, from the stress level ratio storage unit 14, the stress level ratio corresponding to the combination of the parts of speech of the words before and after each accent phrase boundary. The accent component model editing unit 134 uses the stress level ratio read by the stress level ratio reading unit 133 to determine the stress level values of all the accent phrases in the sentence by the method described in the section of the action. And
The stress level of the pitch pattern for each accent phrase read by the accent component model reading unit 131 is changed to generate an accent component of the entire sentence.

Next, the phoneme duration calculation unit 151 calculates the duration of each phoneme to be uttered using the word reading obtained by the morphological analysis unit 121, that is, the sequence of phonemes. This can be realized, for example, by previously storing the average duration of each phoneme in the phoneme duration storage unit 152 and reading out the same. The breath paragraph length calculation unit 153 calculates the duration of the breath paragraph constituting the sentence. Here, the exhalation paragraph is a unit of utterance divided by a pause, and one exhalation paragraph forms one speech tone component. If the pause is 1 in the sentence
If none appear, the sentence is a single exhalation paragraph. If one pause appears in the sentence, the sentence is composed of two exhalation paragraphs. The determination of the position in the sentence where the pause is to be inserted is not directly related to the invention of the present application and will not be described. Expiration paragraph length calculation unit 15
3 is that by adding the duration of all phonemes included in the exhalation paragraph for each exhalation paragraph constituting the sentence,
Calculate the duration of the expiration paragraph.

The speech-tone component calculating section 154 stores a start-end frequency storage section 155 and an end-frequency storage section 156 for the pitch frequencies at the start and end of the speech tone necessary for determining the outline of the speech tone component.
Respectively. Furthermore, the expiration paragraph length calculation unit 15
The inclination of the speech component is calculated using the duration of the expiration paragraph calculated in step 3. That is, the gradient of the tone component [Hz / sec] = (the tone end frequency [Hz] -the tone start frequency [Hz] / the expiration paragraph duration time [sec]). Thus, the outline of the speech component is determined. Finally, the adder 160 calculates the pitch pattern of the input sentence by adding the accent component calculated by the accent component model editing unit 134 and the speech component calculated by the speech component calculation unit 154. , And output from a pitch pattern output terminal 17.

[0023]

As described above, in the present invention, the pitch pattern is determined without using the result of the dependency analysis, which is difficult to analyze correctly, so that a more natural pitch pattern can be generated as compared with the conventional method. It is possible to
Therefore, the present invention is extremely effective as a pitch pattern generation device in a speech synthesis device or the like that reads an arbitrary sentence given as a character string by voice.

[Brief description of the drawings]

FIG. 1 is a block diagram showing one embodiment for realizing a pitch pattern generation device according to the present invention.

FIG. 2 is a detailed block diagram of the embodiment shown in FIG. 1;

FIG. 3 is a diagram illustrating a method for generating a pitch pattern according to the present invention.

FIG. 4 is a diagram illustrating an example of a stress level ratio for each combination of parts of speech.

FIG. 5 is a diagram for explaining a pitch pattern generation method according to the related art.

FIG. 6 is a diagram for explaining the degree of separation.

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 11 Character string input terminal 12 Morphological analysis part 13 Accent component generation part 14 Stress level ratio storage part 15 Speech tone component generation part 16 Pitch pattern generation part 17 Pitch pattern output terminal 21 Accent component 22 Speech tone component 23 Pitch pattern

Claims (3)

(57) [Claims] 1. A stress pattern ratio generating unit in a speech synthesizer for reading an arbitrary sentence aloud, wherein a stress bell ratio storage unit for storing a ratio of the size of an adjacent accent phrase for each combination of part of speech. A morphological analysis unit that divides a sentence into words that constitute the sentence and determines a part of speech and an accent phrase boundary of each word; and the stress level ratio storage unit according to a combination of parts of speech of words before and after each accent phrase boundary. A pitch pattern generation unit comprising: an accent component generation unit that reads a ratio of the size of an accent phrase; and a pitch pattern generation unit that generates a pitch pattern based on the ratio of the size of the read accent phrase. apparatus. 2. The pitch pattern generation device according to claim 1, wherein the pitch pattern generation unit generates the pitch pattern by superimposing an accent component read from the accent component generation unit on a speech tone component of the sentence. . 3. The pitch pattern generation unit according to claim 1, wherein the pitch pattern generation unit gives at least one pitch frequency for each word, determines a shape of each word based on the value, and generates a pitch pattern of the entire sentence. Pitch pattern generator.