JP2679623B2 - Text-to-speech synthesizer - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a text-to-speech synthesizer for reading a text written in characters by voice.

[0002]

2. Description of the Related Art In a conventional text-to-speech synthesizer, when reading a text consisting of a plurality of sentences, every sentence is read with a similar rhythm and intonation.

FIG. 3 is a block diagram showing an embodiment of a conventional text-to-speech synthesizer. Text consisting of a plurality of sentences is input to the sentence cutout unit 102 through the text input terminal 101. The sentence cutout unit 102
Find the character that separates sentences from the text and divide the text into sentences. Characters that can be used as sentence delimiters include "." And ".", But other characters may be defined in advance. The divided sentences are processed by the prosody generation unit 3
Sent to 00. The prosody generation unit 300 independently generates the prosody pattern of each sentence. That is, the prosody pattern is generated based on the characteristics of the sentence, such as the length of the sentence or the structure of the sentence, but the information of other sentences before and after the sentence is not used. The generated prosody pattern is used by the speech synthesis unit 10
Sent to 6. The voice synthesis unit 106 generates a synthetic voice of each sentence based on the prosodic pattern of each sentence, and outputs the synthetic voice from the synthetic voice output terminal 107.

Regarding the text-to-speech synthesizer using such a conventional technique, the Acoustical Society of Japan 1993 Autumn Research Presentation Meeting Proceedings 2-8-13 "Software Japanese Text-to-Speech Synthesis for PC" (Reference 1) And so on.

[0005]

In the above-mentioned conventional text-to-speech apparatus, the prosodic features when reading a sentence are always constant. The prosodic features mentioned here refer to values such as reading speed, voice pitch, and voice volume, and patterns of changes thereof. Even if each sentence is a device that can spontaneously speak, if every sentence is always read out with a constant rhythm and intonation, it gives a monotonous impression to the user. In order to avoid monotone reading, it is necessary to change prosodic features along the flow of text. However, in the conventional text-to-speech synthesizer, the user has to give some instruction to the device in order to change the prosodic features.

On the other hand, the present invention provides a text-to-speech synthesizer capable of automatically changing the prosodic characteristics of each sentence along with the flow of input text and enabling a reading closer to human beings. The purpose is to do.

[0007]

The text-to-speech synthesizer of the first invention is a text input terminal for inputting a text consisting of a plurality of sentences, and a sentence cutout for connecting the text input terminal to divide the text into sentences. Means, a prosody generation means connected to the sentence cut-out means to generate a prosody pattern of the sentence on the basis of a predetermined criterion and used as an operation prosody pattern, and a voice connected to the prosody generation means according to the operation prosody pattern. A text-to-speech synthesizer having a speech synthesizing means for synthesizing , wherein the prosody generation means is connected to the sentence cut-out means and generates a prosody pattern of the sentence on the basis of a predetermined standard to obtain a standard prosody pattern. A prosody generation means, and a prosody pattern storage means that is connected to the standard prosody generation means and stores the standard prosody pattern of all the sentences forming the text. A prosodic pattern changing means for this prosodic pattern storage means connected to a predetermined reference according to each statement of the modification of the standard prosodic pattern production prosody pattern
In a text-to-speech synthesizer including

[0008] The prosody pattern changing means is configured to change the prosody pattern.
Of all the sentences that make up the text by connecting to the
Calculate an approximate curve that approximates the change in the standard prosodic pattern
Difference between the approximate curve and the standard prosodic pattern of each sentence
The approximation curve calculating means for calculating the approximation error
It is connected to the similar curve calculation means, and based on the approximation error of each sentence
An approximation that transforms the standard prosody pattern into an operational prosody pattern
And an error changing unit .

The text-to-speech synthesizer of the second invention is
In the invention of 1, the approximate music calculated by the approximate curve calculation means
The line is a regression line.

The text-to-speech synthesizer of the third invention is
In the first or second invention, the approximation error changing means is
When the approximation error is set to E, the predetermined constants a and b are used.
Operation approximation error E ′ = aE + b
The operational approximation error is added to generate an operational prosodic pattern.
It is characterized by the following.

The text-to-speech synthesizer of the fourth invention is
In the first, second or third invention, the prosody pattern changing hand
Check that the prosodic pattern that the dan changes is the speaking rate of each sentence.
Features.

The text-to-speech synthesizer of the fifth invention is
In the first, second or third invention, the prosody pattern changing hand
The prosodic pattern that the dan changes is the average pitch frequency of each sentence.
It is characterized by that.

The text-to-speech synthesizer of the sixth invention is
In the first, second or third invention, the prosody pattern changing hand
The prosodic pattern that the dan changes is the average power of each sentence
It is characterized by.

[0014]

The physical parameters related to the prosodic features of speech include phoneme duration, pitch frequency pattern,
The amplitude pattern etc. can be raised. When reading a text, humans vary these prosodic features along the flow of the text. The present invention modifies the prosodic features of each sentence in a text consisting of a plurality of sentences, and realizes reading aloud similar to what humans do. The method will be described in detail below.

First , a text consisting of a plurality of input sentences is divided into individual sentences, and prosodic patterns are generated for each sentence. As a method of generating a prosodic pattern of a sentence, the method described in Document 1 can be used. Next, the prosodic patterns of all the sentences that make up the text are temporarily stored. At this time, the prosodic pattern of each sentence varies to some extent depending on the peculiar characteristics of the sentence such as the length of the sentence. However, since the variation is small, it gives the impression that the text is read in the same tone throughout the text. Therefore, the prosodic pattern of each sentence that has been temporarily stored is transformed according to a predetermined standard to obtain the prosodic pattern of the synthetic speech to be output.

The first, second, and third inventions provide a method of transforming the prosodic pattern of each sentence once accumulated. In the first aspect of the invention, an approximate curve is first obtained as a method of modifying the variation of the prosody pattern of each sentence. This approximate curve represents a large flow of prosodic patterns throughout the text. Next, the approximation error of the prosodic pattern of each sentence by this approximation curve is obtained. This can be regarded as a deviation from the overall flow caused by a characteristic unique to each sentence (for example, sentence length, sentence structure, etc.). By subjecting this deviation, that is, processing that emphasizes the approximation error, the prosodic pattern of the entire text is changed. In the second invention, a regression line is used as the approximation curve. In the third invention, when the approximation error is E,
Using the predetermined constants a and b, the transformation of E ′ = aE + b (1) is performed. This constant a serves to expand the approximation error of each sentence of the input text when its absolute value is larger than 1 and to reduce it when it is smaller than 1. The constant b serves to translate the prosodic pattern of the entire input text.
That is, for example, when the target prosodic pattern is the utterance speed, if a positive number is given to b, the reading speed of the entire text becomes faster. On the contrary, if a negative number is given to b, the reading speed becomes slow.

In the fourth, fifth, and sixth inventions, the prosodic patterns to be transformed are the speech rate, the average pitch frequency, and the average power of each sentence, respectively.

As a method of transforming the approximation error E by the regression line, E '= 3E (2) is used to change the speech rate of each sentence.
The operation principle of the present invention will be described in more detail. FIG. 4 is a diagram for explaining the operation principle of the present invention. As input text, consider a text consisting of six sentences.

First, the phoneme duration is calculated for each sentence constituting the text. As this method, for example, a conventionally known method described in Document 1 can be used. The duration of each phoneme of the sentence is added to obtain the duration of the entire sentence.
The number of moras of a sentence is divided by the duration of the whole sentence to obtain the speech rate (mora / second) of the sentence. It is assumed that the pattern indicating the speech rate of each sentence thus obtained is like the standard speech rate pattern 401 in FIG.

Next, a regression line 402 that approximates the relationship between the standard speech rate pattern 401 and the sentence number is obtained. This regression line shows the tendency of the rough change in the speaking rate throughout the input text. An approximation error E'is obtained by transforming the difference (approximation error) E between the regression line 402 and the standard speech rate of each sentence by the equation (2). Regression line 402
The pattern of the speech speed, which is obtained by adding the approximate error E ′ after transformation to
The operation speech rate pattern 403 is set. The operation speech rate pattern 403 obtained in this way is based on the original standard speech rate pattern 401, which indicates that the change tendency of the speech rate of the entire text is the original standard speech rate pattern 403.
Standard speech rate pattern 401 while maintaining the tendency of
Will change more than. Then, the phoneme duration of each sentence is recalculated so that the utterance speed of each sentence becomes the value indicated by the operational utterance speed pattern 403. As a result, it becomes possible to change the reading of text, which was monotonous in the conventional method due to the repetition of the same speech rate.

Although the utterance speed has been described as an example of the prosodic pattern, the average pitch frequency and the average power of each sentence can be treated in the same manner to obtain the same effect.

Further, although the approximation by the regression line is shown as an example, other approximation methods can be used. Regarding the method of transforming the approximation error, the method using only the approximation error of the sentence has been described here, but a method of utilizing the approximation error of the preceding and following sentences is also conceivable.

By using the above method,
It is possible to automatically change the prosody pattern of each sentence of the input text. As a result, it becomes possible to output a natural read-aloud voice, which was impossible with the conventional text-to-speech synthesizer.

[0024]

1 is a block diagram showing a reference example of a text-to-speech synthesizer according to the present invention. A text composed of a plurality of sentences is input to the sentence cutout unit 102 through the text input terminal 101. The sentence cutout unit 102 searches the input text for a character that serves as a sentence break. Then, the input text is divided into the sentences that compose it by the characters that serve as the delimiters, and the standard prosody generation unit 103
Send to The standard prosody generation unit 103 generates a standard prosody pattern of each sentence sent from the sentence cutout unit 102 using a conventionally known method or the like, and sequentially sends these to the storage unit 104. The accumulation unit 104 accumulates the prosody pattern of each sentence sent from the standard prosody generation unit 103. When the prosody pattern of the last sentence of the input text is sent from the standard prosody generation unit 103, all the stored prosody patterns are sent to the changing unit 105. Change part 1
05 uses the information of all the prosodic patterns of the input text to modify the prosodic patterns of the respective sentences to generate the final operational prosodic pattern used when synthesizing the voice, and sends it to the voice synthesizer 106. . The voice synthesizing unit 106 generates a synthetic voice according to the operational prosody pattern sent from the changing unit 105, and outputs it to the synthetic voice output terminal 107.

FIG. 2 is a block diagram showing an embodiment of a text-to-speech synthesizer according to the first invention. Change unit 20
5 is an approximate curve calculation unit 201 and an approximate error changing unit 20.
It is configured to include 2. The standard prosody patterns of all the sentences forming the input text, which are stored in the storage unit 104, are sent to the approximate curve calculation unit 201. The approximate curve calculation unit 201 obtains an approximate curve that approximates the relation between the sentence number of each sentence of the input text and the prosodic pattern of each sentence, and an approximation error due to the approximate curve. Here, the second invention particularly uses a regression line as the approximated curve. The approximated curve and the approximated error obtained are approximated by the approximated error changing unit 202.
Sent to The approximation error changing unit 202 changes the approximation error calculated by the approximation curve calculating unit 201 by a predetermined method. The prosody pattern of each sentence is recalculated from the changed approximation error and the approximation curve sent from the approximation curve calculation unit 201, and this is used as the operational prosody pattern. At this time, with respect to the approximation error calculated by the approximation curve calculation unit 201,
A third aspect of the invention is a modification of the equation (1) described in the section of the action. The operation of the other parts is the same as that of the above-mentioned reference example, and therefore the description thereof is omitted.

Modification unit 1 in the embodiment of FIG. 1 or FIG.
The prosody pattern changed in 05 or 205 is the utterance speed of each sentence constituting the input text.
4th invention.

Modification unit 1 in the embodiment of FIG. 1 or FIG.
The fifth invention is that the prosody pattern changed in 05 or 205 is the average pitch frequency of each sentence constituting the input text.

Further, the sixth invention is that the prosody pattern changed by the changing unit 105 or 205 in the embodiment of FIG. 1 or FIG. 2 is the average power of each sentence constituting the input text.

[0029]

As described above, according to the present invention, when a text consisting of a plurality of sentences is input, prosodic features such as rhythm and intonation of each sentence constituting the text are automatically detected. It is possible to read aloud by changing the input. Therefore, it is very effective for realizing a text-to-speech synthesizer that needs to read a large amount of text, such as a reader.

[Brief description of the drawings]

FIG. 1 is a block diagram showing a reference example of a text-to-speech synthesizer according to the present invention.

FIG. 2 is a block diagram showing an embodiment of a text-to-speech synthesizer according to the first invention.

FIG. 3 is a block diagram showing an embodiment of a conventional text-to-speech synthesizer.

FIG. 4 is a diagram for explaining the operating principle of the present invention.

[Explanation of symbols]

 100, 200, 300 Prosody generation unit 101 Text input terminal 102 Sentence extraction unit 103 Standard prosody generation unit 104 Accumulation unit 105, 205 Change unit 106 Speech synthesis unit 107 Synthetic speech output terminal 201 Approximate curve calculation unit 202 Approximation error change unit 401 Standard Speech rate pattern 402 Regression line 403 Operation speech rate pattern

Claims (6)

(57) [Claims] 1. A text inputting text consisting of a plurality of sentences.
Connected to the text input terminal and the text input terminal.
A sentence segmentation method that divides the text into sentences, and this sentence segmentation
Connected to the means and based on a predetermined standard, the prosody of the sentence
A prosody generating means for generating a pattern to be an operational prosody pattern,
By connecting to this prosody generation means, the above-mentioned operational prosody pattern was used.
Sound having a voice synthesizing means for synthesizing voice
A voice synthesizer, wherein the prosody generation means includes the sentence segmentation
It is connected to the sending means and the rhyme of the sentence is based on predetermined criteria.
A standard prosody generator that generates a temperament pattern and makes it a standard prosody pattern.
Dan and connect to this standard prosody generation means to construct the text.
Prosodic pattern that accumulates the standard prosodic pattern of all sentences
Tan storage means and this prosody pattern storage means
Change the standard prosodic pattern of each sentence according to the defined criteria.
Including prosody pattern changing means for shaping and operating prosody pattern
In the text-to-speech synthesizer, the prosody pattern changing means is stored in the prosody pattern accumulating means.
The standard prosody of all sentences that connect and form the text
Calculate an approximate curve that approximates the change in pattern and
Approximation error, which is the difference between the similarity curve and the standard prosodic pattern of each sentence
Approximate curve calculating means for calculating the difference and this approximate curve calculating means
The standard prosody of each sentence based on the approximation error.
Approximation error changing means for transforming a pattern into an operational prosodic pattern
A text-to-speech synthesizer comprising: 2. An approximate curve calculated by the approximate curve calculating means.
Is a regression line, and the textile according to claim 1,
Stroke synthesizer. 3. The approximation error changing means removes the approximation error.
When E is set, the operation approximation error is made by using the predetermined constants a and b.
The difference E ′ = aE + b is obtained, and the operational approximation is made to the approximate curve.
It is characterized by generating an operational prosodic pattern by adding errors.
The text-to-speech synthesizer according to claim 1 or 2. 4. The prosody pattern changed by the prosody pattern changing means.
Tan is the speech rate of each sentence.
The text-to-speech synthesizer according to 1, 2, or 3. 5. The prosody pattern changed by the prosody pattern changing means.
Ton is the average pitch frequency of each sentence
The text-to-speech synthesizer according to claim 1, 2 or 3. 6. The prosody pattern changed by the prosody pattern changing means.
Tan is the average power of each sentence.
The text-to-speech synthesizer according to 1, 2, or 3.