JP3397406B2 - Voice synthesis device and voice synthesis method - Google Patents

Description

Detailed Description of the Invention

[0001]

[Table of Contents] The present invention will be described in the following order. Industrial applications Conventional technology (Fig. 4) Problems to be Solved by the Invention Means for Solving the Problems (FIG. 1) Action (Fig. 1) Example (FIGS. 1 to 3) The invention's effect

[0002]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a voice synthesizing apparatus and a voice synthesizing method, and is particularly suitable for application to a text voice synthesizing apparatus and a text voice synthesizing method for synthesizing a voice from an input character sequence.

[0003]

2. Description of the Related Art Conventionally, a speech synthesizer generally performs speech synthesis by performing a process as shown in FIG. First, in the speech synthesizer 1, a Japanese text mixed with kanji and kana is input to the text input unit 2 from a predetermined input device, and the text is output to the sentence analysis unit 3.
The sentence analysis unit 3 analyzes the input text using the dictionary 4 as a reference, converts it into a phonetic kana character string, and then detects the accent type specific to the word and decomposes it into words and phrases.

That is, in Japanese, words are not separated into words like English, so words such as "US industry" are "US / industry / world" and "US /
It can be divided into two types, such as "domestic / industry".

For this reason, the sentence analysis unit 3 refers to the dictionary 4 and utilizes the continuity of words and the statistical property of words to decompose the text input into words and phrases. Detects bunsetsu boundaries. Furthermore, the sentence analysis unit 3 outputs to the symbol string analysis unit 5 a symbol string indicating the phonetic kana (information about phonemes) and bunsetsu boundaries and accent types (information about prosody) for each sentence thus obtained. .

The symbol string analysis unit 5 separates and extracts this symbol string into phoneme information and prosodic information, the phoneme information is output to the phoneme duration calculation unit 6 and the parameter connection unit 7, and the prosodic information is generated as a pitch pattern. It is output to the unit 8. Here, the phonological information is information about a sound to be uttered, and the prosody information is information about accents and intonation.

When the phoneme information is input to the phoneme duration calculating unit 6, the phoneme duration calculating unit 6 calculates the duration of each phoneme from the phoneme information. For example, as a method of calculating the phoneme duration, the vowel duration is calculated as

[Equation 1] As described above, there is a method of obtaining it by adding the average duration of vowels and the time length of variation from the average length of vowels in each factor of the phonological environment. This method was published in a March 2012 paper collection of the Acoustical Society of Japan (written by Kaiki et al.).

In this equation (1), α is a coefficient matrix and σ
Is a matrix of coefficients “1” or “0” that indicates which category of each factor the phonological environment of the vowel belongs to, and m is the number of factors (that is, the type of vowel, the type of the front phoneme, the type of the front-to-back phoneme,
The type of the rear phoneme, the type of the rear phoneme, the front, the consonant, the rear, the consonant, the long sound, the expiratory paragraph mora number, the expiratory paragraph position, the sentence mora number, the sentence position), and l represents the number of categories of each factor.

In this case, the parameter given as an input is the phonological environment of the phoneme to be obtained, and the variation from the average length of vowels is realized by analyzing and obtaining a large amount of sentences in advance.

The phoneme duration determined in this manner is output to the parameter connection unit 7 and the pitch pattern generation unit 8. The parameter connection unit 7 connects the phoneme piece data read from the phoneme piece database 9 based on the phoneme information and the calculated duration of each phoneme to generate a parameter string. The generated parameter string is output to the voice synthesizer 10.

On the other hand, the pitch pattern generator 8 generates a pitch pattern based on the prosodic information and the calculated duration of each phoneme, and the generated pitch pattern is output to the voice synthesizer 10.

Here, as an example of pitch pattern calculation,

[Equation 2] There is a model in which the pitch pattern is divided into a phrase component and an accent component, and each is represented by the sum of the secondary critical response outputs driven by an impulse input and a step input as shown in FIG. This model was published in January, 1989
ol.J72-A, No.1 (Fujisaki et al., "Synthesis of Sentence Speech Based on Fundamental Frequency Pattern Generation Process Model").

In this equation (2), G _pi (t), G
_aj (t) is the impulse response of the phrase control mechanism,
It is the step response of the accent control mechanism, and the impulse response and the step response are as follows in the range of t ≧ 0.

[Equation 3] And the following equation

[Equation 4] Becomes

In the equations (3) and (4), F
_min is the lowest pitch frequency, α, β, θ are constants, A _pi is the size of the phrase command, T _li is the position of the phrase command, A _aj
Is the size of the accent command, T _1j is the start position of the accent command, and T _2j is the end position of the accent command. When t <0, G _pi (t) = G _aj (t) = 0.

In this case, the parameters given as inputs are coefficients that determine the time, magnitude and response speed of the input. These values are obtained by analyzing a large amount of sentences in advance and held, and a pitch pattern is generated using appropriate values according to a predetermined rule. Next, the voice synthesizing unit 10 performs a waveform synthesizing process based on the parameter sequence and the pitch pattern, and outputs a synthetic voice signal via the digital analog converting unit (D / A unit) 11.

[0016]

By the way, in such a conventional text-to-speech synthesizing apparatus 1, the pitch pattern is calculated by combining the accent type of each word previously described in the dictionary for text analysis and the word. According to the accent change rule in the case of Natsu, the accent of the phrase is obtained and output to the symbol string. In this case, when words of the same accent type and the same number of mora appear at the same position in the sentence, the accent patterns required are the same.

The phoneme duration is calculated in consideration of the phoneme environment before and after the phoneme to be calculated, the phoneme environment before and after the phoneme, and the position in the sentence, but in the case of different words in the same phoneme environment. The phoneme duration is the same.

However, the pitch pattern when a person actually utters is such that even if words with the same accent type and the same number of mora appear at the same position in the sentence, they are hardly spoken in the same pitch pattern. Even if the phoneme duration is microscopically the same, the phoneme duration is unlikely to be the same when considering a macro phoneme environment such as a word unit. Therefore, there is a problem in that the synthesized voice of the conventional text-to-speech synthesizer becomes monotonous.

The present invention has been made in view of the above points, and an object thereof is to propose a voice synthesizing apparatus and a voice synthesizing method capable of obtaining a synthetic voice close to a human voice.

[0020]

In order to solve such a problem, the present invention is obtained by analyzing a dictionary 24 holding word-specific phoneme duration information and / or accent command values, and an inputted sentence. A symbol string analysis means 25 for separating and extracting the symbol string data of the sentence into phoneme information and prosody information, and a phoneme continuation if necessary to a value obtained by calculating the duration of the phoneme information for a given phoneme. Phoneme duration calculating means 26 for calculating phoneme duration by replacing with phoneme information, parameter producing means 27 for generating a parameter string based on phoneme information and phoneme duration, phoneme duration and prosody information. , Or a pitch pattern generating means 28 for generating a pitch pattern based on the phoneme duration, the prosody information and the accent command value.
And a voice synthesizing means 30 for synthesizing a voice waveform from the parameter sequence and the pitch pattern.

In order to solve such a problem, in the present invention, the phoneme duration information and / or the accent command value peculiar to a word is held in the dictionary 24, and the symbol string of the sentence obtained by analyzing the inputted sentence is obtained. The data is separated and extracted into phonological information and prosodic information, and the value obtained by calculating the duration of the phoneme information for a given phoneme is replaced with the phoneme duration information as necessary to determine the phoneme duration. Calculate, to generate a parameter string based on the phoneme information and phoneme duration, to generate a pitch pattern based on the phoneme duration and prosody information, or the phoneme duration and prosody information and accent command value, A voice waveform is synthesized from the parameter string and the pitch pattern.

[0022]

The phoneme duration calculated by replacing the value obtained by calculating the duration of the phoneme information with respect to a predetermined phoneme with the phoneme duration information unique to the word held in the dictionary 24, and Synthesizes the speech waveform based on the word-specific accent command value held in the dictionary 24, so that even if words with the same accent type and the same mora number appear at the same position in the sentence, or if different words are the same. Even in the phonological environment, it is possible to reliably obtain the pitch pattern peculiar to each word.

[0023]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS An embodiment of the present invention will be described in detail below with reference to the drawings.

In FIG. 1, reference numeral 21 denotes a speech synthesizer as a whole, which is adapted to output a voice close to human utterance by processing the text input to the text input unit 22 by each predetermined processing unit. There is. Here, the speech synthesizer 21 has the same configuration as the conventional speech synthesizer 1, but the information added to the dictionary and the processing content of each processing unit using these information are different.

First, a text input unit 2 is input from a predetermined input device.
A Japanese text mixed with kanji and kana is input to 2 and is output to the sentence analysis unit 23. The sentence analysis unit 23 analyzes the input text using the dictionary 24 as a reference, converts it into a phonetic kana character string, and then detects a word-specific accent type and decomposes it into words and phrases.

That is, the sentence analysis unit 23 uses the continuity of words and the statistical property of words while referring to the dictionary 24, and decomposes the text input into words and phrases. Detect the boundaries of. Further, the sentence analysis unit 23 outputs to the symbol string analysis unit 25 a symbol string indicating the reading kana (information on phoneme), bunsetsu boundary, and accent type (information on prosody) of each sentence thus obtained. .

Here, in the dictionary 24, in addition to the information indicating the accent type, the accent command value peculiar to the word and the phoneme duration information are held in advance, and the phoneme duration calculating unit 26 reads out the phoneme duration information. The pitch pattern generator 27 reads the accent command value. Next, the symbol string analysis unit 2
Reference numeral 5 separates and extracts this symbol string into phoneme information and prosody information, outputs the phoneme information to the phoneme duration calculation unit 26 and the parameter connection unit 27, and outputs the prosody information to the pitch pattern generation unit 28.

The phoneme duration calculation unit 26 calculates the duration of each phoneme from the phoneme information from the symbol string analysis unit 25 and the phoneme duration information from the dictionary 24. For example,
The case where the duration of the third vowel (ie, “i”) of the “white flower” is obtained when the text “white flower (shiroihana)” is input will be described below.

The duration of this vowel "i" can be obtained from the equation (1). That is, specifically, the vowel "i"
Duration = "coefficient for" / i / + previous phoneme is / o "
Coefficient in case of / + coefficient in case where phoneme in front of is / r /
The coefficient when the subsequent phoneme is / h / + the later phoneme is / a /
Coefficient in case of + no coefficient in front + coefficient in case of no subsequent sound + coefficient in case of no long tone + coefficient in case of mora length of expiratory paragraph "s" + coefficient in case of expiratory paragraph "
+ Coefficient when the mora length N of the sentence is + coefficient when in the sentence

Next, the term enclosed by "" in this expression is previously stored in the dictionary 2.
By replacing it with the value held in 4, the duration of the vowel / i / is calculated. That is, the duration of the vowel "i" =
“Value read from dictionary 24” + coefficient in case of mora length N of sentence + coefficient in case of sentence.

In this way, the phoneme duration information necessary for obtaining the duration of each phoneme is prepared in advance in the dictionary 24,
If the duration of each phonological unit is calculated from the expression (1) using this phonological duration information, even if different words have the same phonological environment, a natural duration corresponding to each word is calculated. be able to. Here, since the unknown word has no duration information, in the case of an unknown word, the duration of the phoneme is calculated by the conventional method.

The duration time of each phoneme calculated in this way is output to the parameter connection unit 27 and the pitch pattern generation unit 28. In the parameter connection unit 27, the phoneme information from the symbol string analysis unit 25 and the phoneme duration calculation unit 26.
The phoneme piece data selected from the phoneme piece database 29 is connected based on the phoneme duration time calculated in (3) to generate a parameter string. The generated parameter string is output to the voice synthesizer 30.

On the other hand, in the pitch pattern generation unit 28, based on the prosody information from the symbol string analysis unit 25, the accent command value from the dictionary 24, and the phoneme duration calculated by the phoneme duration calculation unit 26, Generate a pitch pattern.

FIG. 2 shows a conventional process for generating a pitch pattern when there is one phrase command and one accent command, for example. As described above, in the conventional pitch pattern generation process, the parameter given as an input is a value obtained by analyzing a large amount of sentences in advance, and a value suitable for each is selected according to a predetermined rule, and this selection is performed. These values are used as the accent command size A1, the start position t1, the end position t2, and the response speed β1 to generate a pitch pattern.

On the other hand, in this embodiment, the dictionary 24 is prepared in advance.
The accent command value peculiar to each word is stored in the dictionary, and the accent command size A2 of a predetermined word is stored in the dictionary 24.
The start position t3, the end position t4, and the response speed β2 are read out, and a pitch pattern as shown in FIG. 3 is calculated using these to obtain a word-specific pitch pattern.

In this way, if a unique accent command value for each word is prepared in the dictionary 24 in advance and a pitch pattern is generated by using this accent command value, words with the same accent type and the same number of mora are in the sentence. Even if they appear at the same position, the accent pattern peculiar to each word can be obtained, so that the pitch pattern peculiar to the word can be obtained. In the case of an unknown word, the accent command value is calculated by the conventional method as in the case of calculating the phoneme duration.

The pitch pattern generated in this way is output to the voice synthesizing section 30, and the voice synthesizing section 30 synthesizes a voice waveform from the parameter sequence and the pitch pattern, and a synthesized voice signal is generated via the D / A section 31. Output.

In the above structure, the dictionary 24 holds phoneme duration information and accent command values in advance.
When a word registered in the dictionary 24 is input to the text input unit 22, the parameter connection unit 27 calculates the phoneme duration and the phoneme duration calculated using the phoneme duration information stored in the dictionary 24. A parameter string of phoneme piece data is generated based on the phoneme information. In the pitch pattern generation unit 28, the calculated phoneme duration,
A pitch pattern is generated based on the phoneme information and the accent instruction book held in the dictionary 24.

Since the parameter sequence and pitch pattern generated in this way become the parameter sequence and pitch pattern unique to each word, the synthesized speech obtained by synthesizing these parameter sequence and the pitch pattern is much more human voiced. The synthesized speech is close.

According to the above construction, the accent command value of each word and the phoneme duration information are additionally held in the text analysis dictionary 24. When a word registered in the dictionary 24 is input, a parameter string is generated based on the phoneme duration calculated using the phoneme duration information, and the calculated phoneme duration and the phoneme duration are calculated. By generating a pitch pattern based on the accent command value stored in the dictionary 24, it is possible to output a synthesized voice that is much closer to human speech.

In the above embodiment, the case where the dictionary 24 holds both the accent command value and the phoneme duration information has been described, but the present invention is not limited to this, and only one of the information is stored in the dictionary. It may be held at 24.

[0042]

As described above, according to the present invention, the value obtained by calculating the duration of phoneme information with respect to a predetermined phoneme is added to the phoneme duration information unique to the word stored in the dictionary. Since the speech waveform is synthesized based on the phoneme duration calculated by replacement and the word-specific accent command value stored in the dictionary, even words with the same accent type and the same number of mora have the same position in the sentence. , Or even when different words are in the same phonological environment, it is possible to reliably obtain a pitch pattern peculiar to each word, and thus it is possible to output synthetic speech that is much closer to human speech. it can.

[Brief description of drawings]

FIG. 1 is a block diagram showing a functional configuration of an embodiment of a text-to-speech synthesizer according to the present invention.

FIG. 2 is a characteristic curve showing a pitch pattern calculated by a conventional text-to-speech synthesizer.

FIG. 3 is a characteristic curve showing a pitch pattern calculated by the text-to-speech synthesizer in the embodiment.

FIG. 4 is a block diagram showing a functional configuration of a conventional text-to-speech synthesizer.

[Explanation of symbols]

1, 21 ... Speech synthesizer, 2, 22 ... Text input section, 3, 23 ... Sentence analysis section, 4, 24 ... Dictionary, 5,
25 ... Symbol string analysis unit, 6, 26 ... Phonological duration calculation unit, 7, 27 ... Parameter connection unit, 8, 28 ... Pitch pattern generation unit, 9, 29 ... Phoneme unit database, 10, 30 ... Voice synthesizer, 11, 31 ... D / A
Department.

Continuation of front page (56) Reference JP-A-1-126695 (JP, A) JP-A-5-181491 (JP, A) JP-A-5-289686 (JP, A) JP-A-5-289688 (JP , A) (58) Fields investigated (Int.Cl. ⁷ , DB name) G10L 13/08

Claims (2)

(57) [Claims] 1. A dictionary holding word-specific phoneme duration information and / or accent command values, and a symbol string of the sentence obtained by analyzing an inputted sentence.
Symbol string analysis means for separating and extracting data into phoneme information and prosodic information, and calculating a duration of the phoneme information for a given phoneme
If necessary, continue the above phoneme to the value obtained by
Phoneme duration calculating means for calculating phoneme duration by replacing with phoneme information, parameter generating means for generating a parameter string based on the phoneme information and phoneme duration, the phoneme duration and Pitch pattern generating means for generating a pitch pattern based on the prosodic information, or the phoneme duration, the prosody information and the accent command value, and a voice synthesizing means for synthesizing a voice waveform from the parameter string and the pitch pattern. A voice synthesizer comprising: 2. A symbol string of the above sentence, which holds phoneme duration information and / or accent command value peculiar to a word and is obtained by analyzing the inputted sentence.
Data is separated into phonological information and prosodic information and extracted, and the duration of the phonological information for a given phonological element is calculated.
If necessary, continue the above phoneme to the value obtained by
The phoneme duration is calculated by replacing it with time length information, and a parameter string is generated based on the phoneme information and the phoneme duration, and the phoneme duration and the prosody information, or the phoneme duration and A voice synthesizing method comprising generating a pitch pattern based on the prosody information and the accent command value, and synthesizing a voice waveform from the parameter string and the pitch pattern.