JPH0836397A - Voice synthesizer - Google Patents

Abstract

PURPOSE:To provide a voice synthesizer which has high quality, requires a smaller amount of memory and accommodates for the generation of voices having various tone quality. CONSTITUTION:The synthesizer is provided with a natural voice element storage means 2 and a feature parameter storage means 3 which extracts parameter vectors that characterize acoustic characteristics such as a vowel normal section against each element of natural voice element sets and stores the position information of the parameter vectors along with the vectors themselves. The feature parameter vectors corresponding to object natural voice elements are made to be targets. A voice synthesizing means 6 synthesizes the voice transition from the targets to desired phonemes, connects them to the natural voice elements so that soft and various tone qualities and intonation are added relative to vowels and high quality voices are provided for consonants. Moreover, the required capacity for the storage of voice elements is made smaller and the connections between voice elements are made smoother.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a speech synthesizer,
In particular, the present invention relates to a voice synthesizer that converts arbitrary contents into voice.

[0002]

2. Description of the Related Art There are roughly two types of speech synthesis methods for converting arbitrary contents into speech. One is a synthesizing method that understands the vocalization mechanism of the voice, that is, the movements of the vocal cords, the mouth, and the throat, and controls the electric circuit by using the knowledge as a rule. The other method is a method that does not require much knowledge of speech, and prepares a large number of speech segments and connects the appropriate segments according to the input. The former is well known, for example, a combination of a formant synthesis method and a formant control rule. FIG. 6 is a structural example of a combination of the formant synthesis method and the formant control rule. In the figure, a formant synthesizer control rule storage unit 9 stores a plurality of rules for controlling the formant synthesizer, and a formant synthesizer control coefficient generation unit 8 is a formant synthesizer based on the control rule. , A formant synthesizer 10 actually synthesizes speech, a voiced sound source 11 simulates vocal cord vibration, and a series formant synthesizer 13 forms formant resonators in series. The unvoiced sound source unit 12 that connects and synthesizes voiced sounds such as vowels and nasal sounds
Is a turbulence noise source required for synthesis of fricatives and plosives, and the parallel formant synthesizer 14 synthesizes unvoiced consonants such as fricatives and plosives with resonators connected in parallel. Synthesis unit 15
Is a part for synthesizing the output of the serial formant synthesizer 13 and the output of the parallel formant synthesizer 14 and outputting a synthesized sound.

When information about phonetic symbols, accent positions, and intonation necessary for speech synthesis is input to the formant synthesizer control coefficient generator 8, the formant synthesizer control coefficient generator 8 stores the formant synthesizer control rule. The unit 9 refers to a necessary rule and outputs the formant synthesizer control coefficient to the formant synthesizer 10. The inside of the formant synthesizer 10 operates as follows. The voiced sound source unit 11 simulates a pulse-shaped sound source waveform generated in a human vocal cord when synthesizing a voiced sound such as a vowel. This pulsed signal is input to the serial formant synthesizer 13,
The in-line formant synthesizer 13 gives the sound source waveform an appropriate feature as a vowel or a nasal sound by using a plurality of formant resonators, and outputs it to the synthesizer 15. On the other hand, the unvoiced sound source unit 12 sends a noise-like waveform, which is a sound source of fricative or plosive sound, to the parallel formant synthesis unit 14, and the parallel formant synthesis unit 14 uses a plurality of resonators to obtain the frequency characteristics required for each consonant. Are instantly formed and output to the combining unit 15. Synthesis part 5
Outputs the vowels and nasal sounds of the serial formant synthesizer 13 and the consonants of the parallel formant synthesizer 14 and outputs them as synthesized speech.

The advantage of this method is that since all the sounds are created according to rules, it is highly flexible and can synthesize sounds of various tones and intonations. The disadvantage is that it is difficult to generate natural-sounding consonants, especially under the present circumstances where the synthesis rules are not clear for voices with complex vocalization mechanisms such as consonants.

Next, a method of using another conventional speech unit will be described. FIG. 7 is a block diagram of this system. The speech unit selection unit 16 selects, from the speech unit database storage unit 17, a speech unit required for synthesis from the phonetic symbol string and the accent information that are input. In this case, the speech unit is stored after being compressed into a coefficient such as a linear prediction coefficient, or after being subjected to a simple compression process on the time waveform. The selected plurality of speech units are connected by the unit connection synthesizing unit 18 and synthesized into a speech waveform at an appropriate fundamental frequency.

The advantage of this method is that the speech unit is basically cut out from the natural speech that is the model, so that the synthesis quality is extremely high if a smooth connection between the units is possible.
On the other hand, the disadvantage of this method is that it requires a large-capacity storage device to store the speech units, which increases the cost. In addition, there is a problem in that only the voice quality of the model voice can be synthesized and the flexibility is lacking.

Therefore, the method shown in FIG. 8 will be described as a method of merging the above two methods. In the figure, the voiced sound source section 11 is a section for simulating the vibration of the vocal cords and generating a sound source signal. The serial formant synthesis unit 13 is a unit that synthesizes a voiced sound such as a vowel. The voice unit database storage unit 17 stores a voice unit of a consonant cut out from a natural voice, the voice unit selection unit 16 selects and extracts a necessary voice unit, and the unit connection synthesis unit 18 This is a part that synthesizes the output of the serial formant synthesis unit 13 and the output of the speech unit selection unit 16 and outputs the synthesized speech.

The operation of the third conventional example constructed as described above will be described below. The voiced sound source section 11 generates a desired sound source signal from the information about the fundamental frequency in the coefficient for controlling the formant synthesizer, the information about the amplitude information of the sound source, and inputs the desired sound source signal to the serial formant synthesizer 13.
No sound source signal is output in the consonant section or the unvoiced section. The series formant synthesizer 13 determines the characteristics of the resonators arranged in series from the formant frequency information in the formant synthesizer control coefficient, the information about the bandwidth of the formant resonance peak, and the like, and determines the vowel sound from the above sound source signal. Etc. to a voice signal. The output of the serial formant synthesis unit 13 is sent to the segment connection synthesis unit 18. On the other hand, from the information on phonemes in the formant synthesizer control coefficients, the speech unit selection unit 16
Confirms whether or not the phoneme exists in the speech unit database storage unit 17, and if it exists, takes out the speech unit from the speech unit database storage unit 17 and sends it to the unit connection synthesis unit 18. For example, when the phoneme to be synthesized is “k” and the subsequent vowel is “a”, the speech unit selection unit 1
Reference numeral 6 searches the voice unit database storage unit 17 for a voice unit cut out from "ka" with the consonant "k". The segment connection synthesizing unit 18 includes the series formant synthesizing unit 13
The vowel signal from and the consonant signal of the speech unit selection unit 16 are combined by addition processing or superposition processing. By configuring in this way, it is possible to give various sound quality and intonation for vowels by the formant synthesis method, and for consonants, high quality speech that cannot be realized by the formant synthesis method by using the voice unit is provided. it can. Further, since storage as a speech unit is limited to syllables having a short duration, it can be realized by a small-capacity storage device. However, since there is a considerable difference between the sound quality from the formant synthesis section and the sound quality of the segment, there is a problem of unnaturalness and deterioration at the connection part.

[0009]

As described above in the conventional example, in the case of a system in which all sounds are created by rules,
Although it is highly flexible and can synthesize voices of various tones and intonations, it is difficult to synthesize voices with complex voicing mechanisms such as consonants because the synthesis rules are not yet clear. On the other hand, in the case of the method using speech units, although the synthesis quality is extremely high, there is a problem that a large-capacity storage device is required to store the speech units and that only the voice quality of the model speech can be synthesized, and it lacks flexibility. There's a problem. Further, the method in which the modeling synthesis and the segment synthesis are combined has the advantages of the method using the synthesis rule and the method using the segment, but conversely,
There is a problem in that a difference in sound quality occurs at the connection portion, which causes a significant sound quality deterioration factor.

In view of the above problems of the conventional speech synthesizer, the object of the present invention is (1) flexibility in sound quality, and (2) storage capacity can be greatly reduced as compared with the conventional method using speech units. And, (3) an object of the present invention is to provide a speech synthesizing device having a high synthesis quality in which a difference in sound quality due to a connection operation of speech units is not noticeable.

[0011]

A speech synthesis apparatus according to the present invention comprises a natural speech unit storing means for storing a natural speech unit set, and the natural speech units stored in the natural speech unit storing means. Feature parameter storage means for extracting a parameter vector characterizing the acoustic characteristic of the position from a single or a plurality of positions for each segment of the set, and storing the feature parameter vector together with information on the extracted position; From the natural speech unit set stored in the natural speech unit storage means, one of the feature parameter vectors corresponding to a certain natural speech unit is targeted, and the target is changed to the natural speech unit. A speech synthesizing unit for synthesizing a transition of voices to a desired phoneme to be connected, and the natural speech unit and the synthesized speech unit are connected in the vicinity of the position of the target. And a Koemotohen connection means,
This achieves the above object.

The natural speech unit set is a set of units from the start of a consonant of a single syllable to at least a transitional part between consonant vowels or a stationary part of a vowel, and the feature parameter vector is at least a consonant start of the single syllable unit. It may be a vector extracted from a transition part between positions or consonant vowels or a vowel stationary part.

The characteristic parameter vector stored in the characteristic parameter storage means may have at least each formant frequency, each formant band width, and the characteristic of the sound source.

The natural voice storing means cuts out and stores from a consonant start position of a plurality of single syllable original phoneme segments cut out from natural voice to a transition position of a consonant and a succeeding vowel, and the characteristic parameter storing means. , A parameter vector characterizing the acoustic characteristics extracted from the transition position from the consonant to the vowel of each monosyllabic unit stored in the natural voice unit storing means is stored as the first target together with the information of the extracted position. And a parameter vector characterizing the acoustic characteristic extracted from the vowel stationary part of the monosyllabic original phoneme corresponding to the monosyllabic phoneme is stored as the second target together with the information of the extracted position. The synthesizing means interpolates between the first target and the second target and connects the desired sound to the monosyllabic segment from the second target. The speech element connecting means connects the monosyllabic element and the speech element synthesized by the speech synthesizing means in the vicinity of the position of the first target. It may be a configuration.

[0015]

Embodiments of the present invention will be described below with reference to the drawings.

(First Embodiment) FIG. 1 shows the structure of a speech synthesizer according to an embodiment of the present invention. In the figure, a synthesis control unit 1 sends control information such as a timing signal and a fundamental frequency to each unit, and a natural voice element storage unit 2 stores a voice unit set cut out from natural voice after normalizing or compressing it. In this embodiment, the syllable segment set includes a consonant start portion to a vowel stationary portion.
The characteristic parameter storage unit 3 extracts a parameter vector that characterizes acoustic characteristics from a single or a plurality of positions for each unit of the natural speech unit set, and stores the characteristic parameter vector together with information on the extracted position. In the present embodiment, the parameter vector is the formant frequency of the vowel stationary part of each monosyllabic segment, the formant band width, and the average slope of the spectrum as the sound source characteristic. The formant synthesis rule storage unit 4 is a unit that generates an appropriate formant frequency or bandwidth in a certain phoneme environment, and the voiced sound source unit 5 is a unit that simulates vibration of the vocal cords and generates a sound source signal. The formant synthesis unit 6 is a unit that simulates the frequency characteristics of voiced sounds such as vowels with a plurality of resonators to generate synthetic speech. The unit connection unit 7 is a unit that smoothly connects the natural voice unit and the formant synthesized voice.

The operation of the speech synthesizer of this embodiment having the above-mentioned structure will be described below.

The synthesizing operation will be described by taking the word "sora" as an example. First, the synthesis control unit 1 selects the syllable “so” from the natural speech unit storage unit 2, and simultaneously selects the parameter vector corresponding to the stationary part of the vowel “o” of the syllable “so” from the characteristic parameter storage unit 3. To do. Information about the slope of the spectrum in this parameter vector is sent to the voiced sound source unit 5 to generate a sound source signal having a spectrum slope close to the vowel "o" of the syllable "so". Further, information on the formant frequency and the bandwidth in the parameter vector is sent to the formant synthesis unit 6 and stored as target information. Next, the synthesis control unit 1 selects a rule describing the behavior of the vowel “o” under the environment “s-or” from the formant synthesis rule storage unit 4, and this rule is stored in the formant synthesis unit 6. Sent. The formant synthesizing unit 6 synthesizes the voices of the transitions up to the consonant “r” indicated by the synthesis rule while changing the coefficient of the synthesis filter, using the values of the target formant frequency and the bandwidth as the start values. Next, the connection unit 7 connects the natural voice unit “so” and the synthetic voice “o” at the position of the position information of the parameter vector. As a connection method, for example, there is a pitch synchronization window overlapping method as shown in FIG. In this way, the "so" of the natural speech unit and the "o" of the synthesized speech can be connected smoothly, so that the naturalness and clarity of the output speech of the speech synthesis apparatus of this embodiment can be improved. FIG. 3 shows how the formant-type synthesized speech is matched with the spectrum of the natural speech segment at the connection point. Thus, formant-type synthetic speech shows a fairly good match with natural speech when parameter vectors are given appropriately. Hereinafter, similarly, the synthesis control unit 1
Selects "ra", and the formant-type synthesis unit synthesizes "a" to produce the entire synthesized voice. FIG. 4 shows how the synthetic speech and the natural speech segment are connected.

As described above, in the speech synthesizer of the first embodiment of the present invention, the characteristic parameter vector is obtained from the vowel stationary part of the monosyllabic voice, but the characteristic vector at the consonant start position of each syllable is obtained. It is naturally possible that the formant synthesizer 6 has a feature vector and synthesizes a certain vowel targeting the feature vector of the next syllable, and depending on the type of syllable, a large term can be obtained for improving the sound quality.

(Second Embodiment) Next, a speech synthesizer according to a second embodiment of the present invention will be described with reference to FIGS. 1 and 5. In FIG. 1, a natural speech unit storage unit 2 cuts out and stores from a consonant start position of a plurality of monosyllabic original phonemes cut out from a natural voice to a transition position of a consonant and a succeeding vowel, and stores a characteristic parameter. The unit 3 includes a parameter vector characterizing the acoustic characteristic extracted from the transition position from the consonant to the vowel of each monosyllabic unit stored in the natural speech unit storage unit 1 together with the extracted position information and the first target. And the parameter vector characterizing the acoustic characteristic extracted from the vowel stationary part of the monosyllabic original phoneme corresponding to the monosyllabic phoneme is stored as the second target together with the information of the extracted position. FIG. 5 shows this state.

Similar to the first embodiment, the synthesizing operation will be described by taking the word "sora" as an example. First, the synthesis control unit 1 selects the syllable "so" from the natural speech unit storage unit 2, and at the same time, from the characteristic parameter storage unit 3, the consonant / vowel transition part of the syllable "so" and the stationary part of the vowel "o". Two parameter vectors corresponding to are selected as the first target and the second target, respectively. Information about the slope of the spectrum in each of the parameter vector components is sent to the voiced sound source unit 5. Further, information on the formant frequency and the bandwidth in each parameter vector is sent to the formant type speech synthesizer 6 and stored as the first target and second target information, respectively. Next, the synthesis control unit 1 selects a rule describing the behavior of the vowel “o” under the environment “s-or” from the formant synthesis rule storage unit 4, and this rule is stored in the formant synthesis unit 6. Sent. The formant synthesis unit 6 synthesizes the transitions up to the consonant “r” indicated by the synthesis rule with the former formant frequency and the bandwidth value of the first target as starting values and the second target as the passing point of each parameter. Speech synthesis is performed while changing the coefficient of. Next, the connection unit 7 first sets "so" which is a natural speech segment and "o" which is a synthetic speech.
The connection is made at the so-o transition portion, which is the target position.
FIG. 5 summarizes the combining operation and the connecting operation. With this configuration, the length of the syllable stored in the natural voice element storage unit 2 may be a consonant-vowel crossing portion, and the storage capacity of the element set required for synthesis can be significantly reduced. Furthermore, since the vowel stationary part of each phoneme can be accurately reproduced by the second target, it is possible to prevent the sound quality from being deteriorated due to the missing vowel stationary part of the phoneme. Similarly, the synthesis control unit 1 selects "ra", and the formant synthesis unit synthesizes "a" to complete the synthesized speech.

With this configuration, it is possible to give various timbres and intonations flexibly for the vowels by the formant synthesis method, and for the consonants of high quality which cannot be realized by the formant synthesis method by the speech unit method. Can provide audio. Since storage as a voice unit is limited to consonants with a short duration, it can be realized with a small-capacity storage device.

[0023]

As described above, according to the present invention, a vowel signal can be flexibly provided with various tones and intonation by a serial formant synthesis method, and a consonant signal can be formant synthesized by a method using a speech unit. Since it is possible to provide high-quality consonants that cannot be realized by the method, the synthesized voice that combines them can have high quality and can support various voice qualities. In contrast to the conventional method using speech units, this method can be implemented with a small-capacity storage device because it can be stored as a speech unit from the consonant start to the crossover part of a syllable having a short duration. It is possible.

Furthermore, since the acoustic characteristics are matched at the connection points of the elemental pieces, the deterioration due to the difference in the sound quality of the connection portion can be greatly improved.

[Brief description of drawings]

FIG. 1 is a block diagram of a speech synthesizer according to first and second embodiments of the present invention.

FIG. 2 is a diagram showing a state of connection between a natural voice unit “so” and a synthetic vowel “o”.

FIG. 3 is a diagram showing a state of matching between natural speech units and synthetic speech.

FIG. 4 shows the word “so” in a natural speech unit and a synthetic speech unit.
Diagram showing how "ra" is combined

FIG. 5 is a diagram showing a method of connecting a natural speech unit and a synthetic speech unit according to the second embodiment of the present invention.

FIG. 6 is a block diagram of a conventional formant synthesizer.

FIG. 7 is a block diagram of a conventional speech synthesizer using speech units.

FIG. 8 is a block diagram of an apparatus for performing voice synthesis by fusing formants and voice units.

[Explanation of symbols]

 1 synthesis control unit (means) 2 natural speech unit storage unit (means) 3 characteristic parameter storage unit (means) 4 formant synthesis rule storage unit (means) 5 sound source unit (means) 6 formant synthesis unit (means) 7 unit pieces Connection unit (means) 8 Formant synthesizer control coefficient generation unit (means) 9 Formant synthesizer control rule storage unit (means) 10 Formant synthesizer (means) 11 Voiced sound source unit (means) 12 Unvoiced sound source unit (means) 13 Serial formant synthesis section (means) 14 Parallel formant synthesis section (means) 15 Synthesis section (means) 16 Speech element selection section (means) 17 Speech element database storage section (means) 18 Segment connection synthesis section (means) )

Claims (4)

[Claims] 1. A natural speech unit storage means for storing a natural speech unit set, and a single unit for each unit of the natural speech unit set stored in the natural speech unit storage unit. A characteristic vector storage unit that extracts a parameter vector that characterizes the acoustic characteristics of the position from a plurality of positions, and stores the characteristic parameter vector together with the extracted position information, and the natural voice unit storage unit that stores the characteristic parameter vector. From the natural speech unit set, targeting one of the feature parameter vectors corresponding to a certain natural speech unit,
A voice synthesizing unit for synthesizing a transition of a voice from the target to the natural speech unit to a desired phoneme, and a voice connecting the natural speech unit and the synthesized speech unit near the position of the target. A voice synthesizer comprising a segment connecting means. 2. A natural speech segment set is a set of segments from the start of a consonant of a single syllable to at least a transition part between consonant vowels or a vowel steady part, and the feature parameter vector is at least a consonant of the single syllable segment. The speech synthesizer according to claim 1, wherein the vector is extracted from a start position or a transitional part between consonant vowels or a vowel stationary part. 3. The speech synthesizer according to claim 1, wherein the characteristic parameter vector stored in the characteristic parameter storage means has at least each formant frequency, each formant bandwidth, and the characteristic of the sound source. 4. The natural voice storing means cuts out and stores from a consonant start position of a plurality of monosyllabic phoneme segments cut out from a natural voice to a transition position of a consonant and a subsequent vowel, and the characteristic parameter storing means. Stores, as a first target, a parameter vector characterizing an acoustic characteristic extracted from a transition position from a consonant to a vowel of each monosyllabic segment stored in the natural voice segment storage means together with information on the extracted position. And a parameter vector characterizing the acoustic characteristic extracted from the vowel stationary part of the monosyllabic original phoneme corresponding to the monosyllabic phoneme is stored as the second target together with the information of the extracted position, The speech synthesizing means interpolates between the first target and the second target and connects the second target to the monosyllabic segment to obtain a desired phoneme. Said speech unit connecting means connects said monosyllable element and the speech element synthesized by said speech synthesizing means in the vicinity of the position of said first target. 1. The speech synthesizer according to 1.