JP2006227367A - Speech synthesizer - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a speech synthesizer capable of generating a natural synthesized sound by solving conventional problem points regarding voiceless sounding of a speech. <P>SOLUTION: In the speech synthesizer equipped with an speech element dictionary in which speech elements as basic units of a speech are registered, a parameter generation section which generates synthesis parameters of at least a speech element, phoneme duration, and fundamental frequency for a phoneme/rhythm symbol string, and a waveform generation section which generates a synthesized waveform while referring to the speech element dictionary according to the composition parameters from the parameter generation section, the parameter generation section 300 is equipped with a vowel voiceless sounding level decision section 304 which corrects a result of a voiceless sounding decision made according to a normal voiceless sounding rule according to a speaking speed and the kind of a syllable, and the vowel voiceless sounding level decision section 304 makes the voiceless sounding decision based upon at least two or more levels. <P>COPYRIGHT: (C)2006,JPO&NCIPI

Description

  The present invention relates to a text-to-speech synthesizer that takes text of a mixture of vague and kana sentences that are read and written daily and converts them into speech, and more particularly to a process for realizing vowel devoicing.

  Text-to-speech conversion technology inputs vaguely mixed sentences that we read and write every day, converts them into speech, and outputs them. Since there is no restriction on the output vocabulary, it is suitable for recording / playback speech synthesis. As an alternative technology, application in various fields of use can be expected.

  Conventionally, this type of speech synthesizer typically has a processing form as shown in FIG. In FIG. 2, 101 is a text analysis unit, 102 is a parameter generation unit, 103 is a waveform generation unit, 104 is a word dictionary, and 105 is a segment dictionary.

  The text analysis unit 101 inputs a kanji-kana mixed sentence, performs morphological analysis with reference to the word dictionary, determines reading, accent, and intonation, and outputs a phonetic symbol with prosodic symbols (intermediate language).

  The parameter generation unit 102 sets a pitch frequency pattern, phoneme duration, and the like, and the waveform generation unit 103 performs speech synthesis processing. The waveform generation unit 103 selects a speech synthesis unit to be used for speech synthesis from the phonetic dictionary 105 from a phonetic symbol string given in an intermediate language, and combines / transforms the speech synthesis unit according to the parameters determined by the parameter generation unit. Perform synthesis processing. As the speech synthesis unit, phonemes, syllables (CV), VCV, CVC (C: consonant, V: vowel), a unit in which the phoneme chain is expanded, or the like is used.

  In the above configuration, when a vague character / kana mixed sentence (hereinafter referred to as text) that is read and written daily is input, the text analysis unit 101 generates a phoneme / prosodic symbol string from the character information. A phoneme / prosodic symbol string is a text string describing the reading, accent, intonation, etc. of an input sentence (hereinafter referred to as an intermediate language). The word dictionary 104 is a pronunciation dictionary in which word readings, accents, and the like are registered, and the text analysis unit 101 generates an intermediate language while referring to the word dictionary.

  The intermediate language generated by the text analysis unit 101 is a parameter generation unit 102, which includes a speech segment (sound type), phoneme duration (sound length), fundamental frequency (voice pitch, hereinafter referred to as pitch), and the like. Are determined and sent to the waveform generation unit 103. A speech unit is a basic unit of speech for connecting and creating a synthesized waveform, and there are various types depending on the type of sound.

  Various parameters generated by the parameter generation unit 102 are generated with reference to a segment dictionary 105 composed of a ROM or the like that stores speech units and the like in the waveform generation unit 103, and a synthesized waveform is generated through a speaker. Is output. The above is the flow of the text-to-speech conversion process.

  Next, processing in the parameter generation unit 102 will be described in detail with reference to FIG. FIG. 3 is a block diagram showing the configuration of the parameter generation unit 102 of the conventional speech synthesizer. In FIG. 3, the parameter generation unit 102 includes an intermediate language analysis unit 201, a pitch pattern generation unit 202, a vowel devoicing determination unit 203, a phoneme power determination unit 204, a phoneme duration calculation unit 205, and a duration correction unit 206. The

  The intermediate language input to the parameter generation unit 102 is a phonological character string including an accent position, a pose position, and the like. From this, a temporal change in pitch (hereinafter referred to as a pitch pattern), and a duration of each phonology. Parameters necessary for generating a waveform (hereinafter, referred to as a waveform generation parameter) such as a phoneme duration (hereinafter, referred to as a phoneme duration) and a voice power are determined.

  The input intermediate language is analyzed for a character string by the intermediate language analysis unit 201, a word boundary is determined from a word delimiter written on the intermediate language, and a mora position of an accent kernel is obtained from the accent symbol.

  An accent nucleus is a position where the pitch falls within a word. A word having an accent nucleus in the first mora is called a 1-type accent, and a word having an accent nucleus in the n-mora is called an n-type accent. This is referred to as a undulating accent word. Conversely, a word that does not have an accent nucleus (for example, “newspaper” or “computer”) is referred to as a 0-type accent or a flat accent word.

  The pitch pattern generation unit 202 generates a pitch pattern based on phrase symbols / accent symbols in the intermediate language, the number of mora of words, the number of mora of phrases, and the like.

  The vowel devoicing determination unit 203 performs vowel devoicing determination from a phonological symbol or accent symbol in an intermediate language, and sends the result to the phonological power determination unit 204 and the phonological duration calculation unit 205. The vowel devoicing will be described later.

  The phoneme duration calculation unit 205 calculates the duration of each phoneme from the phoneme string and sends it to the duration correction unit 206. When the user designates the utterance speed level, the phoneme duration calculated in 205 is linearly expanded / contracted by the duration correction unit 206 according to the designated level. The phoneme duration extended and contracted according to the utterance speed level by the duration correction unit 206 is sent to a waveform generation unit (not shown).

  The phoneme power determination unit 204 calculates the amplitude value of the waveform and sends it to the waveform generation unit 103 (see FIG. 2).

  These waveform generation parameters described above are sent to the waveform generation unit 103 to generate a composite waveform.

  Next, vowel devoicing will be described in detail. When a person speaks, the air pushed out from the lungs is used as a sound source by opening and closing movements of the vocal cords, and the vocal tract resonance characteristics are changed by moving the chin, tongue, lips, etc. to express various phonemes. The pitch described above corresponds to the vibration period of the vocal cords, and this temporal change is an expression of accent and intonation.

  In addition to this vocal cord vibration, the tongue creates a narrow space in the part of the vocal tract, and when the airflow passes through it, turbulence is generated and a noisy sound is generated. There is also a popping sound that blocks the road and stops the temporary air flow, then opens at once and generates an impulse sound.

  Accompanied by vocal cord vibration such as vowels, plosives / b /, / d /, / g /, friction sounds / j /, / z /, nasal consonants / stream sounds / m /, / n /, / r / The phonemes are collectively referred to as voiced sounds, plosive sounds / p /, / t /, / k /, friction sounds / s /, / h /, / f /, and the phonemes not accompanied by vocal cord vibration are called unvoiced sounds. In particular, focusing on consonants, consonants accompanied by vocal cord vibrations are called voiced consonants, and consonants not accompanied are called unvoiced consonants. For voiced sounds, a periodic waveform due to vocal cord vibration is generated, and for unvoiced sounds, a noisy waveform is generated.

  When voiceless consonants (/ p /, / t /, / k /, / s /, / f /, / h / etc.) Are interleaved with / i /, / u /, leaving only the mouth There is a phenomenon in which the vocal cords are not vibrated and are pronounced only by breathing. This is called vowel devoicing.

  FIG. 4 is a diagram showing an actual sound waveform of “covered”. As shown in FIG. 4, the vowel of “shi” “covered” does not appear as a periodic waveform, but has a waveform that is fused with the silent friction sound “sh”.

  Also in the text-to-speech conversion system, the vowel devoicing expression is important for improving the auditory quality, and the vowel devoicing determination unit 203 in FIG. 3 performs this determination. The vowels determined to be unvoiced here are subjected to special processing according to the phoneme power determination unit 204 and the phoneme duration calculation unit 205. Unlike normal vowels, the unvoiced vowels are sent to the waveform generation unit 103 as phoneme power 0 and phoneme duration 0.

The devoicing determination is performed according to the following rules, for example.
(1) Vowels / i /, / u / sandwiched between unvoiced consonants (including silence) are devoiced (2) However, if there is an accent nucleus, they are not devoiced (3) However, the previous vowel is already unvoiced (4) However, the end of the question is not silent

  However, these are rules derived from general trends, and devoicing does not always occur according to the rules described above in actual speech. There are also differences among individuals. The above rule is extremely appropriate for normal speech rate utterance, but generally speaking, when the speech rate is slow, devoicing is difficult to occur. FIG. 5 shows a voice waveform of / kisoku / when the normal speed and the speech speed are slow. The / kisoku / underlined part matches the devoicing condition, and devoicing occurs at normal speed, whereas vowels clearly appear in any slow utterance.

As one countermeasure against such a phenomenon, Patent Document 1 proposes a speech synthesizer including a vowel devoicing determination unit that changes a determination criterion as to whether or not to perform a vowel devoicing process according to the utterance speed. . In this document, a threshold value is provided for the voice rate, and at the voice rate that is equal to or lower than the threshold value, the voice-off process is not performed even if the voice-off requirement according to the rule is satisfied.
JP-A-11-116263

  While it is certain that devoicing tends to be difficult to occur at lower utterance speeds, this is not necessarily the case for all syllables. In syllables such as “It was: ~ deshita”, even if the utterance speed is slow, the syllable is often made silent. As a tendency, the frictional sounds / s /, / f /, / h /, etc., in which the consonant part can be continuously generated, are likely to be silent even at low speech speeds.

  If the speech rate is slow, even if such syllables are included and devoicing is uniformly suppressed, the synthesized sound becomes distorted and unnatural. In addition, even for a syllable in which devoicing is difficult to occur when the utterance speed is slowed, the utterance property does not change to extremes, such as complete devoicing / no devoicing at the threshold. Originally, the phenomenon of devoicing occurs because the vowel interval between consonants with a narrow aperture area maintains a narrow aperture area, and the expiratory airflow required to vibrate the vocal cords as a voiced sound cannot be secured. It is. Due to the structure of the human vocal organs and the vocalization mechanism, the change in properties changes in stages. However, the conventional uniform and uniform devoicing determination process cannot express the change in the stepwise nature of the devoicing that accompanies the step change in the speech rate, and must be an unnatural synthesized sound. .

  Accordingly, an object of the present invention is to provide a speech synthesizer that can solve a conventional problem related to voice unvoiced and obtain a natural synthesized sound.

  The speech synthesizer according to the present invention includes a unit dictionary in which speech units as basic units of speech are registered, and at least speech units, phoneme durations, and fundamental frequency synthesis parameters for phoneme / prosodic symbol strings. In a speech synthesizer comprising: a parameter generation unit to generate; and a waveform generation unit that generates a composite waveform while referring to a segment dictionary with respect to a synthesis parameter from the parameter generation unit, the parameter generation unit includes a normal devoicing rule A determination correction means for correcting the determination result according to the utterance speed and the type of syllable, and the determination correction means at least according to the degree of devoicing. The determination of devoicing is performed at two or more levels.

  In the speech synthesizer according to the present invention, in the determination of devoicing, at least two devoicing levels are provided, and the devoicing level is assigned step by step according to the utterance speed level. Natural devoicing can be realized at any utterance speed without abruptly changing the degree of devoicing before and after the boundary point. In addition, the devoicing level setting according to the utterance speed level is configured to give a setting rule suitable for each syllable type, so that natural devoicing that reflects the characteristics of consonants with respect to changes in utterance speed is possible. It is feasible, and natural and smooth synthesized speech can be obtained.

  Embodiments of the present invention will be described below with reference to the drawings. The drawings are only schematically shown so that the present invention can be understood.

  FIG. 1 is a block diagram illustrating a configuration of a parameter generation unit of the speech synthesizer according to the embodiment. The characteristic part of the present invention resides in the devoicing determination means and the method for realizing the same. In this embodiment, the text analysis unit 101, the word dictionary 104, the waveform generation unit 103, and the segment dictionary 105 shown in FIG.

  In FIG. 1, a parameter generation unit 300 includes an intermediate language analysis unit 301, a pitch pattern generation unit 302, a vowel devoicing primary determination unit 303, a vowel devoicing level determination unit 304, a phonological power determination unit 305, and a phonological duration calculation unit 306. It has.

  The input to the parameter generation unit 300 is an intermediate language to which prosodic symbols are added and a speech rate parameter designated by the user as in the conventional case. Also, voice quality parameters such as intonation indicating the voice pitch and intonation level may be designated from the outside depending on the user's preference and usage pattern.

  The intermediate language to be synthesized is input to the intermediate language analysis unit 301, and the utterance speed parameter specified by the user is input to the vowel devoicing level determination unit 304. Of the output data from the intermediate language analysis unit 301, parameters such as a phrase delimiter, a word delimiter, and an accent symbol are input to the pitch pattern generation unit 302, such as a phoneme symbol string, a word delimiter, and an accent symbol. These parameters are input to the phoneme power determination unit 305 and the phoneme duration calculation unit 306, and parameters such as phoneme symbol strings and accent symbols are input to the vowel devoicing primary determination unit 303.

  The pitch pattern generation unit 302 calculates data such as the occurrence time and size of the phrase command and the start time / end time and size of the accent command from the input parameters, and generates a pitch pattern. The generated pitch pattern is input to the waveform generation unit 103 (see FIG. 2). Since the pitch pattern generation process is not directly related to the present invention, the description thereof is omitted.

  The vowel devoicing primary determination unit 303 performs vowel devoicing determination based only on input text such as face and accent, and outputs the determination result to the vowel devoicing level determination unit 304.

  The vowel devoicing level determination unit 304 performs final devoicing determination from the primary vowel devoicing determination result and the utterance speed level specified by the user, and the final determination result is used as the phoneme power determination unit 305 and phonological continuation. The result is output to the time calculation unit 306.

  The phoneme power determination unit 305 calculates the amplitude shape of each phoneme from the vowel devoicing determination result and the phoneme symbol string input from the intermediate language analysis unit 301, and outputs it to the waveform generation unit 103.

  The phoneme duration calculation unit 306 calculates the duration of each phoneme from the vowel devoicing determination result and the phoneme symbol string input from the intermediate language analysis unit 301, and outputs the result to the waveform generation unit 103.

  The operation of the speech synthesizer configured as described above will be described below. Since the difference from the prior art is the processing in the parameter generation unit 300, other processing is omitted.

  First, the user designates an utterance speed level in advance. The utterance speed is usually given as a parameter in the form of how many mora is uttered per minute, and for convenience of use, the level value is quantized to about 5 to 10 levels. Depending on this level, processing such as extension of phoneme duration is performed. In addition, parameters for voice quality control such as voice pitch and intonation can be specified. Unless otherwise specified by the user, a predetermined value (default value) is set as the specified value.

  The utterance speed control parameter designated by the user is sent to the vowel devoicing level determination unit 304 and the duration calculation unit 306. The other input intermediate language is sent to the intermediate language analysis unit 301, and the input language string is analyzed by the intermediate language analysis unit 301. The analysis unit here is assumed to be one sentence unit.

  For example, information on the number of phrase commands and the number of mora of each phrase command, the number of accent commands and the number of mora / accent type of each accent command, etc. Is sent to the pitch pattern generation unit 302.

  The pitch pattern generation unit 302 calculates the size of each phrase command / accent command, rise / fall position, and the like from the input parameters, and generates a pitch pattern using a response function defined in advance. The calculated pitch pattern is sent to the waveform generation unit 103 (see FIG. 2).

  Accent symbols / phonological character strings and the like are sent to the vowel devoicing primary determination unit 303, and vowel devoicing determination is performed. In this primary determination, a determination is made only from the arrangement of character strings and the presence or absence of an accent nucleus, and a provisional determination result is sent to the vowel devoicing level determination unit 304.

  The vowel devoicing level determination unit 304 is also input with the utterance speed level designated by the user, and the final determination process is performed together with the primary determination result. In this determination process, it is compared whether or not the utterance speed exceeds a specified value, and the vowel devoicing process is not performed only when it is determined that the utterance speed is slow from the comparison result. After this processing, the determination result is sent to the phoneme power determination unit 305 and the phoneme duration calculation unit 306 in order to make a final determination of vowel devoicing. The processing of the vowel devoicing temporary determination unit 303 and the vowel devoicing level determination unit 304 will be described in detail later.

  The phoneme power determination unit 305 calculates a waveform amplitude value of each phoneme or syllable from parameters such as a phoneme character string input from the intermediate language analysis unit 301 and outputs the waveform amplitude value to the waveform generation unit 103.

  The phoneme duration calculation unit 306 calculates the duration of each phoneme or syllable from parameters such as a phoneme character string input from the intermediate language analysis unit 301.

  Next, the vowel devoicing determination process in the vowel devoicing temporary determination unit 303 and the vowel devoicing level determination unit 304 will be described in detail with reference to a flowchart. FIG. 6 is a flowchart showing a specific example of the vowel devoicing determination process. In the figure, ST indicates each processing step of the flow.

  Up to ST7 corresponds to the processing of the vowel devoicing temporary determination unit 303, and the subsequent processing corresponds to the processing of the vowel devoicing level determination unit 304. The processing of the vowel devoicing primary determination unit up to ST7 is the same as the devoicing determination in the prior art, and in accordance with general devoicing rules, the primary vowel environment is determined based on the phonological environment before and after the target vowel, the presence or absence of an accent nucleus, etc. Make a devoicing decision.

  First, in step ST1, a syllable pointer i for searching the input intermediate language in syllable units is initialized to 0. In step ST2, the type of vowel of the i-th syllable (a, i, u, e, o) is initialized. ), The consonant type of the syllable (unvoiced consonant or unvoiced / voiced consonant), and the consonant type of the subsequent syllable are set to V1, C1, and C2, respectively.

  The processes in steps ST3 and ST4 are determination processes for the precondition that the vowel is unvoiced, and the processes in steps ST5 to ST7 are made unvoiced even if the precondition for vowel devoicing in steps ST3 and ST4 is satisfied. This is a determination process when no occurrence occurs.

  In step ST3, it is determined whether the vowel V1 of the syllable is “i” or “u”. If the vowel V1 of the syllable is “i” or “u”, the process proceeds to step ST4, and if not, it is determined that the syllable vowel V1 is not an object of devoicing, and the process proceeds to step ST15.

  In step ST4, it is determined whether the consonant C1 of the syllable is an unvoiced consonant and whether the subsequent syllable consonant C2 is an unvoiced consonant or a sentence end / pause. If both are satisfied, the process proceeds to step ST5 to determine whether or not an accent nucleus exists in the corresponding syllable. Japanese expresses the position of the accent by changing the pitch. Therefore, in a syllable with an accent nucleus indicating a position of an accent with a transition from a high pitch to a low pitch, devoicing without a pitch structure hardly occurs.

  In step ST5, it is determined whether or not the syllable is an accent nucleus. If there is no accent nucleus in the corresponding syllable (if it is not an accent nucleus), it is determined in step ST6 whether or not the previous syllable is devoiced. Devoicing has the property that it is difficult to occur continuously. If the previous syllable is already devoiced, the subsequent syllable usually does not devoicing.

  In step ST7, it is determined whether or not the corresponding syllable is the question sentence end. At the end of the question, there is no devoicing because the pitch rises rapidly. For example, in the case of “to do” and “to do?”, The syllable at the end of the latter question sentence clearly includes the intention of emphasis, so devoicing does not occur.

  Those that satisfy the conditions of steps ST3 to ST7 and reach ST8 are candidates for devoicing. After ST8, the devoicing level is determined according to the utterance speed and syllable type, which is a feature of the present invention.

  First, syllable types including vowels that are candidates for devoicing are classified. Although it is possible to classify the syllables and make them regular, it is possible to classify the syllables into three types in this embodiment for easy understanding of the processing structure, and individually set the judgment level based on the utterance speed. is doing.

  As described above, a syllable in which the consonant part C1 is a frictional sound is often pronounced unvoiced even when the utterance speed is slow. When the consonant part C1 is a plosive such as / p /, / t /, / k /, it becomes difficult to devoice if the utterance speed is slow. With respect to the consonant sound, the frictional sound / s /, / h /, / f / in which the opening area at the time of utterance is maintained even when the duration is extended is highly likely to be devoiced even if the utterance speed is low. Among these, at / h / and / f / having a relatively large opening area, devoicing is less likely to occur than at / s /. The plosive sound has an extremely narrow opening area at the time of rupture, but then becomes wider. Since it is difficult to utter while keeping the aperture area small, if the utterance speed becomes slow, the aperture area increases with time and vocal cord vibration starts.

  Reflecting such properties, in this embodiment, there are three types of C1 = / s / (ST8, ST10), C1 = / h /, / f / (ST9, ST11), and other (ST12). It classifies into syllable groups, and devoicing determination levels TH1 and TH2 corresponding to the utterance speed are set for each classification. Here, TH1 and TH2 represent utterance speed determination threshold values in units of mora / min. In the subsequent steps, TH1 is used for boundary determination of (no devoicing) ← → (quasi-unvoiced), and TH2 is , (Quasi-voiceless) ← → (voiceless) boundary determination.

  Specifically, the threshold is set to TH1 = Na1, TH2 = Na2 in the case of the first classification (C1 = / s /), and in the case of the second classification (C1 = / h /, / f /), Set TH1 = Nb1, TH2 = Nb2, and in the case of the third classification (other than the above), set TH1 = Nc1, TH2 = Nc2. Each of these threshold values is set to a relationship of Na1 ≦ Nb1 ≦ Nc1, Na2 ≦ Nb2 ≦ Nc2, Na1 ≦ Na2, Nb1 ≦ Nb2, and Nc1 ≦ Nc2, as is apparent from the properties of each syllable.

  In steps from ST13, the devoicing level is determined using the set speech rate determination thresholds TH1 and TH2. The devoicing level is a level in which a plurality of levels are provided according to the degree of devoicing, instead of separating them into two extremes of devoicing / no devoicing with respect to the utterance speed. In the present embodiment, an intermediate level “quasi-devoiced” is provided between “not devoiced” and “completely devoiced” for a total of three levels.

  Unvoiced vowels are realized as phoneme power 0 and phoneme duration 0, unlike normal vowels. “Quasi-voicing” is not to eliminate vowels as in the case of devoicing, but to leave vowels to some extent while their power and duration are smaller than normal vowels. In the present embodiment, a flag variable uvflag [i] representing a vowel devoicing level is provided for each syllable, and by comparing the utterance speed determination thresholds TH1 and TH2 with the utterance speed Tlevel, (no devoicing: uvflag [i] = 0), (quasi-unvoiced: uvflag [i] = 2), and (voicing: uvflag [i] = 1) are finalized.

  In step ST18, the syllable counter i is incremented by 1. In step ST19, it is determined whether the syllable counter i is greater than or equal to the total number of mora sum_mora (i ≧ sum_mora). If i <sum_mora, it is determined that the processing is not completed. The process then returns to step ST2 and the next syllable process is repeated in the same manner.

  The above-described processing is finished for all input text syllables, that is, when the syllable counter i exceeds the total number of mora sum_mora in step ST19.

  As described above, according to this embodiment, in the determination of devoicing, a plurality of devoicing levels are provided, and the devoicing level is assigned step by step according to the utterance speed level. Natural devoicing can be achieved at any utterance speed without abruptly changing the degree of devoicing before and after the speed boundary point. In addition, the devoicing level setting according to the utterance speed level is configured to give a setting rule suitable for each syllable type, so that natural devoicing that reflects the characteristics of consonants with respect to changes in utterance speed is possible. It is feasible, and natural and smooth synthesized speech can be obtained.

It is a block diagram which shows the structure of the parameter production | generation part of the speech synthesizer which concerns on embodiment. It is a functional block diagram of a text voice conversion process. It is a functional block diagram which shows the structure of the parameter production | generation part in a prior art. It is a figure which shows the actual generation | occurrence | production waveform of "covered". It is a figure which shows the actual generation | occurrence | production waveform of "rule". It is a flowchart of the vowel devoicing determination according to the embodiment.

Explanation of symbols

300 Parameter generation unit 301 Intermediate language analysis unit 302 Pitch pattern generation unit 303 Vowel devoicing primary determination unit 304 Vowel devoicing level determination unit 305 Vowel power determination unit 306 Phoneme duration calculation unit

Claims (4)

A speech unit dictionary in which speech units that are basic units of speech are registered; a parameter generation unit that generates at least speech units, phoneme durations, and fundamental frequency synthesis parameters for phoneme / prosodic symbol strings; and the parameters In a speech synthesizer comprising a waveform generation unit that generates a synthesized waveform while referring to the segment dictionary for synthesis parameters from a generation unit,
The parameter generation unit includes a vowel devoicing determination unit that determines whether or not to perform a vowel devoicing process, and the vowel devoicing determination unit determines at least two devoicing levels according to the degree of devoicing. A speech synthesizer characterized by the above.   The speech synthesizer according to claim 1, wherein the devoicing level has at least one intermediate level between “not devoicing” and “unvoiced”.   The vowel devoicing determining means includes primary determining means for performing vowel devoicing determination based on the arrangement of character strings and the presence or absence of accent nuclei, according to the result of the primary determining means and the utterance speed designated by the user, The speech synthesizer according to claim 1, further comprising a devoicing level determination criterion for selecting the devoicing level.   4. The speech synthesizer according to claim 3, wherein the devoicing level determination criterion is determined according to the type of syllable to be devoiced.

Images