JP5915264B2 - Speech synthesizer - Google Patents

Description

  The present invention relates to a technique for synthesizing speech (speech sound or singing sound) using speech segments.

  Conventionally, segment connection type speech synthesis has been proposed in which sound designated as a target for speech synthesis (hereinafter referred to as “synthesis target sound”) is generated by connecting a plurality of speech waveforms collected in advance. For example, in the technique of Patent Document 1, a speech waveform (segment data) collected in advance for each speech unit is stored in a storage device, and each speech waveform corresponding to a pronunciation character (for example, lyrics) of a synthesis target sound is stored. A sound signal of the synthesis target sound is generated by selecting the devices sequentially and connecting them to each other.

JP 2007-240564 A

  In the technique of Patent Document 1, when a time length longer than the speech waveform stored in the storage device is designated as the duration of the synthesis target sound, a speech signal is generated by repeating (looping) the speech waveform. . Therefore, there is a problem that a regular characteristic change (for example, a change in amplitude or period) in which the time length of the voice waveform is one period occurs, and the sound quality perceived by the listener is deteriorated. If the time length of each voice waveform is sufficiently secured so that it is not necessary to repeat the voice waveform, the above problems can be solved, but a huge storage capacity is required to store the voice waveform over a long period of time. There's a problem. In view of the above circumstances, an object of the present invention is to prevent deterioration in sound quality due to repeated speech waveforms while reducing the storage capacity necessary for speech synthesis.

  Means employed by the present invention to solve the above problems will be described. In order to facilitate the understanding of the present invention, in the following description, the correspondence between the elements of the present invention and the elements of the embodiments described later will be indicated in parentheses, but the scope of the present invention will be exemplified in the embodiments. It is not intended to be limited.

  The speech synthesizer of the present invention stores a plurality of unit waveforms (for example, unit waveform u [m]) extracted from different positions on the time axis among voiced speech waveforms (for example, speech waveform Vb). Means (for example, storage device 12), and waveform generation means (for example, speech synthesizer 28) that generates a synthesized waveform (for example, synthesized waveform C [n]) by arranging each of a plurality of unit waveforms on the time axis. To do. In the above configuration, a plurality of unit waveforms extracted from different positions on the time axis in the speech waveform are arranged on the time axis to generate a composite waveform, so that the speech waveform is repeated. Compared with the above configuration, it is possible to prevent deterioration in sound quality due to waveform repetition. Further, since each unit waveform extracted from the speech waveform is stored in the waveform storage means, there is also an advantage that a necessary storage capacity is reduced as compared with a configuration in which all sections of the speech waveform are stored.

  In a preferred aspect of the present invention, the waveform generation means includes a first unit waveform (for example, a first unit waveform Ua [] selected from a plurality of unit waveforms for each of a plurality of processing periods (for example, the processing period R [n]). n]), a first waveform series (for example, the first waveform series Sa [n]) in which the intensity increases with time within the processing period, and the first unit waveform among the plurality of unit waveforms. A second waveform series (for example, the second waveform series Sb, for example) in which a plurality of second unit waveforms (for example, the second unit waveform Ub [n]) different from the above are arranged so that the intensity decreases with time within the processing period. [n]) is added to generate a composite waveform, for example, a composite waveform C [n]). In the above aspect, a composite waveform is generated by adding (cross-fade) the first waveform series in which the first unit waveforms are arranged and the second waveform series in which the second unit waveforms are arranged. The effect that it is difficult to perceive the periodicity of the characteristic change in the segment waveform (for example, the segment waveform Q) in which is arranged is particularly remarkable. The first unit waveform and the second unit waveform do not necessarily have to be different for all the processing periods on the time axis, and there is a configuration in which a processing period in which the first unit waveform and the second unit waveform are common exists. It is included in the scope of the present invention. That is, the “plurality of processing periods” in the above aspect means each processing time in which the first unit waveform and the second unit waveform are different among all the processing periods on the time axis.

  In the specific example of the aspect in which the composite waveform is generated by adding the first waveform series and the second waveform series, the first unit waveform of one processing period of the plurality of processing periods and one of the plurality of processing periods The second unit waveform in the processing period immediately after the processing period is a common unit waveform. According to the above aspect, since a common unit waveform is selected as the second unit waveform in successive processing periods, a comparison is made with a configuration in which both the first unit waveform and the second unit waveform are changed for each processing period. Thus, it is possible to suppress a regular characteristic change for each processing period in the segment waveform.

  In a specific example of the aspect in which the composite waveform is generated by adding the first waveform series and the second waveform series, the waveform generation unit randomly selects the first unit waveform from the plurality of unit waveforms for each processing period. In the above aspect, since the first unit waveform is randomly selected for each processing period, it is possible to suppress a periodic characteristic change for each processing period in the segment waveform.

  In the specific example of the aspect in which the composite waveform is generated by adding the first waveform series and the second waveform series, the waveform generation means includes the time length of one processing period and the time of another processing period. Different from the length. In the above aspect, since the time lengths of the respective processing periods can be different, it is possible to suppress a periodic characteristic change in the fragment waveform as compared with the configuration in which the time lengths of all the processing periods are common. . The above effects become particularly prominent by setting each time length of the plurality of processing periods at random.

  In a preferred aspect of the present invention, each of the plurality of unit waveforms corresponds to one cycle of the speech waveform. In the above aspect, since the unit waveform corresponding to one cycle of the speech waveform is used for generating the synthesized waveform, the effect of achieving both the reduction of the storage capacity and the suppression of the periodicity of the characteristic change is particularly remarkable. Become.

  In a preferred aspect of the present invention, the peak-to-peak value of the intensity (amplitude) of the unit waveform is common to the plurality of unit waveforms. In the above aspect, since the peak-to-peak values of the unit waveforms are common, fluctuations in the amplitude of the composite waveform generated from the plurality of unit waveforms are suppressed. Therefore, there is an advantage that a natural voice whose amplitude is kept constant can be generated.

  In a preferred aspect of the present invention, the unit waveform has a common time length for a plurality of unit waveforms. In the above aspect, since the time length of each unit waveform is common, the fluctuation | variation of the period of the synthetic | combination waveform produced | generated from a several unit waveform is suppressed. Therefore, there is an advantage that a natural voice whose period is kept constant can be generated.

  In a preferred aspect of the present invention, the phases of the plurality of unit waveforms are adjusted such that the cross-correlation function between the unit waveforms is maximized. In the above aspect, each phase is adjusted so that the cross-correlation function between the unit waveforms is maximized. Therefore, the cancellation of the first unit waveform and the second unit waveform is suppressed, and an audibly natural element is obtained. There is an advantage that a half waveform can be generated.

  The speech synthesizer according to each aspect described above is realized by hardware (electronic circuit) such as DSP (Digital Signal Processor) dedicated to speech synthesis, and general-purpose arithmetic processing such as CPU (Central Processing Unit). It is also realized by cooperation between the device and the program. The program of the present invention (for example, the program PGM1) stores a plurality of unit waveforms in a computer including waveform storage means for storing a plurality of unit waveforms extracted from different positions on the time axis in a voiced sound waveform. A waveform generation process for generating a composite waveform by arranging each of them on the time axis is executed. In the waveform generation processing, for example, for each of a plurality of processing periods, a first waveform in which a plurality of first unit waveforms selected from a plurality of unit waveforms are arranged so that the intensity increases with time within the processing period. A combination of a series and a second waveform series in which a plurality of second unit waveforms different from the first unit waveform among the plurality of unit waveforms are arranged so that the intensity decreases with time within the processing period. This is a process for generating a waveform. According to the above program, the same operation and effect as the speech synthesizer of the present invention are realized. The program of the present invention is provided to a user in a form stored in a computer-readable recording medium and installed in the computer, or provided from a server device in a form of distribution via a communication network and installed in the computer. Is done.

  The present invention can also be implemented as a speech processing device that generates a plurality of unit waveforms used in the speech synthesizer according to each aspect described above. The speech processing apparatus according to the present invention includes a waveform extraction unit (for example, a waveform extraction unit 62) that extracts a plurality of unit waveforms from different positions on the time axis in a voiced sound waveform, and a plurality of waveform extraction units. Waveform correction means (for example, a waveform correction unit 64) for correcting the unit waveform so that the acoustic characteristics of each unit waveform approach each other.

  In a preferred aspect of the present invention, the waveform correction means includes period correction means (for example, a period correction unit 74) that adjusts the time length of each of the plurality of unit waveforms to a common predetermined length. In the above aspect, since the cycle of each unit waveform is adjusted to a common predetermined length, fluctuations in the cycle of the combined waveform are suppressed. Therefore, there is an advantage that a natural voice whose period is kept constant can be generated.

  In a preferred aspect of the present invention, the period correction means, for each of a plurality of different candidate lengths, a distortion index value indicating the degree of distortion of each unit waveform when each unit waveform is expanded or contracted to the candidate length on the time axis And an index calculation means (for example, an index calculation unit 742) for calculating the time, and a candidate length that minimizes the degree of distortion indicated by the distortion index value among a plurality of candidate lengths is selected as a predetermined length, and each time of each of the plurality of unit waveforms is selected. Correction processing means (correction processing unit 744) for adjusting the length to a predetermined length. In the above aspect, since the predetermined length after correction is selected so that distortion of each unit waveform is suppressed, there is an advantage that a unit waveform that accurately reflects the acoustic characteristics of the speech waveform can be generated.

  The speech processing apparatus according to a preferred aspect of the present invention corrects the amplitude of each unit waveform so that the amplitude of the unit waveform increases as the predetermined length increases with respect to the time length of the unit waveform extracted by the waveform extraction unit. Distortion correcting means (for example, a distortion correcting unit 78) is provided. In the above aspect, since the fluctuation of the amplitude of the unit waveform caused by the correction by the period correcting unit is corrected, the effect that the unit waveform that faithfully reflects the acoustic characteristics of the speech waveform can be generated is particularly remarkable.

  The audio processing apparatus described above is realized by hardware (electronic circuit) such as a DSP (Digital Signal Processor) dedicated to audio processing, and a general-purpose arithmetic processing apparatus such as a CPU (Central Processing Unit) and a program. It is also realized in collaboration with. The program of the present invention (for example, program PGM2) is a program for generating a plurality of unit waveforms used for speech synthesis, and a plurality of unit waveforms from different positions on the time axis among voiced sound waveforms. And a waveform correction process for correcting the plurality of unit waveforms extracted by the waveform extraction process so that the acoustic characteristics of each unit waveform approach each other. According to the above program, the same operation and effect as the speech processing apparatus of the present invention are realized. The program of the present invention is provided in a form stored in a computer-readable recording medium and installed in the computer, or is provided in a form distributed via a communication network and installed in the computer.

1 is a block diagram of a speech synthesizer according to a first embodiment of the present invention. It is explanatory drawing of the segment data of a variation unit and a stationary unit. It is a time-series schematic diagram of an edit image and a speech unit. It is a flowchart of operation | movement of a speech synthesizer. It is a flowchart of the waveform generation process which produces | generates the segment waveform of a stationary unit. It is explanatory drawing of a waveform production | generation process. It is a block diagram of the speech processing unit concerning a 2nd embodiment. It is explanatory drawing of operation | movement of an amplitude correction part. It is explanatory drawing of operation | movement of a period correction | amendment part. It is explanatory drawing of operation | movement of a phase correction part. It is a block diagram of the period correction | amendment part in 3rd Embodiment. It is explanatory drawing of operation | movement of the period correction | amendment part in 3rd Embodiment. It is a block diagram of the waveform correction | amendment part in 4th Embodiment. It is a block diagram of the speech synthesizer in the fifth embodiment. It is explanatory drawing of the waveform production | generation process which produces | generates the segment waveform of a stationary segment in a modification.

<A: First Embodiment>
FIG. 1 is a block diagram of a speech synthesizer 100 according to the first embodiment of the present invention. The speech synthesizer 100 is a speech processing device that generates a synthesis target sound such as a singing sound or a speech sound by unit connection type speech synthesis. As shown in FIG. This is realized by a computer system including the device 14, the display device 16, and the sound emitting device 18.

  The arithmetic processing unit 10 (CPU) has a plurality of functions (a display control unit 22, an information generation unit 24, an element unit) for generating the audio signal SOUT of the synthesis target sound by executing the program PGM1 stored in the storage device 12. The selection unit 26 and the speech synthesis unit 28) are realized. The audio signal SOUT is an acoustic signal representing the waveform of the synthesis target sound. A configuration in which each function of the arithmetic processing device 10 is distributed over a plurality of integrated circuits, or a configuration in which a dedicated electronic circuit (DSP) realizes each function may be employed.

  The storage device 12 stores a program PGM1 executed by the arithmetic processing device 10 and various pieces of information (element group G, composite information Z) used by the arithmetic processing device 10. A known recording medium such as a semiconductor recording medium or a magnetic recording medium or a combination of a plurality of types of recording media is employed as the storage device 12.

  The segment group G is a set (speech synthesis library) of a plurality of segment data W. Each unit data W is a sample series showing a waveform on the time axis of a speech unit, and is used as a material for speech synthesis. The phoneme unit is a phoneme (corresponding to a minimum unit of linguistic meaning) or a phoneme chain (for example, a diphone or a triphone) in which a plurality of phonemes are connected. In the following, for convenience, silence is described as one phoneme (symbol #).

  Speech segments are classified into stationary segments whose acoustic characteristics are stationary and variable segments whose acoustic characteristics vary with time. A typical example of a stationary element is a voice element of a voiced sound (voiced vowel or voiced consonant) composed of one phoneme, and a typical example of a variable element is an unvoiced sound composed of one phoneme ( An unvoiced consonant) or a plurality of phonemes (voiced sound or unvoiced sound) and includes a transition between phonemes (phoneme chain).

  The part (A) in FIG. 2 shows the speech waveform (envelope) Va of the variable unit, and the part (B) in FIG. 2 shows the speech waveform (envelope) Vb of the stationary unit. Yes. As shown in part (A) of FIG. 2, for a speech unit classified as a variable unit, a sample sequence over the entire section of the speech waveform Va when a specific speaker speaks the speech unit is a prime unit. It is stored in the storage device 12 as a piece of data W. On the other hand, as shown in the part (B) of FIG. 2, with respect to a speech unit classified as a stationary unit, a speech waveform Vb (acoustic characteristics are steady when a specific speaker speaks the speech unit). A set of sample sequences of M (three in the following example) unit waveforms u [1] to u [M] extracted from different positions on the time axis. It is stored in the storage device 12 as a piece of data W. In the first embodiment, each unit waveform u [m] (m = 1 to M) corresponding to one stationary element is equivalent to one period (for example, several milliseconds) of a voice waveform Vb of a voiced sound that is continuous in time. This is a section of the time length T0 corresponding to the degree. Each of the M unit waveforms u [1] to u [M] has similar acoustic characteristics to the extent that they are perceived as common speech segments by the listener. However, since the voice waveforms Vb are extracted from different time points, the acoustic characteristics (waveforms) of the M unit waveforms u [1] to u [M] continuously utter a single speech segment. In this case, they are different from each other within the range of fluctuation (fluctuation) of the acoustic characteristics.

  The synthesis information Z stored in the storage device 12 of FIG. 1 is information (score data) that specifies the synthesis target sound in time series. As shown in FIG. 1, the synthesis information Z designates a pitch Zb, a pronunciation time Zc, a duration Zd, and a volume Ze for each of a plurality of speech segments Za constituting the synthesis target sound. In addition to the information exemplified above (or in place of the above information), information such as volume and velocity can be designated by the composite information Z.

  The input device 14 is a device (for example, a mouse or a keyboard) that receives an instruction from a user. The display device 16 (for example, a liquid crystal display device) displays an image instructed from the arithmetic processing device 10. The sound emitting device 18 (for example, a speaker or headphones) emits a sound wave corresponding to the sound signal SOUT generated by the arithmetic processing device 10.

  The display control unit 22 in FIG. 1 causes the display device 16 to display the editing screen 40 of the part (A) in FIG. 3 that is visually recognized by the user for generating and editing the composite information Z. As shown in part (A) of FIG. 3, the editing screen 40 is an image (staff paper type or piano roll type) in which a time axis (horizontal axis) and a pitch axis (vertical axis) intersect each other are set. Image). By appropriately operating the input device 14 while referring to the editing screen 40, the user can change the arrangement of the note images 42 obtained by graphicizing the synthesis target sound, change the position and size of each note image 42, and each synthesis target sound. It is possible to instruct the speech synthesizer 100 to specify a pronunciation character (for example, a syllable of lyrics). The format of the edit screen 40 is arbitrary. For example, it is also possible to display a list of numerical values of each piece of information (speech segment Za, pitch Zb, pronunciation time Zc, duration Zd, volume Ze) of the synthesis information Z as the editing screen 40.

  The information generation unit 24 in FIG. 1 generates or updates the composite information Z in accordance with an instruction from the user with respect to the editing screen 40. Specifically, the information generation unit 24 sets each speech segment Za of the synthesis information Z according to the phonetic character specified in the note image 42. For example, for the phonetic character “ma [ma]” illustrated in the part (A) of FIG. 3, as shown in the part (B) of FIG. 3, [# -m], [m-a], [a], It is converted into four speech segments Za called [a- #] (#: silence). In the above example, a diphone is exemplified, but the pronunciation character “ma [ma]” is converted into two speech segments Za, such as [m] and [a], when using a monophone, for example. When using a phone, it is converted into two speech segments Za called [# -m-a] and [a- #]. Further, the information generating unit 24 sets each pitch Zb according to the position on the pitch axis of the note image 42, and sets the pronunciation time Zc of each speech element Za according to the position on the time axis of the note image 42. The continuation length Zd is set according to the length of the note image 42 on the time axis. Similarly, the volume Ze is set according to an instruction from the user.

  The segment selection unit 26 selects the segment data W corresponding to each speech segment Za specified by the synthesis information Z at the time corresponding to the sound generation time Zc of each speech segment Za. Select sequentially. The speech synthesizer 28 generates a speech signal SOUT using the segment data W selected by the segment selector 26. Specifically, the speech synthesis unit 28 designates the synthesis information Z for each selected speech unit for each speech unit of the segment data W selected by the segment selection unit 26 (hereinafter referred to as “selected segment”). The segment waveform Q adjusted to the pitch Zb, the duration Zd, and the volume Ze is generated from the segment data W, and the adjacent segment waveforms Q are connected to each other to generate the audio signal SOUT. FIG. 4 is a flowchart of a process in which the speech synthesizer 28 generates the segment waveform Q. Each time the element selection unit 26 selects the element data W, the process of FIG. 4 is executed.

  When the segment selection unit 26 selects the segment data W, the speech synthesis unit 28 determines whether the selected segment is a stationary segment (SA1). The method for distinguishing between the stationary unit and the variable unit is arbitrary. For example, information indicating the type of the speech unit (stationary unit / variable unit) is added to the unit data W in advance, and the information is stored. A configuration in which the speech synthesizer 28 distinguishes between a stationary element and a variable element by referring to the reference may be adopted. When the selected segment is a variable segment (SA1: NO), the speech synthesizer 28 synthesizes the segment data W selected by the segment selector 26 (the speech waveform Va of part (A) in FIG. 2). The segment waveform Q of the selected segment is generated by adjusting the pitch Zb, duration Zd, and volume Ze specified by the information Z for the selected segment (SA2).

  On the other hand, when the selected segment is a stationary segment (SA1: YES), the speech synthesizer 28 selects the M unit waveforms u [1] to u [M] included in the segment data W of the selected segment. A process of generating the segment waveform Q (hereinafter referred to as “waveform generation process”) by selectively arranging each of them on the time axis is executed (SA3).

  FIG. 5 is a flowchart of the waveform generation process (process SA3 of FIG. 4), and FIG. 6 is an explanatory diagram of the waveform generation process. When the processing of FIG. 5 is started, the speech synthesizer 28 sets the continuation length Zd designated by the synthesis information Z for the selected segment to N processing periods R [1] to R [N] as shown in FIG. Classify (SB1). The time length Lr [n] of each processing period R [n] (n = 1 to N) is set at random. However, each time length Lr [n] corresponds to an integer multiple of the time length T0 of the unit waveform u [m], and the total of the N time lengths Lr [1] to Lr [N] is the continuation length Zd. (Lr [1] + Lr [2] +... + Lr [N] = Zd).

  The time length Lr [n] of the first embodiment is defined as an added value of the reference length L0 and the fluctuation length d [n] (Lr [n] = L0 + d [n]). The speech synthesizer 28 randomly sets each of the N fluctuation lengths d [n] within a predetermined range, and adds each fluctuation length d [n] to a predetermined reference length L0, thereby processing period R [ The time length Lr [n] of n] is set. Therefore, the time length Lr [n] of each processing period R [n] can be different. Further, the number N of the processing periods R [n] changes according to the continuation length Zd.

  As shown in FIG. 6, the speech synthesizer 28 selectively arranges M unit waveforms u [1] to u [M] included in the segment data W of the selected segment on the time axis. The combined waveform C [n] having the time length Lr [n] is generated for each processing period R [n] (SB2 to SB6). A waveform obtained by concatenating N synthesized waveforms C [n] is applied to the generation of the audio signal SOUT as a segment waveform Q. In FIG. 6, the time change of the intensity (amplitude or power) of each unit waveform u [m] is schematically illustrated.

  The speech synthesizer 28 initializes a variable n designating one processing period R [n] to 1 (SB2). Then, the speech synthesizer 28 sets two different unit waveforms u [m] among the M unit waveforms u [1] to u [M] included in the segment data W of the selected segment as the first unit. The waveform Ua [n] and the second unit waveform Ub [n] are selected (SB3).

  Specifically, the speech synthesizer 28 uses the first unit waveform Ua [n-1] in the immediately previous processing period R [n-1] as the second unit waveform Ub [n] in the current processing period R [n]. ], And the first unit of the processing period R [n] is randomly selected from (M−1) of the M unit waveforms u [1] to u [M] except for the second unit waveform Ub [n]. Select unit waveform Ua [n]. For the first processing period R [1], any one of M unit waveforms u [1] to u [M] (for example, one randomly or fixedly selected from M) is used. The unit waveform u [m] is selected as the second unit waveform Ub [n].

  For example, as shown in FIG. 6, in the first processing period R [1] within the duration Zd, the unit waveform u [3] is selected as the first unit waveform Ua [1] and the unit waveform u [2] is selected. Selected as the second unit waveform Ub [1]. In the processing period R [2] immediately after, the unit waveform u [1] is selected as a new first unit waveform Ua [2], and the unit waveform u [3] is selected as the second unit waveform Ub [2]. Selected from [1]. In the processing period R [3], the unit waveform u [2] is selected as the new first unit waveform Ua [3], and the unit waveform u [1] is selected as the second unit waveform Ub [3]. Selected from [2].

  When the first unit waveform Ua [n] and the second unit waveform Ub [n] in the processing period R [n] are selected as described above, the speech synthesis unit 28, as shown in FIG. The processing period R [n] is a crossfade of the first waveform series Sa [n] in which the waveforms Ua [n] are arranged and the second waveform series Sb [n] in which the plurality of second unit waveforms Ub [n] are arranged. The synthesized waveform C [n] is generated (SB4). Specifically, the first waveform series Sa [n] includes the number of first unit waveforms Ua [n] over the time length Lr [n] of the processing period R [n] (Lr [n] / T0), It is a time series arranged so that the intensity (amplitude) of each first unit waveform Ua [n] increases with time. On the other hand, the second waveform series Sb [n] includes the second unit waveforms Ub [n] of the number (Lr [n] / T0) of the second unit waveforms Ub [n] over the time length Lr [n] of the processing period R [n]. This is a time series arranged so that the intensity (amplitude) of the unit waveform Ub [n] decreases with time. The speech synthesis unit 28 generates a synthesized waveform C [n] by adding the first waveform series Sa [n] and the second waveform series Sb [n].

  The speech synthesizer 28 determines whether or not the synthesized waveform C [n] (C [1] to C [N]) has been generated for all of the N processing periods R [1] to R [N] ( SB5). When the result of the process SB5 is negative, 1 is added to the variable n (SB6), and the process period R [n] corresponding to the updated variable n (that is, the synthesized waveform C [n-1] is generated immediately before). The composite waveform C [n] is generated by executing the process SB3 to the process SB5 for the process period R [n] immediately after the process period R [n-1].

  When the generation of N synthesized waveforms C [1] to C [N] is completed by repeating the above processing (SB5: YES), the speech synthesizer 28 sets the N synthesized waveforms C [1] to C [N]. ] Are arranged on the time axis to generate a segment waveform Q0 (SB7). Then, the speech synthesizer 28 generates the segment waveform Q by adjusting the segment waveform Q0 generated in the process SB7 to the pitch Zb and volume Ze specified by the synthesis information Z as the selected segment (SB8). . As understood from the above description, the segment waveform Q of the pitch Zb and the volume Ze over the continuation length Zd specified by the synthesis information Z for the selected segment is generated for the selected segment. As described above, the speech signal SOUT is generated by concatenating the segment waveform Q generated in the process SA2 for the variable segment and the segment waveform Q generated in the waveform generation process SA3 (process SB8) for the stationary segment. .

  As understood from the above description, in the first embodiment, M unit waveforms u [1] to u [M] extracted from different positions on the time axis in the speech waveform Vb are appropriately selected. As a result, the combined waveform C [n] is generated. Therefore, for example, when compared with a configuration in which one speech waveform Vb is repeated at the time of generating a stationary phoneme (for example, the configuration of Patent Document 1), the periodicity of the characteristic change that occurs in the speech signal SOUT due to the repetition of the speech waveform Vb. Is less perceptible to the listener (that is, a high-quality sound signal SOUT can be generated).

  In the first embodiment, in particular, since the composite waveform C [n] is generated by cross-fading the first waveform series Sa [n] and the second waveform series Sb [n], for example, a plurality of unit waveforms u [m] As compared with the configuration in which the synthesized waveform C [n] is generated by selectively arranging the waveforms, the effect that it is difficult to perceive the periodicity of the characteristic change in the segment waveform Q is particularly remarkable. In the first embodiment, since each processing period R [n] can be set to a different time length Lr [n], N processing periods R [1] to R [N] have the same time length. Compared to the set configuration, the effect that the periodicity of the characteristic change in the segment waveform Q is difficult to be perceived is particularly remarkable. In the first embodiment, the unit waveform u [m] selected as the first unit waveform Ua [n-1] in the processing period R [n-1] continues in the processing period R [n] immediately after the second. The unit waveform Ub [n] is selected. Therefore, the segment waveform Q is compared with the configuration in which both the first unit waveform Ua [n] and the second unit waveform Ub [n] are selected regardless of the selection target in the immediately preceding processing period R [n]. There is an advantage that the periodicity of the characteristic change is reduced.

  In the first embodiment, since a plurality of portions (unit waveform u [m]) extracted from the speech waveform Vb are stored in the storage device 12, the entire section of the speech waveform Vb is stored in the storage device 12. As compared with the above, there is an advantage that the storage capacity required for the storage device 12 is reduced. In particular, in the first embodiment, since one period of the voice waveform Vb is stored in the storage device 12 as each unit waveform u [m], the effect of reducing the storage capacity is particularly remarkable. Note that in a portable device such as a mobile phone or a portable information terminal, the storage capacity is more limited than, for example, a stationary information processing apparatus. Is particularly effective when installed in a portable device.

<B: Second Embodiment>
FIG. 7 is a block diagram of a speech processing apparatus 200 according to the second embodiment of the present invention. The speech processing apparatus 200 generates M unit waveforms u [1] to u [M] that are used to generate the unit waveform Q of stationary phonemes in the speech synthesizer 100 of the first embodiment.

  As shown in FIG. 7, the audio processing device 200 is realized by a computer system including an arithmetic processing device 50 and a storage device 52. The storage device 52 stores a program PGM2 executed by the arithmetic processing device 50 and various types of information stored by the arithmetic processing device 50. For example, a speech waveform Vb that is a material of M unit waveforms u [1] to u [M] is stored in the storage device 52. The voice waveform Vb is a sample series indicating a voice produced by uttering a voiced voice segment continuously in time. For example, a voice waveform Vb picked up by a sound collecting device (not shown) connected to the voice processing device 200 or a voice waveform Vb supplied from various recording media such as an optical disk or a communication network such as the Internet is stored in the storage device 52. Stored. In the following description, only one speech waveform Vb is referred to for convenience, but actually, a plurality of speech waveforms Vb corresponding to different speech segments are stored in the storage device 52, and a plurality of examples illustrated below. The unit waveform u [m] is sequentially generated for each voice waveform Vb.

  The arithmetic processing unit 50 executes a plurality of functions (waveform extraction unit 62) for generating M unit waveforms u [1] to u [M] from the speech waveform Vb by executing the program PGM2 stored in the storage device 52. , A waveform correction unit 64) is realized. A configuration in which each function of the arithmetic processing unit 50 is distributed over a plurality of integrated circuits, or a configuration in which a dedicated electronic circuit (DSP) realizes each function may be employed.

  FIG. 8 shows a speech waveform Vb (envelope) stored in the storage device 52. As shown in FIG. 8, the waveform extraction unit 62 includes M (three in the following example) unit waveforms x [1] from different positions on the time axis in the speech waveform Vb stored in the storage device 52. Extract ~ x [M]. Each unit waveform x [m] is a section corresponding to one cycle of the speech waveform Vb. A known technique is arbitrarily employed for extracting the unit waveform x [m].

  Even when the speaker continuously utters one speech unit, the acoustic characteristics (amplitude and period) of the actual speech waveform Vb vary with time, so each unit waveform x extracted from the speech waveform Vb. The acoustic properties of [m] can be different. The waveform correction unit 64 in FIG. 7 generates M unit waveforms u [1] to u [M] by correcting (normalizing) the acoustic characteristics of the unit waveforms x [m] to be similar to each other. To do. As shown in FIG. 7, the waveform correction unit 64 includes an amplitude correction unit 72, a period correction unit 74, and a phase correction unit 76.

  As shown in FIG. 8, the peak-to-peak value A [m] of the intensity (amplitude) of each unit waveform x [m] can be different due to the temporal variation of the amplitude in the speech waveform Vb. The peak-to-peak value A [m] means a difference (total amplitude) between the maximum value and the minimum value of the intensity of the unit waveform x [m]. The amplitude correction unit 72 corrects each unit waveform x [m] so that the peak-to-peak value A [m] of the unit waveform x [m] is adjusted to a predetermined value A0 (for example, the unit waveform x [m] is amplitude-adjusted). The unit waveform yA [m] (yA [1] to yA [M]) is generated by expanding and contracting in the direction. The method of correction by the amplitude correction unit 72 is arbitrary. For example, the unit waveform x [m] is multiplied as a correction value by the ratio (A0 / A [m]) of the predetermined value A0 to the peak-to-peak value A [m]. Is preferred.

  Further, due to the temporal variation of the period in the voice waveform Vb, the time length of each unit waveform x [m] (one period of the voice waveform Vb) T [m] may be different. 7 corrects each unit waveform yA [m] so that the period T [m] of the unit waveform yA [m] corrected by the amplitude correction unit 72 is adjusted to a predetermined value T0. A unit waveform yB [m] (yB [1] to yB [M]) is generated. Although the correction method by the period correction unit 74 is arbitrary, for example, the following method is preferable.

  Part (A) of FIG. 9 is a waveform diagram of the unit waveform yA [m] after correction by the amplitude correction unit 72. First, as illustrated in the part (B) of FIG. 9, the period correcting unit 74 expands and contracts each unit waveform yA [m] on the time axis to thereby unit time y ′ ′ of time length T ′ [m]. [m] (yA ′ [1] to yA ′ [M]) is generated. The time length T ′ [m] is an integral multiple of the sampling period of the voice waveform Vb and is the closest to the time length T [m] of the unit waveform yA [m] (for example, the time length T [m] Integer part). Each unit waveform yA ′ [m] is generated so that the intensity (signal value) becomes zero at the start point ts and the end point te. Secondly, as illustrated in the part (C) of FIG. 9, the period correcting unit 74 expands and contracts the unit waveform yA ′ [m] on the time axis to thereby unit time yB [m] (yB [1] to yB [M]). The time length T0 is set to, for example, the mode value of the time length T ′ [m] of each unit waveform yA ′ [m] (thus, an integer multiple of the sampling period).

  The peak-to-peak value A0 and the time length T0 of the M unit waveforms yB [1] to yB [M] are normalized (shared) by the above processing, but the waveform extraction unit 62 of the speech waveform Vb is set to 1. Depending on the position on the time axis of each unit waveform x [m] extracted as a period, the correlation of the waveform of each unit waveform yB [m] may be low. For example, the unit waveform yB [1] in part (A) of FIG. 10 has a maximum point (peak) immediately after the start point, whereas the unit waveform yB [2] in part (B) of FIG. There is a difference that a dip comes soon after. 7 corrects each unit waveform yB [m] so that the correlation between the waveforms increases among the M unit waveforms yB [1] to yB [M] corrected by the period correction unit 74. The unit waveform u [m] (u [1] to u [M]) is generated by correcting the phase.

  The phase correction unit 76 selects one unit waveform yB [m] among the M unit waveforms yB [1] to yB [M] corrected by the period correction unit 74 as the reference waveform yREF. FIG. 10 illustrates the case where the unit waveform yB [1] illustrated in the part (A) is the reference waveform yREF. The phase correction unit 76 calculates a cross-correlation function Fm (τ) with the reference waveform yREF for each of (M−1) unit waveforms yB [m] other than the reference waveform yREF. The variable τ is a time difference (shift amount) of the unit waveform yB [m] with respect to the reference waveform yREF. As illustrated in part (C) of FIG. 10, the phase correction unit 76 sets the start point ts of the unit waveform yB [m] on the time axis for the time of the variable τ at which the cross-correlation function Fm (τ) is maximum. The unit waveform u [m] is generated by moving (shifting the unit waveform yB [m]). As shown in part (C) of FIG. 10, the section of the unit waveform yB [m] before the start point ts after movement is added to the end of the unit waveform yB [m]. Note that two periods of the voice waveform Vb are extracted as the unit waveform x [m] by the waveform extraction unit 62, and the start point ts of the unit waveform yB [m] is the time of the variable τ that maximizes the cross-correlation function Fm (τ). It is also possible for the phase correction unit 76 to extract a unit waveform u [m] for one period starting from the time point elapsed since

  As understood from the above description, the M unit waveforms u [1] to u [M] in the first embodiment share a peak-to-peak value A0 and a time length T0, and have a cross-correlation function Fm. The phase is adjusted so that (τ) is maximized. The M unit waveforms u [1] to u [M] generated by the waveform correction unit 64 are stored in the storage device 52 as shown in FIG. 7, and the first unit waveforms u, for example, via a communication network or a portable recording medium. It is transferred to the storage device 12 of the speech synthesizer 100 of the embodiment.

  In the second embodiment, since the peak-to-peak values of the M unit waveforms u [1] to u [M] are adjusted to the predetermined value A0, the peak-to-peak values are different for each unit waveform u [m]. Compared with the configuration, amplitude fluctuations in the combined waveform C [n] (segment waveform Q) generated using the unit waveform u [m] are suppressed. Further, since the time lengths of the M unit waveforms u [1] to u [M] are adjusted to the predetermined value T0, the unit waveform u is compared with the configuration in which the time lengths of the unit waveforms u [m] are different. Variations in the period (pitch) in the synthesized waveform C [n] generated using [m] are suppressed. Therefore, it is possible to generate a sound with an acoustically natural impression in a section (stationary part) of a stationary segment with small fluctuations in amplitude and period among synthesis target sounds.

  When the correlation between the unit waveforms u [m] is low, the first unit waveform Ua [n] is added (cross-fade) to the first waveform series Sa [n] and the second waveform series Sb [n]. And the second unit waveform Ub [n] cancel each other, and the reproduced sound of the synthesized waveform C [n] may be audibly unnatural. In the second embodiment, since the phase of each unit waveform u [m] is adjusted so that the cross-correlation function Fm (τ) is maximized, it is possible to generate a sound with an acoustically natural impression. .

  Note that the order of processing by each element of the waveform correction unit 64 is appropriately changed. For example, a configuration in which the amplitude correction unit 72 corrects the amplitude after the cycle correction by the cycle correction unit 74 may be employed. Further, each element of the waveform correction unit 64 is omitted as appropriate. That is, the waveform correction unit 64 is included as an element including at least one of the amplitude correction unit 72, the period correction unit 74, and the phase correction unit 76.

<C: Third Embodiment>
As described in the second embodiment, the period correction unit 74 adjusts the period T [m] of each unit waveform yA [m] to a predetermined value T0. The third embodiment is a specific example of the period correction unit 74 focusing on a method for selecting a time length (predetermined length T0) of each unit waveform yB [m]. FIG. 11 is a block diagram of the period correction unit 74 of the third embodiment, and FIG. 12 is an explanatory diagram of the operation of the period correction unit 74 of the third embodiment. As shown in FIG. 11, the period correction unit 74 of the third embodiment includes an index calculation unit 742 and a correction processing unit 744.

As shown in FIG. 12, the index calculation unit 742 calculates the distortion index value D [k] (k = 1 to K) for each of a plurality (K) of different candidate lengths X [1] to X [K]. Calculate. The candidate length X [k] is a time length that is a candidate for the predetermined length T0, and is set to a time length that is an integral multiple of the sampling period of the speech waveform Vb. For example, the candidate length X [1] is set to the time length T ′ [1] of the unit waveform yA ′ [1] described in the second embodiment, and the candidate length X [2] is the unit waveform yA ′ [2]. The candidate length X [3] is set to the time length T ′ [3] of the unit waveform yA ′ [3] (K = M = 3). The distortion index value D [k] is each unit when each of the M unit waveforms yA [1] to yA [M] is expanded or contracted from the initial period T [m] to a common candidate length X [k]. This is an index indicating the degree of distortion of the waveform yA [m] on the time axis (the degree of deformation of the unit waveform yA [m] before and after expansion / contraction). Assuming that there are three unit waveforms yA [m] as shown in FIG. 12 (M = 3), the distortion index value D [k] is calculated by the following equation (1), for example.
D [k] = | T [1] −X [k] | / X [k] + | T [2] −X [k] | / X [k] + | T [3] −X [k] | / X [k] (1)
As can be understood from Equation (1), the difference between the period T [m] of each unit waveform yA [m] and the candidate length X [k] is large (the waveform of the waveform when expanded or contracted to the candidate length X [k]). The greater the deformation), the larger the distortion index value D [k].

  As shown in FIG. 12, the correction processing unit 744 in FIG. 11 minimizes the degree of distortion expressed by the distortion index value D [k] among the K candidate lengths X [1] to X [K]. The candidate length X [k] (that is, the candidate length X [k] corresponding to the minimum distortion index value D [k]) is selected as the predetermined length T0, and each unit waveform yA [m after correction by the amplitude correction unit 72 is selected. ] Is adjusted to a common predetermined length T0 to generate a unit waveform yB [m]. The expansion / contraction method of each unit waveform yA [m] is the same as in the second embodiment.

  As described above, in the third embodiment, each unit after adjustment so that the degree of expansion / contraction (distortion index value D [k]) of M unit waveforms yA [1] to yA [M] is minimized. Since the predetermined length T0 of the waveform yB [m] is variably set, the difference between the unit waveform yA [m] before correction by the period correction unit 74 and the unit waveform yB [m] after correction (the sound of the audio waveform Vb) There is an advantage that the deviation from the characteristic is reduced.

  In the second embodiment, the time length T ′ [m] of each unit waveform yA ′ [m] is calculated by rounding off the decimal part of the period T [m] of each unit waveform yA [m]. It is also possible to calculate the time length T ′ [m] of each unit waveform yA ′ [m] by rounding up the decimal part of the period T [m] of the unit waveform yA [m]. Therefore, in the third embodiment, as illustrated below, the time length Ta ′ [m] obtained by discarding the fractional part of the period T [m] of each unit waveform yA [m] and the period of each unit waveform yA [m] The time length Tb ′ [m] obtained by rounding up the decimal part of T [m] can be set as each candidate length X [k].

  For example, the candidate length X [1] is set to the time length Ta ′ [1] obtained by truncating the fractional part of the period T [1] of the unit waveform yA [1], and the candidate length X [2] is set to the unit waveform yA. A time length Tb ′ [1] obtained by rounding up the decimal part of the period T [1] of [1] is set. The candidate length X [3] is set to the time length Ta ′ [2] obtained by discarding the decimal part of the period T [2] of the unit waveform yA [2], and the candidate length X [4] is set to the unit waveform yA [2 ] Is set to a time length Tb ′ [2] obtained by rounding up the decimal part of the period T [2]. Similarly, the candidate length X [5] is set to the time length Ta ′ [3] obtained by truncating the fractional part of the period T [3] of the unit waveform yA [3], and the candidate length X [6] is the unit waveform. The time length Tb ′ [3] is set by rounding up the fractional part of the cycle T [3] of yA [3]. That is, six candidate lengths X [1] to X [6] corresponding to the combination of each unit waveform yA [m] and the period T [m] are set.

  The index calculation unit 742 calculates the distortion index value D [k] (D [1] to D [6]) for each candidate length X [k] by the calculation of the mathematical formula (1) described above, and the correction processing unit 744 The candidate length X [k] that minimizes the distortion index value D [k] among the six candidate lengths X [1] to X [6] is determined as the adjusted predetermined length T0. In the above configuration, the same effect as that of the third embodiment is realized.

In addition, the calculation method of each distortion index value D [k] is changed as appropriate. For example, in Equation (1), the absolute value | T [m] −X [k] | of the difference between the period T [m] and the candidate length X [k] is calculated in order to make each term a positive number. However, each term can be made a positive number by squaring the ratio of the difference between the period T [m] and the candidate length X [k] and the candidate length X [k] as shown in the following equation (2). Is possible.
D [k] = {(T [1] −X [k]) / X [k]} ²
+ {(T [2] −X [k]) / X [k]} ² + {(T [3] −X [k]) / X [k]} ² (2)

<D: Fourth Embodiment>
FIG. 13 is a block diagram of the waveform correction unit 64 in the fourth embodiment. As shown in FIG. 13, the waveform correction unit 64 of the fourth embodiment has a distortion correction unit 78 added to the elements (amplitude correction unit 72, period correction unit 74, phase correction unit 76) exemplified in the above-described embodiments. It is a configuration.

  When the period correction unit 74 expands and contracts the period T [m] of each unit waveform yA [m] to the time length T0, the peak-to-peak value A [m] of each unit waveform yB [m] expands and contracts on the time axis. The peak-to-peak value A0 immediately after the correction by the amplitude correction unit 72 (before the correction by the period correction unit 74) can be varied. That is, distortion occurs in each unit waveform yB [m] after correction by the period correction unit 74. Specifically, the time length T0 of the unit waveform yB [m] after correction by the period correction unit 74 is longer than the period T [m] of the unit waveform yA [m] before correction (the degree of expansion is high). ), The peak-to-peak value A [m] of the unit waveform yB [m] is smaller than the peak-to-peak value A0 immediately after correction by the amplitude correction unit 72, and the unit waveform after correction by the period correction unit 74 The shorter the time length T0 of yB [m] is compared to the period T [m] of the unit waveform yA [m] before correction (the higher the degree of contraction), the peak-to-peak value A of the unit waveform yB [m]. [m] is a larger value than the peak-to-peak value A0. In consideration of the above tendency, the distortion correction unit 78 of the fourth embodiment adjusts the peak-to-peak value A [m] of each unit waveform yB [m] corrected by the period correction unit 74 as described above. Correct the waveform distortion described.

  Specifically, the distortion correction unit 78 calculates the ratio (T0 / T [m]) of the time length T0 to the initial period T [m] of the unit waveform yA [m] after correction by the period correction unit 74. The peak-to-peak value A [m] of the unit waveform yB [m] is acted as a correction value (typically multiplied). As understood from the above description, the time length T0 of the unit waveform yB [m] after correction by the period correction unit 74 is longer than the period T [m] of the unit waveform yA [m] before correction (period As the degree of expansion by the correction unit 74 increases, the peak-to-peak value A [m] of the unit waveform yB [m] is corrected to a larger value by the processing by the distortion correction unit 78. Therefore, there is an advantage that the waveform distortion caused by the correction by the period correction unit 74 can be suppressed. The process by which the phase correction unit 76 corrects each unit waveform yB [m] after correction by the distortion correction unit 78 to generate each unit waveform u [m] is the same as in the second embodiment.

  In the fourth embodiment described above, the peak-to-peak value A [m] of each unit waveform yB [m] is corrected according to the degree of expansion / contraction of the unit waveform yA [m] by the period correction unit 74. There is an advantage that the unit waveform u [m] that accurately reflects the acoustic characteristics of the voice waveform Vb can be generated. Note that the method for selecting the predetermined length T0 in the fourth embodiment is arbitrary, and for example, the third embodiment described above in which the time length T0 is set according to the distortion index value D [k] is suitably employed.

<C: Fifth Embodiment>
In the fifth embodiment, the speech synthesizer 28 of the first embodiment is replaced with the speech synthesizer 28A of FIG. As shown in FIG. 14, the speech synthesis unit 28A includes a synthesis processing unit 82, an anharmonic component generation unit 84, a filter unit 86, and a synthesis unit 88. The synthesis processing unit 82 operates in the same manner as the speech synthesis unit 28 of the first embodiment and generates the audio signal HA. The audio signal HA corresponds to the audio signal SOUT of the first embodiment, and includes abundant harmonic components (basic tone component and harmonic component) corresponding to the pitch Zb and volume Ze specified by the synthesis information Z. As described above, the reproduction sound of the audio signal HA rich in harmonic components may be an artificial impression. Therefore, in the fifth embodiment, the audio signal SOUT is generated by adding the anharmonic component HB to the audio signal HA.

  The anharmonic component generator 84 generates an anharmonic component H0. The anharmonic component H0 is a noise component such as white noise or pink noise. The filter unit 86 generates an anharmonic component HB from the anharmonic component H0. For example, a comb filter that selectively passes band components other than the harmonic frequencies (fundamental frequencies and harmonic frequencies) corresponding to the pitch Zb in the anharmonic component H0 is suitable as the filter unit 86. The synthesis unit 88 adds the audio signal HA generated by the synthesis processing unit 82 and the anharmonic component HB generated by the filter unit 86 to generate the audio signal SOUT.

  In the fifth embodiment described above, since the anharmonic component HB is added to the audio signal HA generated by the synthesis processing unit 82, compared with the configuration in which the audio signal HA is output alone as the audio signal SOUT. There is an advantage that sound with a natural impression can be generated. A configuration in which the filter unit 86 in FIG. 14 is omitted (a configuration in which the anharmonic component H0 is added to the audio signal HA) may be employed.

<D: Modification>
Each of the above forms can be variously modified. Specific modifications are exemplified below. Two or more aspects arbitrarily selected from the following examples can be appropriately combined.

(1) Modification 1
The method of generating the composite waveform C [n] using the M unit waveforms u [1] to u [M] is appropriately changed. For example, a configuration in which unit waveforms u [m] sequentially selected from M unit waveforms u [1] to u [M] are arranged on the time axis to generate a composite waveform C [n] may be employed. . As understood from the above description, the speech synthesizer 28 of the first embodiment generates an audio signal SOUT by arranging M unit waveforms u [1] to u [M] on the time axis. It is an example of (waveform generation means).

  Further, in each of the above embodiments, the configuration in which each processing period R [n] is continued on the time axis is illustrated, but as shown in FIG. 15, a holding period in which a plurality of unit waveforms u [m] are arranged. It is also possible to insert E [n] between the processing period R [n] and the immediately subsequent processing period R [n + 1]. In the holding period E [n], a plurality of first unit waveforms Ua [n] selected in the immediately preceding processing period R [n] are arranged without changing the intensity. The time length Le [n] of each holding period E [n] can be set, for example, at random like the time length Lr [n] of the processing period R [n], but can also be set to a common fixed value. It is. As can be understood from the illustration of FIG. 15, a configuration in which consecutive processing periods R [n] are continued on the time axis is not essential in the present invention.

(2) Modification 2
The method of setting each processing period R [n] to a different time length Lr [n] is appropriately changed. For example, the time length Lr [n] is calculated by adding or subtracting a predetermined value to or from the time length Lr [n] to calculate the time length Lr [1] to Lr [N] of each processing period R [n]. ] Can be made different. In the first embodiment, the variation length d [n] of the time length Lr [n] is set to a random number. However, a configuration in which the time length Lr [n] itself is a random number may be employed. However, the time lengths Lr [1] to Lr [N] can be set to equal times.

(3) Modification 3
A method for selecting the first unit waveform Ua [n] and the second unit waveform Ub [n] for each processing period R [n] is arbitrary. For example, a configuration in which M unit waveforms u [1] to u [M] are sequentially selected as the first unit waveform Ua [n] for each processing period R [n] may be employed. In the first embodiment, the unit waveform u [m] selected as the first unit waveform Ua [n-1] in the processing period R [n-1] is continued in the immediately subsequent processing period R [n]. Although the unit waveform Ub [n] is selected, both the first unit waveform Ua [n] and the second unit waveform Ub [n] can be independently selected for each processing period R [n].

(4) Modification 4
In the second embodiment, the speech processing apparatus 200 separate from the speech synthesis apparatus 100 is illustrated, but the speech processing apparatus 200 that generates M unit waveforms u [1] to u [M] from the speech waveform Vb. The functions (the waveform extraction unit 62 and the waveform correction unit 64) can be mounted on the speech synthesizer 100.

DESCRIPTION OF SYMBOLS 100 ... Speech synthesizer, 200 ... Speech processing device, 10, 50 ... Arithmetic processing device, 12, 52 ... Storage device, 14 ... Input device, 16 ... Display device, 18 ... Sound emitting device, 22... Display control section 24. Information generation section 26. Segment selection section 28. Speech synthesis section 40. Editing screen 42. Note image 62 62 Waveform extraction section 64. ... Waveform correction section, 72 ... Amplitude correction section, 74 ... Period correction section, 76 ... Phase correction section, 82 ... Synthesis processing section, 84 ... Anharmonic component generation section, 86 ... Filter section, 88 ... ... Synthesizer, u [m] (u [1] to u [M]) ... Unit waveform, Ua [n] ... First unit waveform, Ub [n] ... Second unit waveform, Sa [n] ... 1st waveform series, Sb [n] ... 2nd waveform series, C [n] ... Composite waveform, Q ... Fragment waveform, Va, Vb ... Voice waveform, SOUT ... Voice signal.

Claims (6)

Waveform storage means for storing a plurality of unit waveforms extracted from different positions on the time axis among voiced sound waveforms;
For each of a plurality of processing periods, a first waveform series in which a plurality of first unit waveforms selected from the plurality of unit waveforms are arranged so that the intensity increases with time within the processing period; A composite waveform is generated by adding a second waveform series in which a plurality of second unit waveforms different from the first unit waveform among the unit waveforms are arranged so that the intensity decreases with time within the processing period. A speech synthesizer comprising: a waveform generating means; The speech synthesizer according to claim 1, wherein each of the plurality of unit waveforms corresponds to one cycle of the speech waveform. The peak-to-peak value and time length of the unit waveform are common to the plurality of unit waveforms,
The speech synthesizer according to claim 1, wherein the phase of each of the plurality of unit waveforms is adjusted so that a cross-correlation function between the unit waveforms is maximized. The first unit waveform of one processing period of the plurality of processing periods and the second unit waveform of the processing period immediately after the one processing period of the plurality of processing periods are common unit waveforms. The speech synthesizer according to any one of claims 1 to 3. The waveform generating means includes
Selecting the first unit waveform randomly from the plurality of unit waveforms for each processing period;
The speech synthesizer according to any one of claims 1 to 4, wherein a time length of each of the plurality of processing periods is set at random. A speech processing device for generating a plurality of unit waveforms used in the speech synthesizer according to claim 1,
Waveform extraction means for extracting a plurality of unit waveforms from different positions on the time axis among voiced sound waveforms;
A speech processing apparatus comprising: a waveform correcting unit that corrects the plurality of unit waveforms extracted by the waveform extracting unit so that the acoustic characteristics of each unit waveform approach each other.

Images