JPH08263090A - Synthesis unit accumulating method and synthesis unit dictionary device - Google Patents

Abstract

PURPOSE: To provide a synthesis unit dictionary device in which the retrieval efficiency is increased during the selection of a synthesis unit that is to be used in a voice synthesis and local peak position information is efficiently accumulated. CONSTITUTION: A synthesis unit dictionary device 2 consists of an accent information file 25 in which accent relative position information, that expresses the separation in terms of relative number of moras between the starting sound syllable of an individual synthesis unit and the accent position of a sentence paragraph that becomes the basis of the synthesis unit, is accumulated corresponding to the synthesis unit, synthesis unit selecting means 21 and 22 which select the candidates for the synthesis units employing the accent relative position information during the retrieving of the synthesis units and a waveform change information file 26 in which voiced sound interval information is accumulated in a relatively small accumulating unit among the voiced sound interval information defined by the interval of a local peak and voiceless sound-silence interval information, voiceless sound-silence interval information is accumulated in a relatively large accumulating unit and an identification code of the accumulating unit is added to the prescribed part of each accumulating unit.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a speech synthesis unit dictionary device used for regular speech synthesis.

[0002]

2. Description of the Related Art A conventional synthesis unit dictionary device used for regular speech synthesis includes a plurality of synthesis units cut out from a speech waveform, prosody information such as pitch frequency, elapsed time, section power of each synthesis unit, and each synthesis unit. The phonological environment information before and after is accumulated. When synthesizing a voice, a desired synthesis unit candidate is searched by using the above-mentioned prosodic information and phonological environment as key information, and the parameter difference between the searched synthesis unit candidate and a target prosody pattern is evaluated by a predetermined evaluation formula. And a specific synthesis unit is determined.

Further, in the case of producing a more natural synthesized speech, the prosody of each synthesis unit to be connected is modified so that the prosodic pattern mismatch does not occur at the connection position of the synthesis unit. As a method of transforming this prosody, for example, Hirokawa, Hakoda, "Study of pitch control method in waveform editing type rule synthesis method", March 2002, Acoustical Society of Japan, Proc. A pitch synchronization waveform superimposing method using a local peak of the waveform is known.

[0004]

However, in the conventional synthesis unit dictionary device, the search range of the synthesis unit becomes wider as the storage amount of the synthesis unit increases. Further, when the prosodic pattern is used as the evaluation criterion, there are many synthetic units that are numerically similar, and it is difficult to judge superiority or inferiority. As a countermeasure, in addition to the above-mentioned parameters, it is possible to use phonological environment information of several phonemes before and after as an evaluation standard.
There is a problem that the search efficiency of the voice synthesis unit dictionary device becomes poor.

Further, in the case of realizing the above-mentioned "pitch synchronization waveform superimposing method using a local peak" in order to secure the naturalness of synthesized speech, local peak position information is extracted from the speech waveform in advance. However, it is necessary to store this. However, there is a problem that when the storage amount of the synthesis unit increases, the local peak position information to be stored increases in proportion to the increase.

In view of the above problems, an object of the present invention is to improve search efficiency when selecting a synthesis unit closest to a prosodic pattern of natural speech and efficiently store local peak position information with a small capacity. , And a synthesis unit dictionary device created by this method.

[0007]

The present invention firstly provides a method for accumulating synthetic units. In this method, when a plurality of synthesis units to be speech-synthesized are stored in a dictionary device, the starting syllable of each synthesis unit is relatively mora (syllabic Accent relative position information indicating whether or not they are apart is stored in association with the synthesis unit.

Further, when the local peak position information used for the prosodic control of the synthesis unit is also stored, the local peak (the maximum point of the amplitude in the pitch structure, the same applies hereinafter).
It is characterized in that different accumulation units are assigned to the voiced sound section information and the unvoiced sound / unvoiced section information, which are determined by the interval, and the identification code of the storage unit is attached to a predetermined part of each allocated storage unit. The voiced sound section information and the unvoiced sound / unvoiced section information are accumulated in the form of the number of samples between local peaks of PCM data obtained by, for example, PCM converting a synthesis unit cut out from a voice waveform.

The present invention also provides a synthesis unit dictionary device obtained by implementing the above method. This device is a synthesis unit dictionary device in which a plurality of synthesis units to be synthesized are accumulated, and the starting syllable of each synthesis unit is relatively separated from the accent position of the syllable serving as the basis of the synthesis unit by a number of moras. A synthesis for selecting a candidate for a synthesis unit using the accent information file in which accent relative position information indicating whether or not is stored in association with the synthesis unit and the accent relative position information stored in the accent information file when searching for the synthesis unit And unit selection means.

A synthesis unit dictionary device according to another configuration of the present invention is a synthesis unit dictionary device in which a plurality of synthesis units to be synthesized are accumulated, and voiced sound section information and unvoiced sounds / silences determined by the intervals of local peaks. The voiced sound section information is stored in a relatively small storage unit of the section information, unvoiced sound / unvoiced section information is stored in a relatively large storage unit, and the identification code of the storage unit is stored in a predetermined portion of each storage unit. It is characterized by having a local peak position information file attached.

[0011]

In general, a sentence is divided into several utterance units, and each utterance unit is superposed with a pitch frequency peculiar to a phrase, that is, an accent component, on a characteristic (speech tone component) in which the frequency drops with a constant slope. It is represented by a model. The individual models are described by the pitch arrangement of the sound in each mora, but the pitch pattern of the Japanese model tends to have a generally H-shape with the accent position as the apex. Therefore, if the type of accent type, that is, the position of the accent with respect to the composition unit is known, it is possible to know the outline of the pitch pattern of the bunsetsu which is the basis of the composition unit. In addition, there is a correlation between the number of morae in the bunsetsu and the duration of the phoneme, and the more the number of morae in the bunsetsu, the shorter the duration of the phoneme. That is, since the number of mora and the relative accent position roughly reflect the pitch and the duration length which are the reference when selecting the composition unit, by using the accent relative position information having such information, the composition unit It is possible to perform a search that roughly takes into account the duration of time and pitch shape. In this case, since it is not necessary to monitor the phoneme environment before and after an unnecessarily long time, it is possible to narrow the search section when selecting the synthesis unit and improve the search efficiency.

Further, when accumulating the local peak position information, the voiced sound section information is accumulated in a relatively small accumulation unit, and the unvoiced sound / unvoiced section information is accumulated in a large accumulation unit, so that a uniform accumulation unit is obtained. The required dictionary capacity is reduced compared to the case of storing. Further, the absolute position of the local peak is obtained by adding the size (data length) of each section information derived from the identification code of the storage unit. Therefore, the position and interval of the local peak can be easily derived in a short time.

[0013]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS A preferred embodiment of the synthetic speech unit dictionary device of the present invention will be described in detail below with reference to the drawings. FIG. 1 is a configuration diagram of a regular speech synthesizer according to an embodiment of the present invention. First, an overall processing outline of the rule speech synthesizing device 1 will be described with reference to FIG.

In the regular voice synthesizer 1, when synthesizing a voice, a character string to be synthesized, an accent type,
And a prosody pattern to the input terminal IN. Input terminal IN
The character string or the like input from is sent to the synthesis unit search unit 3. In the composition unit search unit 3, the input character string,
The speech synthesis unit is converted into predetermined key information to be used when searching, and based on the key information, the unit of speech synthesis unit to be synthesized from the synthesis unit dictionary device 2, the prosodic information of each segment and the relative position of the accent. Find information and information. The synthesis unit search unit 3 also embeds the prosodic information and the accent relative position information of the retrieved candidate segment into one by using an appropriate evaluation formula. Then, the acoustic parameter and the waveform deformation information of the narrowed down element are acquired. In this way, all the synthesis units necessary for synthesis are found. The synthesis unit connection unit 4 smoothly connects the retrieved synthesis units to prevent noise from mixing, and further transforms the prosody so that it is close to the target shape and sends it to the synthesis speech output unit 5. To do.
The synthetic voice output unit 5 converts the acoustic parameters sent from the synthesis unit connection unit 4 into a voice signal and outputs the voice signal via a speaker or the like.

Next, the synthesis unit dictionary device 2 according to the present embodiment.
Will be described. As shown in FIG. 2, the synthesis unit dictionary device 2 is connected to the synthesis unit search unit 3 so that the synthesis unit management unit 21 is bidirectionally accessible, and the synthesis unit management unit 21 is bidirectionally accessible. The composite unit search table 22, the acoustic parameter file 23, the prosody information file 24, the accent information file 25, and the waveform modification information file 26 are included.

The composition unit management unit 21 efficiently stores information to be registered and also functions as an interface for providing information to other processing units in response to a request. Further, the composition unit search table 22 is a table for searching a desired composition unit at high speed.
The acoustic parameter file 23 is an area for storing acoustic parameters used when generating synthetic speech. The prosody information file 24 is an area for storing information such as the average pitch frequency of each synthesis unit, duration time, section power, and type of adjacent phoneme. In this embodiment, the PCM (Pulse code
Since the digitized audio waveform is stored by the modulation method, the above various parameters are assumed to correspond to these digital data (PCM data).

The accent information file 25 includes accent relative position information indicating how many moras each synthesizing unit is cut out from, and how many moras each synthesizing unit is relatively separated from the accent position of the bunsetsu. This file is stored separately.

Details of the accent information file 25 will be described with reference to FIGS. 3 and 4. As described above, if the relative position of the accent with respect to the composition unit is known, it is possible to know the outline of the pitch pattern of the bunsetsu which is the basis of the composition unit. For example, in the case where the number of mora is 3, the flat type 3 mora 0 type, the 3 mora 1 type with an accent at the beginning, and the accent in the central part, as shown in FIGS. In the 3-mora type 2 and the 3-mora type 3 having an accent at the end, if the accent relative position with respect to the head position of each composition unit is known, the pitch pattern of the bunsetsu which is the basis of the composition unit can be known. Therefore,
In the present embodiment, the relative accent position is determined to be "0" when the head position is the accent position, "-1" when the position is one syllable before, and "+1" when the position is one syllable behind. In addition, for the flat type without accent, the second mora is set as the accent position for convenience. Then, this information is added for each synthesis unit. FIG. 4 shows a table image example 401 of the accent information file 25 created in this way.

In FIG. 4, the synthesizing unit number is a value given to each synthesizing unit by the rule speech synthesizing apparatus 1 of the present embodiment, and uniquely associates the number of mora and the relative position of accent. Therefore, it is possible to use the number of mora and the relative position of the accent when searching the composition unit number, and it is also possible to roughly know the pitch and the duration length which are the reference when selecting the composition unit.

Next, the detailed contents of the waveform modification information file 26 will be described with reference to FIGS. In the voiced section, the voice waveform has a periodic shape. This is generally called a pitch structure. Figure 5 for each pitch
As shown in, a peak (502,
503, 504) appears. These peaks are called local peaks. The pitch period of voice is approximately 2.5 ms
It is distributed over 10 ms, and the shorter the cycle, the higher the voice, and conversely, the longer the cycle, the lower the voice sounds. Therefore, the pitch of the voice can be controlled by changing the pitch period of the original voice. The pitch intervals in the actual speech waveform are as shown in the lower part of FIG. 5, and in the illustrated example, the cases of 3600 samples 505, 97 samples 506, and 92 samples 507 are shown. Each pitch interval is an example when voice with a sampling frequency of 10 kHz is used. Further, by performing the thinning-out and the turning-back process in pitch units, it becomes possible to control the duration time of the composition unit. The above-mentioned pitch synchronization waveform superposition method is a method for performing these prosody control, and in this method, the local peak position is used as a mark for each pitch.

FIG. 6 is an explanatory diagram of the storage procedure of the local peak position information in the waveform deformation information file 26, that is, the local peak position information file. In this embodiment, the local peak position information is represented by the number of samples between peaks in the PCM data. However, the local peak position at the beginning is the interval between the start end of the audio data and the peak position. In this way, the absolute position of the local peak can be obtained by adding the peak intervals from the beginning. The peak intervals tend to be roughly divided into short intervals in voiced sound intervals and unvoiced sounds or long intervals in unvoiced intervals. In the voiced sound section, the pitch period is about 2.5 ms to 10 ms, and when the sampling frequency is 10 kHz, local peaks appear at intervals of 25 to 100 sample. On the other hand, the unvoiced sound section is generally about 100 ms to 1 s, and the appearance interval of the local peak is 1000 to 10000 sample.
It is a degree.

Therefore, in this embodiment, a short storage unit and a long storage unit are prepared in the local peak position information file (waveform modification information file 26), and the length of the storage unit is selected according to each peak interval. . In the example shown in FIG.
8 bits are allocated to the short storage unit 601 shown in (a). However, the first 1 bit is a flag 6 indicating the storage unit length.
03, a logical 0 indicates a short storage unit, and the remaining 7
Allocate bit 602. As the expression ability of the peak interval, 0 to 12 sample can be expressed. Further, assuming that the peak interval cannot be too short (for example, about 0 to 2 ms), it is possible to further improve efficiency by assigning a constant value to a value that can be expressed. On the other hand, 16 bits are allocated to the long storage unit 604 shown in (b). Even in a long storage unit, the leading 1 bit is used as a flag 606 indicating the storage unit length. in this case,
Logic 1 is used as the identification code of the long storage unit. Then, the remaining 15 bits 605 are allocated. The peak interval can be expressed as 0 to 32786 sample.

In the synthesis unit 607 actually used, as shown in (c), first, the flag 60 of the head bit is used.
Check 8 to identify the length of the data to be acquired. Here, the peak interval is acquired and the data length of the first read 6
From 10, the read position of the next data (head flag of the next data) 609 is calculated. By performing this processing as needed, a series of local peak position information can be read. The local peak position is given for each synthesis unit, and the end position in each synthesis unit is the end value when a constant value is detected. Further, the reading position (address) of the starting end is designated by using predetermined management information.

In this way, the accent information file 25
By using, it becomes possible to perform a search in consideration of the pitch and the duration, and it is possible to narrow the search section when narrowing down the candidates of the composition unit. Since the accent information file 25 also holds the number of mora that has a great influence on the duration, the duration of the synthesis unit selected by using this as the evaluation criterion at the time of search is close to natural speech. Becomes In addition, since the local peak position information can be retrieved quickly, it is possible to control the prosody in real time, and since appropriate storage units are assigned according to the intervals of the local peaks, simply the same It is possible to store the same information with a capacity of about 50% and 25% of that of the storage unit. This makes it possible to make the size of the waveform dictionary much smaller than before.

The number of moras, relative positions of accents, and the number of samples of pitch intervals used in this embodiment are mere examples, and are not necessarily limited to the values in this embodiment.

[0026]

As is apparent from the above description, according to the synthesis unit accumulating method of the present invention and the synthesis unit dictionary device obtained by the method, a natural number is selected from a large number of synthesis units used in the waveform rule speech synthesizer. When selecting a synthesis unit that is closest to the prosodic pattern of speech, it is possible to suggest a rough search range, which has the effect of dramatically increasing search efficiency.

Further, there is an effect that the local peak position information of the voice waveform can be efficiently accumulated, and the necessary local peak position information can be quickly extracted.
For example, a remarkable effect can be exerted in the application of changing the utterance speed or the pitch of the synthesized voice in real time.

[Brief description of drawings]

FIG. 1 is a block diagram showing an example of a voice synthesizing device using a synthesis unit dictionary device according to an embodiment of the present invention.

FIG. 2 is a diagram showing an example of a block configuration of a composition unit dictionary device of the present embodiment.

FIG. 3 is an explanatory diagram of a relationship between a mora and an accent relative position according to the present embodiment.

FIG. 4 is an explanatory diagram of a table image of an accent information file according to the present embodiment.

FIG. 5 is an explanatory diagram of local peak and peak-to-peak information.

FIG. 6 is an explanatory diagram of an accumulated state of local peak position information in a waveform modification information file according to the present embodiment.

[Explanation of symbols]

 1 Ruled Speech Synthesizer 2 Synthesis Unit Dictionary Device 3 Synthesis Unit Search Unit 4 Synthesis Unit Connection Unit 5 Synthesis Speech Output Unit 2l Synthesis Unit Management Unit 22 Synthesis Unit Search Table 23 Acoustic Parameter File 24 Prosodic Information File 25 Accent Information File 26 Waveform Modification Information file

Claims (5)

[Claims] 1. When accumulating a plurality of synthesis units to be speech-synthesized in a dictionary device, the start syllable of each synthesis unit is relatively separated from the accent position of the syllable serving as the basis of the synthesis unit by a number of moras. A composition unit accumulating method characterized by accumulating accent relative position information indicating whether or not it is associated with the composition unit. 2. A method of accumulating local peak position information used for prosodic control of a synthesis unit when accumulating a synthesis unit to be voice-synthesized in a dictionary device, in a voiced sound section determined by an interval between local peaks. A synthetic unit accumulating method characterized by allocating different accumulating units to information and unvoiced sound / silent section information, and assigning an identification code of the accumulating unit to a predetermined part of each allocated accumulating unit. 3. The voiced sound section information and the unvoiced sound / unvoiced section information are the number of samples between local peaks of PCM data obtained by PCMizing a synthesis unit cut out from a voice waveform. The method for accumulating synthetic units described. 4. A synthesis unit dictionary device comprising a plurality of synthesis units to be synthesized, wherein the starting syllable of each synthesis unit is relatively far away from the accent position of the syllable serving as the basis of the synthesis unit. A candidate for a synthesis unit is selected by using the accent information file in which the accent relative position information indicating whether or not it is stored is associated with the synthesis unit and the accent relative position information stored in the accent information file when searching for the synthesis unit. A synthesizing unit selecting device, comprising: synthesizing unit selecting means. 5. A synthesizing unit dictionary device in which a plurality of synthesizing units to be synthesized are accumulated, which is stored in a relatively small accumulating unit of voiced sound section information and unvoiced sound / silent section information which is determined by a local peak interval. It is equipped with a local peak position information file that stores voice sound section information and also stores unvoiced sound / silent section information in a relatively large storage unit, and attaches an identification code of the storage unit to a predetermined part of each storage unit. A synthesis unit dictionary device characterized by the above.