JP2005352327A - Device and program for speech synthesis - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a device and program for speech synthesis that can obtain a natural long sound more easily. <P>SOLUTION: A text as an object of speech synthesis is inputted (S1), language analysis processing is performed (S2), and normal voice unit data are used for an output result to calculate pitches and sound volume by syllables (S3, S5). Then it is judged whether each syllable is a long sound (S9), has an accent (S13), and includes a pause (S17), thereby leaving normal speech unit data as they are (S11) when a syllable is not a long sound (NO at S9), continuing a last phoneme (S15) by using the normal speech unit data when the syllable is a long sound and has no accept (YES at S9 and NO at S13), changing the normal speech unit data into 1st speech unit data for a long sound (S19) when the syllable is a long sound, has an accent, and includes no pause (YES at S9, YES at S13, and NO at S17), and changing the data into 2nd speech unit data for a long sound (S21) when the syllable is a long sound, has an accent, and is before a pause (YES at S9, YES at S13, and YES at S17). <P>COPYRIGHT: (C)2006,JPO&NCIPI

Description

  The present invention relates to a speech synthesizer.

2. Description of the Related Art Conventionally, a speech synthesizer that generates synthesized speech corresponding to text input and a speech synthesis program for causing a computer to function as such a speech synthesizer have been developed. In such a speech synthesizer and program, as one of the points for improving quality, there is a problem of how to process a long sound. For example, in Patent Document 1, an input text is analyzed with reference to a language dictionary, and a portion with a possibility of lengthening is detected with reference to a lengthening candidate list prepared in advance for the analysis result. For the part where there is a candidate for the long sound, refer to the long sound / non-long sound rule, determine whether long sound / non-long sound is set, and if long sound is set The long sound processing is performed. Then, in the case of performing the lengthening process, the phoneme of the immediately preceding vowel is extended.
JP 2001-343987 A

  However, when a method of extending the immediately preceding vowel when the sound is made longer as in the conventional speech synthesizer, the long sound may sound unnaturally like a buzzer sound. In addition to the method of extending the immediately preceding vowel, it may be possible to superimpose the immediately preceding vowel, but in this case, it may sound discontinuous. In order to eliminate such unnaturalness, it is conceivable to store the previous phoneme and the cluster of long sounds as phonemes in the database, but if you try to store all possible combinations in the database, There is a problem that the necessary capacity becomes enormous.

  The present invention has been made to solve the above problems, and an object thereof is to provide a speech synthesizer and a speech synthesis program capable of obtaining a more natural long sound with a simple configuration.

  In order to achieve the above object, a speech synthesizer according to claim 1 of the present invention comprises normal speech unit data storage means for storing normal speech unit data corresponding to text reading information, and input text. Analyzing using a dictionary and generating reading information with accents, and applying normal voice unit data stored in the normal voice unit data storage means to reading information generated by the reading information generating means First speech unit data storage unit for long sound storing first sound unit data for long sound corresponding to reading information of long sound, and the reading information A long sound detecting means for detecting a long sound from the reading information generated by the generating means, and an access for determining whether or not the long sound detected by the long sound detecting means includes an accent. And the first sound unit data for long sound stored in the first sound unit data storage means for long sound, instead of the normal sound unit data, when the accent determination means determines that there is an accent. And a voice unit data changing unit to be applied.

  In addition to the configuration of the invention described in claim 1, the speech synthesizer described in claim 2 of the present invention stores the second sound unit data for long sound corresponding to the reading information of the long sound accompanied by a pause immediately after. Second sound unit data storage means for long sound, and pause determination means for determining whether or not there is a pause in the reading information immediately after the long sound detected by the long sound detection means, the sound unit data changing means When the pause determination means determines that there is a pause, the second sound unit data for long sound stored in the second sound unit data storage means for long sound is applied instead of the normal sound unit data, When the pause determination unit determines that there is no pause, the first sound unit data for long sound stored in the first sound unit data storage unit for long sound is applied.

  The speech synthesis program according to claim 3 of the present invention is generated in a reading information generation step of analyzing input text using a dictionary and generating accented reading information, and the reading information generation step. In speech synthesis for causing the computer to execute a speech synthesis step of outputting synthesized speech by applying the normal speech unit data stored in the normal speech unit data storage means so as to correspond to the reading information of the text, In a long sound detecting step for detecting a long sound from the reading information generated in the reading information generating step, an accent determining step for determining whether or not the long sound detected in the long sound detecting step includes an accent, and in the accent determining step If it is determined that there is an accent, the normal voice unit data The computer further executes a sound unit data changing step of applying the first sound unit data for long sound stored in the first sound unit data storage unit for long sound so as to correspond to the reading information of the long sound. To do.

  Further, in the speech synthesis program according to claim 4 of the present invention, in addition to the configuration of the invention according to claim 3, is there a pause in the reading information immediately after the long sound detected in the long sound detection step? The computer further executes a pause determination step for determining whether or not, and in the voice unit data changing step, when it is determined that there is a pause in the pause determination step, the pause is performed immediately instead of the normal voice unit data. When the second sound unit data for long sound stored in the second sound unit data storage means for long sound is applied so as to correspond to the reading information of the long sound accompanied by the sound, and it is determined that there is no pause in the pause determination step, The first sound unit data for long sound is applied.

  According to the speech synthesizer described in claim 1 of the present invention, two patterns of voice unit data for long sound, with and without accent, are prepared, and a text to be synthesized is input. First, a long sound is detected, and then whether or not the long sound includes an accent is determined, and then the voice unit data is applied to each of the cases with and without the accent, so that a more natural synthesized sound can be obtained. Can output.

  Further, according to the speech synthesizer according to claim 2 of the present invention, in addition to the effect of the invention according to claim 1, a long sound when there is an accent is further determined by whether or not there is a pause immediately thereafter. Separately, three patterns are prepared as sound data for long sound, so that a more natural synthesized sound can be output.

  According to the speech synthesis program of claim 3 of the present invention, two patterns of voice unit data for long sound are prepared, with and without accent, and the text to be synthesized is input. If a long sound is detected, it is first determined whether or not the long sound includes an accent, and then the voice unit data is applied to the case with or without the accent. Sound can be output.

  According to the speech synthesis program described in claim 4 of the present invention, in addition to the effect of the invention described in claim 3, a long sound when there is an accent is further determined by whether or not there is a pause immediately thereafter. Separately, three patterns are prepared as sound data for long sound, so that a more natural synthesized sound can be output.

  Next, the best mode for carrying out the present invention will be described with reference to the drawings. First, the configuration of a speech synthesizer 1 to which the present invention is applied will be described with reference to FIG. FIG. 1 is a block diagram showing an electrical configuration of the speech synthesizer 1.

  As shown in FIG. 1, the speech synthesizer 1 is provided with a CPU 10 that controls the entire speech synthesizer 1. The CPU 10 includes a ROM 12 that stores BIOS and the like, and a work area for various arithmetic processes executed by the CPU 10. RAM 13 used, an external storage device 20 such as a hard disk, an audio unit 31 for converting synthesized voice into an analog voice waveform signal and performing predetermined amplification, and operations for inputting various operations A panel 41 is connected via a bus. Further, the CPU 10 is connected with a timer 11 that generates an interrupt to the CPU 10 at predetermined time intervals. In addition, a speaker 32 is connected to the audio unit 31.

  Here, the external storage device 20 includes a speech synthesis program 21 for the speech synthesizer 1 to perform speech synthesis processing, a language dictionary 22 for language analysis of the input text, and three types of speech constituting the synthesized speech. Unit data 23 to 25 are stored. The language dictionary 22 stores data necessary for language analysis, such as morpheme, morpheme reading, part of speech, accent type, and grammar. The voice unit data includes normal voice unit data 23 used for voice synthesis processing of normal sound, first long voice unit data 24 used for accented long sound, and a pause immediately after the accent. Three types of second sound unit data 25 for long sound used for long sound are stored.

  In the normal speech unit data 23, speech unit data obtained by uttering each Japanese phoneme is extracted, and a frequency (pitch), duration length, and power (volume) corresponding to the phoneme are stored. Here, the phoneme is a unit of sound used in a certain language (Japanese in this embodiment), and is the smallest unit that causes a difference in meaning. The normal sound unit data 23 does not have long sound-specific sound unit data. Therefore, when speech synthesis is performed using normal speech unit data 23 for a long sound, the previous phoneme is extended and used.

  Further, the first sound unit data 24 for long sound stores sound unit data that is collected by cutting out utterances of long sounds having the same phoneme as the immediately preceding phoneme and having no pause immediately after. Further, the second sound unit data for long sound 25 stores sound unit data obtained by collecting and cutting out utterances of long sound having the same phoneme as the immediately preceding phoneme and having a pause immediately after. In Japanese accents, the pitch drops sharply from the accented syllable to the next syllable. Since there is a difference in how the pitch is lowered depending on whether or not there is, there are two types of voice unit data used for synthesizing a long sound depending on the presence or absence of a pause, thereby making it possible to obtain a more natural synthesized sound.

  Next, speech synthesis processing performed by the speech synthesizer 1 having the above configuration will be described with reference to FIGS. FIG. 2 is a flowchart showing the flow of the speech synthesis process. FIG. 3 is an explanatory diagram illustrating an example of the read character string 100 that is a target of the speech synthesis process executed by the speech synthesizer 1.

  First, a text to be subjected to speech synthesis processing is input (S1). Here, the target text is usually input as a kanji-kana mixed sentence. Input may be performed from the operation panel 41 or may be input from a keyboard connected to a keyboard. Further, the speech synthesizer 1 may be connected to a network and input from the outside, or a file stored in the external storage device 20 may be read or input from a storage medium such as a flexible disk or a CD-ROM. Well, the input method is not limited. For example, it is assumed that a kanji-kana mixed sentence “Joe, I covered New York for a week” is input.

  Next, the target text input in S1 is subjected to morphological analysis using a language dictionary 22 by a well-known method, and language analysis processing for adding reading, accents, and poses is executed (S2). When this language analysis processing is executed, reading, accent, and pose are added to the target text, and a reading character string 100 is output as shown in FIG. Here, the reading is indicated by a katakana character string, “|” is a pause, “′” is an accent, and “>” is an accent delimiter. Although not shown, the linguistic analysis process also outputs syllable breaks.

  Next, for each syllable of the reading character string 100 as the output result of S2, the pitch of each syllable is calculated by a known method using the normal speech unit data 23 (S3), and the volume is calculated ( The result is stored in a buffer in the RAM 13 (S7).

  Next, for each syllable of the output result 100, it is determined whether or not the syllable is a long sound, and sound unit data changing processing for changing sound unit data in the case of a long sound is executed (S9 to S21). First, it is determined whether or not the processing target syllable is a long sound (S9). If it is not a long sound (S9: NO), it is not necessary to change the voice unit data used for the syllable, so the normal voice unit data 23 is selected as it is (S11), and the process proceeds to S23.

  If the syllable to be processed is a long sound (S9: YES), it is further determined whether or not the syllable has an accent (S13). If there is no accent in the syllable (S13: NO), the sound unit data for long sound is not used, but the normal sound unit data 23 is used to extend the duration of the previous phoneme to make a long sound ( S15). Then, the process proceeds to S23.

  If the syllable to be processed is a long sound (S9: YES) and there is an accent (S13: YES), it is further determined whether or not the syllable is immediately before a pause (S17). If it is not immediately before the pause (S17: NO), the voice unit data stored in the first voice unit data 24 for long sound is selected as the voice unit data used for the long sound (S19). Then, the process proceeds to S23. As described above, the first sound unit data 24 for long sound is the sound unit data that has the same phoneme as the immediately preceding phoneme and is collected by cutting out the utterances of the long sound without the pause immediately thereafter. Thus, natural long sound can be synthesized in the ear.

  When the syllable to be processed is a long sound (S9: YES) and there is an accent (S13: YES) and the syllable is immediately before a pause (S17: YES), a long sound is used as voice unit data used for the long sound. The audio unit data stored in the second audio unit data 25 is selected (S21). As described above, the second sound unit data for long sound 25 is sound unit data obtained by collecting the utterances of long sounds that are paused immediately after, and the pitch decreasing method is different from the first sound unit data for long sound 24. Since the second sound unit data 25 for long sound is used when there is a pause, a more natural long sound can be synthesized.

  Through the processes of S9 to S21, the voice unit data to be used for the processing target syllable is selected. According to the selection, the pitch and volume calculated in S3 and S5 are changed (S23) and stored in the buffer of the RAM 12 (S25). Then, it is determined whether or not processing has been completed for all syllables (S27). If all syllables have been completed (S27: YES), processing is terminated. The synthesized speech signal stored in the buffer is sent to the audio unit 31, converted into a speech waveform signal, and output from the speaker 32 as an acoustic signal. If unprocessed syllables still remain (S27: NO), S9 to S25 are repeated with the next syllable as a processing target syllable.

  Next, the above processing will be specifically described with reference to the example shown in FIG. “Joe” in the first syllable 101 is a long sound (S9: YES), there is an accent (S13: YES), and there is a pause immediately after that (S17: YES), so normal voice data “jo” and long sound are used. The second audio unit data 25 “o-˜” is selected (S21), the pitch / volume is changed using the second sound unit data for long sound 25 which is the model for long sound before pause (S23), and stored in the buffer. (S27).

  Since the next syllable “I” is not a long sound (S9: NO), normal voice unit data is selected as it is (S11), and the pitch and volume are not changed (S23). The “shoe” of the third syllable 102 is a long sound (S9: YES) but has no accent (S13: NO). Accordingly, the duration of the vowel part “u” of “shu”, which is the selected normal voice unit data, is extended (S15).

  Since the fourth to sixth syllables “Kan” “B” “K” “Li” are not long sounds (S9: NO), normal voice unit data is selected as it is (S11), and the pitch and volume are changed. No (S23).

  “New” in the seventh syllable 103 is a long sound (S9: YES), but there is no accent (S13: NO). Therefore, the duration of the vowel part “u” of “nyu”, which is the corresponding voice unit data stored in the normal voice unit data 23, is extended (S15), and the pitch / volume is changed accordingly (S23). ) And stored in the buffer (S27).

  “Yaw” in the eighth syllable 104 is a long sound (S9: YES), has an accent (S13: YES), and has no pause after that (S17: NO), so the first sound unit data for long sound 24 "O-" is selected (S19), the pitch / volume is changed using the first sound unit data 24 for long sound, which is a long sound model without pause (S23), and stored in the buffer (S27).

  Since the ninth to last syllables have no long sound (S9: NO), normal voice unit data is selected as it is (S11), and the pitch and volume are not changed (S23). The voice unit data to be used for all syllables is determined / changed, the pitch and volume are set, and the synthesized voice signal is generated and stored in the buffer. Is sent to the audio unit 31. In response to this, the audio unit 31 converts the synthesized speech signal into a speech waveform signal and outputs it as an acoustic signal from the speaker 32.

  As described above, as long sound unit data, different models are prepared in advance depending on the presence or absence of an accent and the presence or absence of a pause immediately after that, and the sound unit data corresponding to each is used by detecting a long sound or a pause. It is possible to output a synthesized voice of a more natural long sound. That is, even in the case of syllables with the same long sound, but with accents, the pitch drops sharply before and after the accent. It is easier to hear naturally than using normal voice unit data. Moreover, since the way of decreasing the pitch differs depending on whether there is a pause immediately after that, if a long sound model is prepared separately depending on whether there is a pause, a more natural long sound can be obtained.

  In the above embodiment, the CPU 10 that executes language analysis processing in S2 of the flowchart of FIG. 2 functions as reading information generation means of the present invention, and the CPU 10 that executes pitch calculation / volume calculation processing in S3 and S5. The CPU 10 that functions as the speech synthesis means of the invention, determines whether or not the sound is a long sound in S9, functions as the sound detection means of the present invention, and the CPU 10 that determines the presence or absence of an accent in S13 functions as the accent determination means of the present invention, The CPU 10 that determines the presence or absence of a pause in S17 functions as the pause determination means of the present invention, selects the first sound unit data for long sound or the second sound unit data for long sound in S19 and S21, and changes the pitch and volume in S23. CPU 10 that functions as the voice unit data changing means of the present invention.

2 is a block diagram showing an electrical configuration of the speech synthesizer 1. FIG. It is a flowchart which shows the flow of a speech synthesis process. It is explanatory drawing which shows the example of the reading character string 100 used as the object of the speech synthesis process performed with the speech synthesizer.

Explanation of symbols

1 Speech Synthesizer 10 CPU
13 RAM
20 external storage device 21 speech synthesis program 22 language dictionary 23 normal speech unit data 24 first sound unit data for long sound 25 second sound unit data for long sound

Claims (4)

Normal speech unit data storage means for storing normal speech unit data corresponding to text reading information;
Reading information generation means for analyzing input text using a dictionary and generating accented reading information;
In a speech synthesizer comprising speech synthesis means for outputting synthesized speech by applying normal speech unit data stored in the normal speech unit data storage means to the reading information generated by the reading information generating means,
First sound unit data storage unit for long sound that stores first sound unit data for long sound corresponding to long sound reading information;
A long sound detecting means for detecting a long sound from the reading information generated by the reading information generating means;
An accent determining means for determining whether or not the long sound detected by the long sound detecting means includes an accent;
When the accent determination means determines that there is an accent, the voice unit data to which the first sound unit data for long sound stored in the first sound unit data storage means for long sound is applied instead of the normal sound unit data Change means,
A speech synthesizer characterized by comprising: Second sound unit data storage means for long sound storing second sound unit data for long sound corresponding to reading information of a long sound accompanied by a pause immediately after,
Pause determination means for determining whether or not there is a pause in the reading information immediately after the long sound detected by the long sound detection means,
The sound unit data changing means, when the pause determining means determines that there is a pause, the second sound for long sound stored in the second sound unit data storing means for long sound instead of the normal sound unit data. The unit data is applied, and when the pause determination unit determines that there is no pause, the first sound unit data for long sound stored in the first sound unit data storage unit for long sound is applied. Item 2. The speech synthesizer according to Item 1. A reading information generation step of analyzing input text using a dictionary and generating accented reading information;
A speech synthesis step of outputting synthesized speech by applying the normal speech unit data stored in the normal speech unit data storage means to the reading information generated in the reading information generation step so as to correspond to the text reading information. In a speech synthesis program to be executed by a computer,
A long sound detecting step for detecting a long sound from the reading information generated in the reading information generating step;
An accent determination step for determining whether or not the long sound detected in the long sound detection step includes an accent;
If it is determined that there is an accent in the accent determination step, the first long sound first data stored in the long sound first sound unit data storage means so as to correspond to the long sound reading information instead of the normal sound unit data. A speech synthesis program for causing a computer to further execute a speech unit data changing step for applying speech unit data. Causing the computer to further execute a pause determination step for determining whether or not there is a pause in the reading information immediately after the long sound detected in the long sound detection step;
In the sound unit data changing step, when it is determined that there is a pause in the pause determining step, the second long sound second information is used so as to correspond to the long sound reading information immediately following the pause instead of the normal sound unit data. The second sound unit data for long sound stored in the sound unit data storage means is applied, and when it is determined that there is no pause in the pause determination step, the first sound unit data for long sound is applied. The speech synthesis program according to claim 3.

Images