JP2002189489A - Speech synthesizer - Google Patents

Abstract

PROBLEM TO BE SOLVED: To achieve cost reduction by eliminating rhythm processing, shortening processing time and suppressing the increase of a memory capacity. SOLUTION: A phonemic sequence output part 2 outputs the phonemic sequence of phonemic symbols identical with the phonemic sequence in which input text information is decomposed and which is stored in the speech database 1 to a phonemic sequence waveform selection part 3. The phonemic sequence waveform part 3 extracts a candidate phonemic sequence from the speech database 1 while successively paying attention to each phonemic sequence inputted from the phonemic sequence output part 2, and selects and outputs the optimal phonemic sequence. A fundamental frequency pattern generating means 4 divides the input text information into a plurality of phonemic sequences having one accent, and generates a fundamental frequency pattern about this. A selection means 5 narrows the number of candidate phonemic sequences by comparing the fundamental frequency of the waveform data of each candidate phonemic sequence with the value of the fundamental frequency pattern, discriminating whether the difference between the two is in a prescribed range, and discriminating further whether each candidate phonemic sequence satisfies a prescribed reference.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a speech synthesizing apparatus for outputting speech based on input of text information.

[0002]

2. Description of the Related Art In a speech synthesizer, usually, input text information for speech synthesis is decomposed into phoneme strings having phoneme symbols equal to the phoneme strings stored in a speech database. The optimal waveform data is extracted from the audio database, and the extracted waveform data is connected to perform audio output of the text information.

[0003]

However, recent speech synthesizers are required to output natural and smooth speech as close as possible to human speech. In order to perform such natural and smooth audio output, it is necessary to generate optimal waveform data by performing prosody processing and the like, but the prosody processing is a complicated process and requires a considerable amount of processing time. . Therefore, in the case of a speech synthesizer which is required to output input text information by voice in real time, there is a drawback that if the input text becomes longer, voice cannot be output continuously.

Further, in order to perform prosody processing, a large memory capacity for storing prosody information must be secured, which has been one factor that hinders cost reduction.

The present invention has been made in view of the above circumstances, and has as its object to provide a speech synthesizer capable of obtaining a natural and smooth speech output without performing prosody processing.

[0006]

As means for solving the above-mentioned problems, the invention according to claim 1 is a sound database storing waveform data and characteristic data of a phoneme sequence used for speech synthesis, A phoneme sequence output unit that inputs text information for speech synthesis, decomposes the text information, and sequentially outputs phoneme sequences of phoneme symbols equal to the phoneme sequence stored in the speech database; and After inputting each phoneme sequence, a target phoneme sequence for which waveform data is to be generated is sequentially determined with respect to the input phoneme sequence, and when a certain target phoneme sequence is determined, the target phoneme sequence is equal to the target phoneme sequence. A phoneme string having a phoneme symbol is set as a candidate phoneme string, and the waveform data and feature data of the candidate phoneme string are extracted from the speech database, and any one of the extracted waveform data is extracted. Is selected as the waveform data of the optimal phoneme sequence for this phoneme sequence of interest, a phoneme sequence waveform selection unit that sequentially outputs the optimal waveform data for each phoneme sequence of interest, and each phoneme from the phoneme sequence waveform selection unit. A speech data synthesizing apparatus comprising: a waveform data connection unit that connects the waveform data of the columns; and a speech output unit that performs speech output of the text information based on input of the connection waveform data from the waveform data connection unit. The phoneme string waveform selection unit divides the text information into units of phoneme strings having only one accent based on the morphological analysis of the text information, and creates a fundamental frequency pattern for each of the divided phoneme strings. Pattern creating means, and waveform data included in the feature data of each candidate phoneme string for the determined one phoneme string of interest. The fundamental frequency of, the accent for this one target phoneme string 1
The phoneme sequence having only one phoneme sequence is compared with the value of the fundamental frequency at a predetermined position of the basic frequency pattern, and the difference between the two is within a predetermined range and the waveform data of the candidate phoneme sequence that satisfies a predetermined criterion is converted to the target phoneme. Selecting means for selecting the optimal waveform data for the column and sequentially outputting the optimal waveform data for each phoneme sequence of interest corresponding to the input order of the text information.

According to a second aspect of the present invention, in the first aspect of the invention, the selecting means includes a plurality of waveform data of a candidate phoneme string having a difference between the two within a predetermined range and satisfying a predetermined standard. In this case, the formant mean square error between the waveform data already selected as the optimal waveform data by the previous selection operation and the waveform data of each candidate phoneme string is obtained, and the one with the smallest value is determined as the optimal waveform. Is selected as data.

According to a third aspect of the present invention, in the first or second aspect of the invention, the selecting means determines whether or not the predetermined criterion is satisfied. A determination is made as to whether or not a phoneme sequence before and after a column matches a phoneme sequence before and after each of the candidate phoneme sequences stored in the speech database. After the determination, it is determined whether or not the difference between the two values of the fundamental frequency is within a predetermined range.

According to a fourth aspect of the present invention, in the first or second aspect of the invention, the selecting means determines whether or not the predetermined criterion is satisfied before and after the target phoneme sequence in the text information. Is compared with the phoneme strings before and after each of the candidate phoneme strings stored in the voice database, and the phoneme strings that match in the pre-phoneme string of the target phoneme string and the pre-phoneme string of each of the candidate phoneme strings are compared. The number of symbols is determined for each pre-phoneme sequence of each candidate phoneme sequence, and the number of phoneme symbols that match in the post-phoneme sequence of the target phoneme sequence and the post-phoneme sequence of each candidate phoneme sequence is calculated for each of the candidate phoneme sequences. Is calculated for each of the following phoneme strings, and for each candidate phoneme string, the sum of the number of matching phoneme symbols in the preceding phoneme string and the number of matching phoneme symbols in the subsequent phoneme string is calculated. The candidate phoneme sequence is
After determining whether or not the sum is a candidate phoneme string having the maximum value, and after determining whether or not the sum is a candidate phoneme string having a maximum value, the fundamental frequency Is determined whether or not the difference between the two values is within a predetermined range.

According to a fifth aspect of the present invention, in the first or second aspect of the invention, the selecting means determines whether or not the predetermined criterion is satisfied before and after the target phoneme sequence in the text information. Is compared with the phoneme strings before and after each of the candidate phoneme strings stored in the voice database, and the phoneme strings that match in the pre-phoneme string of the target phoneme string and the pre-phoneme string of each of the candidate phoneme strings are compared. The number of symbols is determined for each pre-phoneme sequence of each candidate phoneme sequence, and the number of phoneme symbols that match in the post-phoneme sequence of the target phoneme sequence and the post-phoneme sequence of each candidate phoneme sequence is calculated for each of the candidate phoneme sequences. For each candidate phoneme string, an evaluation value relating to the balance between the number of matching phoneme symbols in the preceding phoneme string and the number of matching phoneme symbols in the subsequent phoneme string is determined in advance for each candidate phoneme string. To the set rules Calculated Zui, one candidate phoneme strings,
A determination is made as to whether or not the balance is a candidate phoneme string with the best evaluation value, and further, a determination is made as to whether or not the balance is a candidate phoneme string with the best evaluation value. And determining whether the difference between the two values of the fundamental frequency is within a predetermined range.

According to a sixth aspect of the present invention, in the first aspect of the present invention, the selecting means determines whether or not the predetermined criterion is satisfied by a previous selecting operation. Whether the difference between the average amplitude value over a predetermined section near the end of the waveform data determined as the waveform data of the above and the average amplitude value over a predetermined section near the front end of the waveform data of each candidate phoneme string is within a predetermined range. The determination of whether the difference between the amplitude average values is equal to or less than the predetermined value after the determination of the preceding and following phoneme sequences is performed. This is performed after or before the determination as to whether or not it is within the range is performed.

According to a seventh aspect of the present invention, in the invention according to the sixth aspect, the average amplitude value of the determined waveform data over a predetermined section near the end portion becomes larger as approaching the end portion. The average amplitude value over a predetermined section near the tip in the waveform data of each of the candidate phoneme strings is such that the amplitude value increases as approaching the tip. Calculated based on waveform data weighted to

[0013]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS An embodiment of the present invention will be described below with reference to the drawings. FIG. 1 is a block diagram illustrating the configuration of the speech synthesis device according to the first embodiment. In this figure,
The speech database 1 stores waveform data and characteristic data (data such as a fundamental frequency, a phoneme symbol, and a formant) of a phoneme string used for speech synthesis.

Then, text information is output to a phoneme string output unit 2 and a phoneme string waveform selection unit 3 from text information input means (not shown). The phoneme string output unit 2 decomposes the input text information and outputs a phoneme string of a phoneme symbol equal to the phoneme string stored in the speech database 1 to the phoneme string waveform selection unit 3.

After inputting the phoneme sequence from the phoneme sequence output unit 2, the phoneme sequence waveform selecting unit 3 outputs a phoneme sequence for which waveform data is to be generated. The sequence is referred to as a “phonemic sequence of interest” in this specification.)
Are sequentially determined. When the phoneme string waveform selecting unit 3 determines a certain phoneme string of interest, a candidate phoneme string for this noteworthy phoneme string is stored in the speech database 1 (stored in the speech database 1 and the phoneme symbol of interest is equal to the phoneme symbol). A phoneme sequence is extracted, and one of the extracted candidate phoneme sequences is selected and output as waveform data of an optimal phoneme sequence for the target phoneme sequence.

The waveform data connection unit 6 connects the respective waveform data input from the phoneme sequence waveform selection unit 3 and outputs them to the audio output unit 7. And
The audio output unit 7 outputs audio of text information based on the connection waveform data input from the waveform data connection unit 6.

The phoneme string waveform selecting section 3 has a fundamental frequency pattern creating means 4 and a selecting means 5. The fundamental frequency pattern creating means 4 inputs text information from text information input means (not shown), performs morphological analysis on the text information, divides the text information into a plurality of phoneme strings having one accent, A basic frequency pattern is created for the divided phoneme string of one accent.

The selection means 5 compares the fundamental frequency of the waveform data of each candidate phoneme string extracted from the speech database 1 with the value of the created fundamental frequency pattern, and the difference between the two is within a predetermined range. By determining whether or not each candidate phoneme string satisfies a predetermined criterion, the number of candidate phoneme strings is reduced. In this case, it is determined whether or not the predetermined criterion is satisfied. For the phoneme symbols (the whole phoneme symbol, the first phoneme symbol, and the last phoneme symbol), the phoneme strings before and after the phoneme string of interest in the text information and the speech database There is a determination as to whether or not the phoneme strings before and after each candidate phoneme string stored in No. 1 match. By determining such a fundamental frequency, it is possible to select a phoneme string as close as possible to a fundamental frequency pattern preset as a preferred fundamental frequency in performing natural speech synthesis. Further, by determining the phoneme symbols of the phoneme strings before and after the phoneme string of interest, it is possible to select a phoneme string whose state before and after is as similar as possible to the phoneme string of interest.

If the selection means 5 has not been able to narrow down the candidate phoneme sequence to one even after performing the above determination, the selecting means 5 sets the formant average between the previously selected previous waveform data and each candidate waveform data. The square error is calculated, and the one with the smallest value is selected as the optimal waveform data. The order of selection by the selection means 5, that is, the order of focusing on the phoneme sequence of the text information is the input order of the text information. As described above, the reason for selecting the one with the smallest formant change is that the formant changes according to the movement of the tongue, jaw, lips, etc. when a human speaks, and the human tongue, jaw, lips, etc. Is continuous, and it is considered that it is preferable that the formants present on the time axis be continuous in order to perform natural speech synthesis.

Next, the operation of the first embodiment configured as described above will be described. FIGS. 2 to 5 are flowcharts for explaining the operation of the phoneme string waveform selection unit 3. The basic frequency pattern creation means 4 of the phoneme string waveform selection unit 3 receives text information (for example, “Today is fine.”) From text information input means (not shown). Is subjected to morphological analysis to divide into a plurality of phoneme strings having one accent (step 1). In this embodiment, a plurality of phoneme strings having one accent are “kyo-wa”, “har”
The case of three phoneme strings "e" and "desune" will be described. Then, the fundamental frequency pattern creating means 4 outputs these three phoneme strings “kyo-wa”, “hare”, “desune”
Is created (step 2). For example, since "kyo-wa" has an accent at the position of "o", its fundamental frequency pattern is a mountain-shaped pattern having a peak at the position of "o".

On the other hand, the phoneme string output unit 2 also inputs text information “Today is fine.” From text information input means (not shown), decomposes the text information, and stores it in the speech database 1. A phoneme string having a phoneme symbol equal to the phoneme string present is output to the phoneme string waveform selection unit 3. In this case, the phoneme string output by the phoneme string output unit 2 does not always match the phoneme strings “kyo-wa”, “hare”, and “desune” decomposed by the fundamental frequency pattern creating unit 4 based on morphological analysis. For example, "kyo-", "wa", "hare", "desu",
"Ne". However, in this case, the phoneme string output unit 2 also outputs “kyo-
It is assumed that three phoneme strings “wa”, “hare”, and “desune” are output. The selection means 5 of the phoneme string waveform selection unit 3
Each phoneme string is input from the phoneme string output unit 2 (step 3).

Next, the selection means 5 first determines "kyo-wa" as the phoneme sequence of interest (step 4). here,
The relationship between the phoneme sequence of interest and the candidate phoneme sequence will be specifically described. The candidate phoneme sequence is a phoneme sequence in text information, and is a candidate for a phoneme sequence that is currently focused on in order to generate speech waveform data from it (that is, a target phoneme sequence), and is a phoneme sequence in the text information. Even in the case of a phoneme string that is not currently focused on, it is not a candidate phoneme string. That is, assuming that the decomposed phoneme strings of the text information "Today is fine." Are "kyo-wa", "hare", and "desune", the speech database 1 contains a plurality of phoneme symbols "kyo-wa". Waveform data (although there may be one), a plurality of waveform data of a phoneme symbol "hare",
There are a plurality of waveform data of the phoneme symbol “desune”. In the case where the waveform data of “kyo-wa” is currently to be generated with the target phoneme sequence “kyo-wa” among the three phoneme sequences “kyo-wa”, “hare”, and “desune”, Are only candidate phoneme strings, and a plurality of phoneme strings of a phoneme symbol “hare” and a plurality of phoneme strings of a phoneme symbol “desune” are candidate phoneme strings. Not a column. Similarly, "ha
When the waveform data is to be generated using “re” as the target phoneme sequence, only a plurality of phoneme sequences of the phoneme symbol “hare” in the voice database 1 are candidate phoneme sequences,
A plurality of phoneme strings of the phoneme symbol "kyo-wa" and "desu
The plurality of phoneme strings of the phoneme symbol “ne” are not candidate phoneme strings.

Therefore, the decomposed phoneme strings "kyo-wa", "hare", "desun" of the text information "Today is fine."
When creating each waveform data of "e", first determine this as a phoneme sequence of interest to pay attention to "kyo-wa",
The phoneme string having the same phoneme symbol as “kyo-wa” is treated as a candidate phoneme string, and its waveform data is extracted from the speech database 1. In this case, when there are a plurality of waveform data of the candidate phoneme sequence, the optimum waveform data must be selected from the plurality of data. However, in order to determine which waveform data is optimal, one candidate phoneme sequence is required. Until it is narrowed down, data on fundamental frequencies, phoneme symbols, and formants, which will be described later, are checked. After selecting the optimal waveform data for "kyo-wa" in this way, next, "hare" is determined as the phoneme sequence of interest, and the waveform data of the candidate phoneme sequence "hare" of the same phoneme symbol is determined. From the audio database 1 and the optimum waveform data is selected from among them. Similarly, after selecting the optimal waveform data for "hare", next, "desune" is determined as the phoneme sequence of interest, and the waveform data of the candidate phoneme sequence "desune" of the same phoneme symbol is stored in the voice database. 1 and select the most suitable waveform data from them. As described above, the phoneme sequence of interest is a phoneme sequence output by the phoneme sequence output unit 2 after decomposing the text information, and serves as an index as to which part of the text information to generate waveform data at present. On the other hand, the candidate phoneme sequence is a concept having a meaning as a name in the voice database 1 of the waveform data to be generated.

Now, returning to step 4, the selecting means 5
After determining “kyo-wa” as the phoneme sequence of interest, all candidate phoneme sequences “kyo-wa” (in this case, for example, the first to third (For example, there are three candidate phoneme strings) (for example, data about fundamental frequency, phoneme symbol, formant, etc.) (step 5). Then, the selection unit 5 indicates the value fzt of the pattern created by the fundamental frequency pattern creation unit 4 (the peak position “o” and the values of the respective basic frequencies at a plurality of positions before and after the fundamental position). .) And the respective values fzt
Is compared with each value fzt1 of the first candidate phoneme string (step 6), and it is determined whether or not the following equation (1) holds (step 7). Note that a and b in the equation (1) are fixed values set in advance.

A ≦ fzt−fzt1 ≦ b (1) The selecting means 5 determines in step 7 that fz
"YES" when expression (1) is satisfied for t and fzt1
If formula (1) is not satisfied at any one of the locations, it is determined as "NO". And selecting means 5
Stores this determination result in its own storage means (step 8), and determines whether or not all the candidate phoneme strings have been completed (step 9). In this case, since the second and third candidate phoneme strings still remain, the selecting means 5 determines "NO", and in step 6, each value fzt of the created pattern and the second candidate Each value of the phoneme sequence is compared with fzt1.

The selecting means 5 makes a determination on the second and third candidate phoneme strings "kyo-wa" in the same manner as described above. After making a determination of "YES" in step 9, the equation (1) is obtained. Is determined whether or not there is a candidate phoneme string satisfying
0). If there is a candidate that satisfies the expression (1), the candidate phoneme sequence that does not satisfy the expression (1) is removed (step 11), and it is determined whether or not the candidate phoneme sequence is reduced to one (step 11). 12). If the number is reduced to one, the candidate phoneme sequence is determined as optimal waveform data (step 13),
In step 56, "NO" is determined as to whether or not the selection has been completed for all the target phoneme strings, and the process returns to step 4 to determine the next target phoneme string ("hare"). In the case of the above example, it is assumed that all three candidate phoneme strings satisfy Expression (1). Therefore, the determination result in step 10 is “Y
ES ”, the number of candidate phoneme strings to be removed in step 11 becomes zero, and the determination result in step 12 becomes“ NO ”. Then, the selecting means 5 determines whether or not the phoneme sequence of interest is the first one (step 14). In step 10, if there is no one satisfying the expression (1), the process immediately proceeds to step 14.

Since the phoneme sequence of interest "kyo-wa" is the first phoneme sequence of the input text information, the selecting means 5 executes step 1
In step 4, "YES" is determined. Therefore, the selecting means 5 proceeds to step 24, and first compares the post-phoneme sequence of the target phoneme sequence "kyo-wa" with the post-phoneme sequence of the first candidate phoneme sequence "kyo-wa" for all the phoneme symbols. Then, it is determined whether or not they match (step 25), and the result of the determination is stored in the storage means (step 26). Then, in a step 27, it is determined whether or not the comparison has been completed for all the candidate phoneme strings, and "NO" is determined. The phoneme sequence is compared with the succeeding phoneme sequence. After completing the comparison determination up to the third candidate phoneme string “kyo-wa” in the same manner, the selection means 5 determines “YES” in step 27 and determines whether there is a match. (Step 28). In the case of the above example, the post-phoneme sequence of the first candidate phoneme sequence “kyo-wa” is “ha”.
jimete ", the post-phoneme sequence of the second candidate phoneme sequence" kyo-wa "is" hima ", and the post-phoneme sequence of the third candidate phoneme sequence" kyo-wa "is" hen ". Therefore, step 28
Is "NO". If the result of the determination in step 28 is "YES", the candidate phoneme strings that do not match in the determination in step 25 are removed (step 29), and it is determined whether or not the number of candidate phoneme strings has been reduced to one. (Step 30). If the number is reduced to one, the candidate phoneme sequence is determined as the optimum waveform data (step 3).
1).

After determining "NO" in step 28, the selecting means 5 determines whether or not the phoneme sequence of interest "kyo-wa" is the first one (step 32). The result of this determination is "Y".
ES ”, the flow proceeds to step. Selection means 5
Is the first one of the phoneme sequence after the phoneme sequence of interest "kyo-wa".
The phoneme symbol is compared with the first phoneme symbol of the succeeding phoneme sequence of the first candidate phoneme sequence "kyo-wa" to determine whether or not the two match (step 43), and the result of the determination is stored. (Step 44). Then, in step 45, it is determined whether or not the comparison has been completed for all the candidate phoneme strings, and “NO” is determined. Then, the processing returns to step 42, where the succeeding phoneme string of the target phoneme string “kyo-wa” and the second candidate The phoneme sequence is compared with the succeeding phoneme sequence. After completing the comparison determination up to the third candidate phoneme string “kyo-wa” in the same manner, the selection unit 5 determines “YES” in step 45 and determines whether there is a match. (Step 46). In the case of the above example, the first phoneme symbol of each of the subsequent phoneme strings “hajimete”, “hima”, and “hen” of the first to third candidate phoneme strings “kyo-wa” is “h”. , Matches the first phoneme symbol “h” of the phoneme sequence “hare” after the target phoneme sequence “kyo-wa”. Therefore, the determination result of step 46 is "YES".

If the result of the determination at step 46 is "YES", the selecting means 5 removes the candidate phoneme string that does not match at step 43 (step 47), and narrows down the candidate phoneme string to one. It is determined whether or not it has been performed (step 48). If the number is reduced to one, the candidate phoneme sequence is determined as the optimal waveform data (step 49). In the case of the above example, step 43 is performed for all three candidate phoneme strings.
Is "YES", the number of candidate phoneme strings to be removed in step 47 is zero, and the determination result in step 48 is "NO".

Next, the selecting means 5 selects the phoneme sequence of interest "kyo-
It is determined whether or not "wa" is the first phoneme string (step 5).
0), the formant mean square error M between the waveform data determined as the optimal waveform data by the previous selection operation and the waveform data of the first candidate phoneme sequence of the current phoneme sequence.
Calculate se (step 52). In this case, since the phoneme sequence of interest "kyo-wa" is the first phoneme sequence, the result of the determination in step 50 is "YES", and the selection means 5 reads the preset formant value from, for example, its own storage means. get. This is because there is no determined previous waveform data for the first phoneme sequence yet, and it is necessary to set data corresponding to the previous waveform data in advance.

Therefore, the selection means 5 determines in step 51
After obtaining the value of the set formant in step 52, the mean square error Mse is calculated by the following equation (2) using the value of the set formant in step 52 (step 52). In the present embodiment, the values of each formant stored in the audio database 1 are assumed to be three types of values for the first to third formants, and have been determined as optimal waveform data by the previous selection operation. Are represented as f11, f12 and f13, and the first to third formant values of the waveform data of one candidate phoneme sequence of the current phoneme sequence are represented as f21, f22 and f23. . Since the current phoneme sequence is the first phoneme sequence, f11, f12, f1
Each formant value of 3 becomes each value of the set formant.

[0032]

(Equation 1) Then, the selecting means 5 stores the determination result of step 52 in its own storage means (step 53), and determines whether or not all the candidate phoneme strings have been completed (step 54). In this case, since the second and third candidate phoneme strings still remain, the selecting means 5 determines "NO" and returns to step 52 to return to the determined waveform data and the waveform data of the second candidate phoneme string. Is calculated from the formant mean square error Mse.

The selecting means 5 performs the operation of step 52 on the second and third candidate phoneme strings "kyo-wa" in the same manner, and after the determination of "YES" in step 54,
A candidate phoneme string having the smallest formant mean square error Mse is selected as optimal waveform data (step 55).
In the above case, for example, the second candidate phoneme string “kyo-wa” is selected as the optimum phoneme string. Then, the selection means 5 determines whether or not the selection has been completed for all the phoneme strings of interest. At this time, the selection means 5 still has “hare” and “desun”.
Since "e" remains, "NO" is determined.

Therefore, the selection means 5 returns to step 4 and determines "hare" as the next phoneme sequence of interest. The subsequent operations of Steps 5 to 13 are the same as those described above, and thus redundant description will be omitted. Now, Step 12
Is "NO", the phoneme sequence of interest "hare" is not the first phoneme sequence, so the selecting means 5 determines "NO" in step 14 and proceeds to step 15. Steps 15 to 22 correspond to steps 24 to 31 described above.
The description is omitted because it is substantially the same as (1). The only difference is that the phoneme sequence to be compared is the front phoneme sequence or the rear phoneme sequence. If the determination result in step 21 is “NO”, the selecting means 5 determines in step 23 whether or not the phoneme sequence of interest is the last phoneme sequence.
e ”is not the last phoneme sequence, so the result of the determination is“ N
O ", and the process proceeds to step 24. Step 24-
Since 31 has already been described, its description is omitted.

If the result of the determination in step 30 is "NO", since the phoneme sequence of interest "hare" is not the first phoneme sequence, the selecting means 5 determines "NO" in step 32 and proceeds to step 33. I do. Steps 33 to 40 are substantially the same as steps 42 to 49 described above, and a description thereof will be omitted (the difference is that the comparison target is the last one phoneme symbol of the previous phoneme sequence or the first one phoneme symbol of the rear phoneme sequence). Only the point.) If the result of the determination in step 39 is “NO”, the selecting means 5 determines in step 41 whether or not the phoneme sequence of interest is the last phoneme sequence.
e ”is not the last phoneme sequence, so the result of the determination is“ N
O ", and the routine goes to Step 42. Step 42 ~
49 has already been described, and a description thereof will be omitted.

If the result of the determination in step 48 is "NO", the selecting means 5 determines whether or not the phoneme sequence of interest "hare" is the first phoneme sequence.
O ", the operation of step 52 is immediately performed. In this case, the “determined waveform data” is the second
Since the waveform data for the candidate phoneme string “kyo-wa” has been selected, this waveform data and the first candidate phoneme string “ha
The formant mean square error Mse of “re” is calculated.

Thereafter, the selecting means 5 determines "YES" in step 54, and then determines the formant mean square error M
The candidate phoneme sequence having the smallest value of se is selected as the optimal waveform data (step 55), and it is determined whether or not the selection has been completed for all the target phoneme sequences (step 56). At this point, "desune" still remains, so "NO"
Is determined.

Therefore, the selection means 5 returns to step 4 and determines "desune" as the next phoneme sequence of interest. The description of subsequent steps 5 to 13 will be omitted. Step 1
If the determination result of step 2 is “NO”, the phoneme sequence of interest “desune” is not the first phoneme sequence, so the selecting unit 5 determines “NO” in step 14 and proceeds to step 15. The description of steps 15 to 22 is omitted.

If the result of the determination in step 21 is "NO", the selecting means 5 determines in step 23 whether or not the phoneme sequence of interest is the last phoneme sequence. Since the phoneme sequence is "", the determination result is "YES", and the routine goes to Step 32.

Since the phoneme sequence of interest "desune" is not the first phoneme sequence, the selecting means 5 determines "NO" in step 32 and proceeds to step 33. The description of steps 33 to 40 is omitted. If the result of the determination in step 39 is "NO", the selecting means 5 determines in step 41 whether or not the phoneme sequence of interest is the last phoneme sequence. Therefore, the determination result is “YES”, and the routine goes to Step 50.

The operations in steps 50 to 55 are the same as those described above, and thus the description thereof will be omitted. Then, the selecting means 5 determines "YES" in step 56, and ends all the operations.

In order to make the description easier to understand, the phoneme string output unit 2 firstly outputs the fundamental frequency pattern
Output the same three phoneme strings "kyo-wa", "hare", and "desune" that were decomposed based on morphological analysis, but the phoneme string output unit 2 outputs a different phoneme string. The same operation may be performed when “kyo-”, “wa”, “hare”, “desu”, and “ne” are output (in fact, this case is more normal). For example, the first phoneme sequence of interest "kyo-"
In the case of, the value of the fundamental frequency of each candidate phoneme string “kyo-” extracted from the voice database 1 and the value of the fundamental frequency at the location corresponding to “kyo-” in the pattern for creating “kyo-wa” are described above. What is necessary is just to apply to Formula (1).

The phoneme string waveform selecting section 3 operates as described above, selects the optimum phoneme string waveform data from the speech database 1, takes it out, and outputs it to the waveform data connection section 6. The waveform data connection unit 6 connects the input waveform data and outputs it to the audio output unit 7. The audio output unit 7 outputs the text information “Today is fine” based on the input connection waveform data. Sound output.

In the speech synthesizer of FIG. 1 described above, the phoneme string waveform selecting section 3 selects an optimum phoneme string waveform using only data such as the fundamental frequency, phoneme symbols, and formants stored in the speech database 1. It is supposed to.
Therefore, it is not necessary to perform complicated prosody processing, so that the processing time can be shortened, and the memory capacity can be suppressed from increasing, thereby achieving a sufficient cost reduction.

Next, a second embodiment of the present invention will be described. The configuration of the second embodiment is the same as that of the first embodiment, that is, the configuration shown in FIG. 1, and the order of the discriminating operation when the selecting means 5 selects a candidate phoneme string from the voice database 1 is different. It is. That is, in the first embodiment, as shown in the flowcharts of FIGS. 2 to 5, after first comparing the fundamental frequencies, comparing the preceding and succeeding phoneme strings, and then calculating the formant error. Although the optimum waveform data is selected, in the second embodiment, the comparison of the preceding and succeeding phoneme strings is performed first, the fundamental frequency is compared, and then the error of the formant is calculated. ing.

As described above, by first comparing the preceding and succeeding phoneme strings, first, it is possible to optimize the articulation at the time of speech synthesis of the text. Here, “articulation” refers to an intermediate waveform data portion that does not belong to any of the waveforms of two consecutive phonemes when uttered for two consecutive phonemes. For example, consider the case of uttering “aiueo” consecutively.
There is intermediate waveform data that does not belong to "A" or "I", and similarly, "I"
There is intermediate waveform data that does not belong to "c" or "c". Therefore, if the input text information is "aiueo"
In the case where “A” is selected from a plurality of candidate phoneme strings stored in the voice database, a candidate phoneme string in which “A” and “U” are stored before and after “A”, respectively, is selected. Is most preferable. Therefore, in the second embodiment, the highest priority is given to the comparison of the preceding and succeeding phoneme strings, and this comparison is performed first.

Second, by comparing the preceding and succeeding phoneme strings first as described above, the speaker's natural (“natural” here) related to the waveform data stored in the voice database is determined. (For example, if the speaker has an accent, it is not necessarily natural for a third party, but "natural" for the speaker.)
Intonation can often be extracted.

Next, the operation of the second embodiment will be described. FIGS. 6 to 10 are flowcharts showing the operation of the selecting means 5 in the second embodiment.
In these figures, steps having the same contents as the respective steps in FIGS. 2 to 5 are denoted by a suffix “a”, and the connection signs “A”, “ B ”,…,“ G ”include“ 1 ”,“ 1 ”
Subscripts such as "1" and "12" are added. These flowcharts in FIGS. 6 to 10 differ from the flowcharts in FIGS. 2 to 5 only in the order of some steps, and the contents of each step have been described in FIGS. 2 to 5. Therefore, detailed description of individual steps in the flowcharts of FIGS. 6 to 10 will be omitted, and only outlines will be briefly described.

First, when the phoneme string waveform selecting unit 3 inputs text information from text information input means (not shown), the fundamental frequency pattern creating means 4 performs a morphological analysis on this text information to obtain one text. It is divided into a plurality of phoneme strings having accents (step 1a), and a fundamental frequency pattern of each phoneme string is created (step 2a). On the other hand, the phoneme string output unit 2 also inputs the text information, decomposes the text information, and outputs a phoneme string of a phoneme symbol equal to the phoneme string stored in the speech database 1 to the phoneme string waveform selection unit 3. The selecting means 5 inputs each phoneme string from the phoneme string output unit 2 (step 3a). Then, the selecting means 5 determines the phoneme sequence of interest (step 4a), and acquires data on all the candidate phoneme sequences from the voice database 1 (step 5a). Up to here, the processing order is the same as that of the first embodiment.

Next, the selection means 5 determines whether or not the phoneme sequence of interest is the first one in the text (step 1).
4a) If it is the head, the processing of steps 24a to 32a or the processing of steps 42a to 49a is further performed.
That is, the number of candidate phoneme strings is reduced by comparing the phoneme string after the target phoneme string with the phoneme string after the candidate phoneme string for all of the phoneme symbols. Then, the number of candidate phoneme strings is narrowed down by comparing the first one phoneme symbol of the last phoneme string of the candidate phoneme string with the first one phoneme symbol of the latter phoneme string of the candidate phoneme string.

On the other hand, if it is determined in step 14a that the phoneme string of interest is not the first one, the selection means 5 performs the processing in steps 15a to 23a, or further performs steps 33a to 33a.
41a is performed. That is, for all phoneme symbols, the number of candidate phoneme strings is reduced by comparing the previous phoneme string of the phoneme string of interest with the previous phoneme string of the candidate phoneme string. The last one phoneme symbol of the previous phoneme sequence of the last and the last one of the previous phoneme sequence of the candidate phoneme sequence
By comparing with phoneme symbols, the number of candidate phoneme strings is narrowed down.

Thereafter, the selecting means 5 performs a process of narrowing down the number of candidate phoneme strings by comparing the value of the pattern with the value of the candidate phoneme string for the fundamental frequency (step 6a).
~ 13a) If you still can't narrow down to one candidate,
A formant mean square error with the previously determined waveform data is calculated, and a candidate phoneme string having the smallest formant mean square error is selected (step 50).
a-56a).

As described above, in the second embodiment, first, the selecting means 5 determines the phoneme sequence of the phoneme symbol before and after the phoneme sequence of interest in the text information and the candidate phoneme sequences stored in the speech database. Since it is determined whether or not the preceding and succeeding phoneme strings match, it is possible to optimize the articulation at the time of text-to-speech synthesis of the text. Intonation can often be extracted.

Next, a third embodiment of the present invention will be described. This third embodiment has the same configuration as that of the first embodiment shown in FIG.
In addition, a calculation on the amplitude (volume) of the waveform data is added as a discriminating operation when selecting a candidate phoneme string from. That is, when selecting one of a plurality of candidate phoneme strings, the selecting means 5 in this embodiment sets the amplitude of the leading end of the waveform data to the amplitude of the terminal end of the previously selected previous waveform data. I try to select something that is as close as possible. This makes it possible to perform natural and smooth speech synthesis.

FIG. 11 is an explanatory diagram of processing performed when obtaining the amplitude of waveform data in the present embodiment.
FIG. 11A shows the waveform data of the previously selected previous waveform data or the candidate phoneme string stored in the voice database 1. First, the selection means 5 is configured as shown in FIG.
Absolute value filter processing is performed on the waveform data shown in
By folding the waveform portion on the negative side to the positive side, FIG.
As shown in (b), the waveform data is formed only by the positive part. Next, the selecting means 5 performs a maximum value filtering process on the waveform data shown in FIG. 11B using a window having a predetermined small window width to obtain an envelope as shown in FIG. 11C. To The reason for obtaining such an envelope is to obtain the average value of the amplitude of the waveform data as accurately as possible. In other words, the amplitude means the height of the peak point of the waveform data.
If the average value of the amplitude is calculated for the waveform data of (b), the height of the valley portion of the waveform data is included, so that the average value of the amplitude becomes inaccurate. is there.

FIG. 12 is a diagram for explaining a method of obtaining the average amplitude value of the waveform data in the present embodiment. The selecting unit 5 performs the processes described with reference to FIG. 11 to first obtain the envelope DA of the waveform data determined by the previous selection.
To get Thereafter, the selection means 5 obtains the envelope DB of the waveform data of the target candidate phoneme string.

The waveform data of the candidate phoneme string to be connected to the previously determined waveform data includes the tip hB of the envelope DB.
Is preferably as close as possible to the amplitude of the terminal end eA of the envelope DA. However, if the determination is made only based on the amplitude of the tip hB, there is a possibility that data whose amplitude changes abruptly after the tip hB may be selected. Therefore, in the present embodiment, the amplitude average values of both the section I including the end portion eA and the section J including the front end portion hB are compared.

In FIG. 12, i and j indicate sampling positions of each waveform data, and x (i) and y (j) indicate amplitude values at each sampling position. Now, the envelope DA
, The position of the end part eA is set to i = 100, the position of the front end part hB of the envelope DB is set to j = 1, the section I is i = 96 to 100, and the section J
Is j = 1 to 5. Then, the amplitude value in section I is x (9
6) to x (100), and the amplitude value in section J is y (1) to y
(5). The selection means 5 calculates these amplitude values x (96) to x
An average value of (100) and y (1) to y (5) is obtained, and it is determined whether or not the difference is equal to or less than a certain value. However, from the viewpoint of selecting an optimal one from a plurality of candidate phoneme strings, it is not appropriate to treat all of these amplitude value data equally.
This is because, in order to determine whether or not the amplitude near the end portion eA and the amplitude near the front end portion hB are similar, the end portion e
This is because data closer to A and the front end hB has a greater value, and data farther from the terminal end eA and the front end hB has a smaller value. Therefore, in the present embodiment, the largest weighting coefficient Wx is assigned to the amplitude value x (100) in the envelope DA.
(100), and the amplitude value x (96) has the smallest weight coefficient Wx
(96). Similarly, the largest weight coefficient Wy (1) is given to the amplitude value y (1) in the envelope DB, and the smallest weight coefficient Wy (5) is given to the amplitude value y (5).

The selecting means 5 calculates the average amplitude value Zx in the section I of the envelope DA and the average amplitude value Zy in the section J of the envelope DB using the following equations (3) and (4). FIG.
In the example shown in FIG. 2, i in Expression (3) is i = 96 to 100, and j in Expression (4) is j = 1 to 5. Next, the selection means 5 obtains the absolute value α of the difference between the average amplitude values Zx and Zy, and determines whether this α is smaller than the preset value α0, that is, whether the equation (5) is satisfied. I do. And
The candidate phoneme sequence for which the expression (5) does not hold is removed, and (5)
Leave the candidate phoneme string for which the expression has been established as it is.

[0060]

(Equation 2) In the third embodiment, as described above, the amplitude average value over a predetermined section near the end of the waveform data that has been determined as the optimal waveform data by the previous selection operation, and the vicinity of the leading end in the waveform data of each candidate phoneme string In this embodiment, the determination as to whether or not the difference from the amplitude average value over a predetermined section is within a predetermined range is added to the second embodiment. In this case, the difference between the amplitude average values is determined as follows:
The determination may be made after or before determining whether or not the difference between the fundamental frequencies is within a predetermined range (steps 6a to 12a).
a-48a). As described above, the discrimination between the front and rear phoneme strings is a discrimination that is important in optimizing articulation and giving priority to extracting the intonation of the speaker in the speech database.

FIGS. 13 and 14 are partial flowcharts showing the operation of the selecting means 5 in the third embodiment. 6 to 8 and 10, which are flowcharts according to the second embodiment, are also applied to the third embodiment. The outline of the operation of the third embodiment will be described with reference to these drawings.

First, similarly to the second embodiment, the processing of steps 1a to 5a and step 14a shown in FIG. 6 is performed, and the phoneme symbols are stored in the phoneme strings before and after the phoneme string of interest and the speech database. The number of candidates is narrowed down based on the determination as to whether the stored phoneme strings before and after each of the candidate phoneme strings match (steps 15a to 49a shown in FIGS. 7 and 8). When the number of candidates cannot be narrowed down to one by these processes, the selection unit 5 narrows down the number of candidates based on the comparison of the fundamental frequencies, similarly to the second embodiment (FIG. 13). Steps 6a to 13a). The difference between FIG. 9 which is the flowchart of the second embodiment and FIG. 13 which is the flowchart of the third embodiment is only that the concatenated code under step 12a is changed from F12 to F13. .

Next, when the number of candidates cannot be narrowed down to one even when the fundamental frequencies are compared (when the determination in step 12a is "NO"), the selecting means 5 sets the number in FIG.
As shown in (1), processing for narrowing down the number of candidates based on the determination of the amplitude of the waveform data is performed.

That is, the selecting means 5 first calculates the average amplitude value Zx near the end of the previously determined waveform data by the equation (3), and also calculates the waveform data of the currently targeted candidate phoneme string. Average amplitude Zy near the tip
Is calculated by equation (4) (steps 57 and 58). Then, it is determined whether or not the absolute value of the difference between these amplitude average values is smaller than the set value α0, that is, whether or not the equation (5) is satisfied (step 59). It is stored in the storage means (step 60).

Thereafter, the selecting means 5 determines whether or not the processing of steps 57 to 60 has been completed for all the candidate phoneme strings (step 61). And the above processing is repeated. Also,
If the above processing has been completed for all the candidate phoneme strings, it is determined whether or not there is a candidate phoneme string that satisfies the expression (5) (step 62). ) The candidate phoneme strings that do not satisfy the expression are removed (step 63). However, if all the candidate phoneme strings satisfy Expression (5), the number of removals is zero. Then, it is determined whether or not the candidate phoneme string has been narrowed down to one (step 64), and if it has been narrowed down to one, it is determined to select this as the optimal waveform data (step 65).

On the other hand, if the determination in step 62 or step 64 is "NO", the selection means 5 performs the processing of steps 50a to 56a shown in FIG. 10 as in the second embodiment. In the third embodiment, the candidate number narrowing process based on the determination of the amplitude of the waveform data (steps 57 to 65) is performed after the candidate number narrowing process based on the comparison of the fundamental frequencies (steps 6a to 13a). However, it may be performed before. However, in any case, it must be performed after the candidate number narrowing process (steps 15a to 48a) based on the comparison of the front and rear phoneme strings.

As described above, in the third embodiment, the amplitude average value over a predetermined section near the end of the waveform data determined as the optimum waveform data by the previous selection operation, and the waveform data of each candidate phoneme sequence In the second embodiment, the processing for determining whether or not the difference from the amplitude average value over a predetermined section near the front end is within a predetermined range is added to the second embodiment. It is possible to perform a more natural and smooth speech synthesis than in the case.

Next, a fourth embodiment of the present invention will be described. This fourth embodiment also has the same configuration as that of the first embodiment shown in FIG.
This is a further improvement of the second embodiment with respect to the content of the discriminating operation when selecting a candidate phoneme string from. In the second embodiment, the operation for narrowing down candidates based on a comparison between a phoneme sequence before and after a phoneme sequence of interest and a phoneme sequence before and after a candidate phoneme sequence is shown in FIGS. 7 and 8. Is summarized and described again (however, a case in which the phoneme sequence of interest is an intermediate one which is neither the first nor the last in the text information will be described).

That is, first, the phoneme sequence of interest is compared with the front phoneme sequence of a certain candidate phoneme sequence to narrow down the candidates. If it is not possible to narrow down the candidates to one, the rear phoneme sequence is first compared. To reduce the number of candidates to one. And if you still can't narrow down to one candidate,
If the last phoneme symbol of the previous phoneme sequence of the phoneme sequence of interest and the last phoneme symbol of the previous phoneme sequence of the candidate phoneme sequence are compared, the candidates are narrowed down. The candidates are narrowed by comparing the first one phoneme symbol of the post-phoneme sequence of the sequence with the first one phoneme symbol of the post-phoneme sequence of the candidate phoneme sequence. Therefore, in the second embodiment, if the candidates are narrowed down to one in the comparison between the pre-phoneme strings, the comparison between the post-phoneme strings is not performed. You are doing it.

However, in such a selection that largely depends on the preceding phoneme sequence, there may be cases where it is not always possible to optimize the articulation when text is synthesized. A specific example of such a case will be described with reference to the table of FIG. For example, the phoneme string output unit 2 decomposes the text “Oh my good morning” input from text information input means (not shown), and generates a phoneme string of a phoneme symbol equal to the phoneme string stored in the speech database 1. "O",
Consider the case where “ha”, “yo-”, “go”, “zai”, and “masu” are output to the phoneme string waveform selection unit 3 and the target phoneme string is “yo-”.

In this case, the voice database 1 contains
The first to fourth candidate phoneme strings “yo-” extracted from the text content shown in FIG. 5 are stored, and the respective pre-phoneme strings and candidate phoneme strings are as illustrated. Here, the previous phoneme sequence of the first candidate phoneme sequence is “a” instead of “ha”, and the fourth candidate phoneme sequence is “u” instead of “ku”. It often happens that waveform data is stored in the voice database 1 with consonants removed irrespective of the text content. The first to fourth candidate phoneme strings “yo
When selecting the optimal waveform data from "-", in the second embodiment, the target phoneme sequence is first compared with the previous phoneme sequence of the target phoneme sequence and the previous phoneme sequence of each candidate phoneme sequence to narrow the candidates. The third candidate phoneme string “yo-” having “ha” whose phoneme symbol matches the previous phoneme string of the row is selected as the optimal waveform data, and at the time of this selection, the other first phoneme string “yo-” is selected. , The second and fourth candidate phoneme strings “yo-” will be truncated.

However, if the voice synthesis of “Oh hayashi wa” is performed using the waveform data of the third candidate phoneme sequence “yo-”, even if the connection between “ha” and “yo-” is improved, This is not always a good thing because the connection between "yo-" and "go" is not taken into account at all. Therefore, in the fourth embodiment, the selecting unit 5
Does not select the optimum waveform data based only on the front phoneme sequence, but selects the optimum waveform data based on both the front phoneme sequence and the rear phoneme sequence.

That is, the selecting means 5 of the present embodiment compares the front phoneme sequence and the postphoneme sequence of the target phoneme sequence, counts the total number N of matching phoneme symbols, and selects the candidate having the maximum total number N. The waveform data of the phoneme sequence is selected as the optimum waveform data. For example, the previous phoneme sequence “a” of the first candidate phoneme sequence matches “a” in the previous phoneme sequence “ha” of the target phoneme sequence, and the post-phoneme sequence “go” of the first candidate phoneme sequence is Since it matches the succeeding phoneme string “go” after the target phoneme string, the total number N of matching phoneme symbols is N = 3 in the case of the first candidate phoneme string.
Becomes Similarly, when the total number N of the other candidate phoneme strings is obtained, N = 0 for the second candidate phoneme string, N = 2 for the third candidate phoneme string, and N = 2 for the fourth candidate phoneme string. If N = 2
Becomes Therefore, the first candidate phoneme string “yo−” in which the maximum total number N is N = 3 is selected as the optimal waveform data.

Since the first candidate phoneme string “yo-” is extracted from the same text content “Ohashio-aigo” as the input text, it is the most preferable waveform data for speech synthesis. Has been stored in the audio database 1 in a state where is "a", and thus is not selected as the optimal waveform data in the second embodiment. However, according to the fourth embodiment, as described above, the selecting means 5 can select the optimum waveform data.

Next, the operation of the fourth embodiment is limited to only the selection operation based on the comparison between the preceding phoneme sequence and the rear phoneme sequence, and will be described with reference to the flowcharts of FIGS. (FIG. 16 is equivalent to FIG. 6 of the second embodiment, and FIG. 17 is the same as FIG. 7 of the second embodiment.
8 and FIG. ). The operation of the other parts is the same as in the second embodiment, and will not be described.

First, when the phoneme string waveform selecting section 3 inputs text information from text information input means (not shown), the fundamental frequency pattern creating means 4 performs a morphological analysis on this text information to obtain one text. It is divided into a plurality of phoneme strings having accents (step 1a), and a fundamental frequency pattern of each phoneme string is created (step 2a). On the other hand, the phoneme string output unit 2 also inputs the text information, decomposes the text information, and outputs a phoneme string of a phoneme symbol equal to the phoneme string stored in the speech database 1 to the phoneme string waveform selection unit 3. The selecting means 5 inputs each phoneme string from the phoneme string output unit 2 (step 3a). Then, the selecting means 5 determines the phoneme sequence of interest (step 4a), and acquires data on all the candidate phoneme sequences from the voice database 1 (step 5a). Up to this point, the processing order is the same as that of the second embodiment and further the first embodiment.

Next, the selection means 5 determines whether or not the phoneme sequence of interest is the first one in the text (step 5).
01), since the preceding phoneme string does not exist at the beginning, the phoneme string after the phoneme string of interest and the phoneme string after the candidate phoneme string are compared, and the total number N of matching phoneme symbols is stored in its own memory. It is stored (step 502). If the target phoneme sequence is not the first one, it is determined whether or not it is the last one (step 503). If the last phoneme sequence is not the last one, both the front phoneme sequence and the rear phoneme sequence should exist. The pre-phoneme sequence and the post-phoneme sequence of the phoneme sequence of interest are compared with the pre-phoneme sequence and the post-phoneme sequence of the candidate phoneme sequence, and the total number N of matching phoneme symbols is stored in its own memory (step 504). If the target phoneme sequence is the last one, there is no post-phoneme sequence. Therefore, the front phoneme sequence of the target phoneme sequence and the front phoneme sequence of the candidate phoneme sequence are compared, and the total number of matching phoneme symbols N Is stored in its own memory (step 505).

As described above, the selecting means 5 determines whether or not the storage of the total number N has been completed for all the candidate phoneme strings (step 506), and if not, returns to step 501 to return to the next step 501. The total number N of the candidate phoneme strings is obtained and stored. Then, when the selection unit 5 finishes storing the total number N for all the candidate phoneme strings,
The candidate phoneme strings other than the candidate phoneme string having the maximum total number N are removed (step 507).

As a result of this removal, the selection means 5 determines that the candidate is 1
It is determined whether or not it has been narrowed down to one (step 508), and if it has been narrowed down to one, it is determined as optimal waveform data (step 509), and the processing of step 56a (see FIG. 10) is performed. Also, if the candidates are not narrowed down to one,
The processing after step 6a (see FIG. 9) is performed. That is, as in the case of the second embodiment, the selection unit 5 performs a process of narrowing down the number of candidate phoneme strings by comparing the value of the pattern with the value of the candidate phoneme string for the fundamental frequency (steps 6a to 6a). 13a) If still one candidate cannot be narrowed down, a formant mean square error with the previously determined waveform data is calculated, and a candidate phoneme sequence having the smallest formant mean square error value is selected. (Steps 50a to 56a).

As described above, in the fourth embodiment, the selecting means 5 does not select the optimum waveform data based only on the pre-phoneme sequence, but uses both the pre-phoneme sequence and the post-phoneme sequence as a reference. Since the optimum waveform data is selected in this way, it is possible to optimize the articulation which could not always be optimized in the second embodiment.

Next, a fifth embodiment of the present invention will be described. In the fifth embodiment, the discriminating operation of the selecting means 5 in selecting a candidate phoneme string in the fourth embodiment is further improved. In the fourth embodiment, the phoneme strings before and after the phoneme string of interest are compared with the phoneme strings before and after each candidate phoneme string,
The articulation is optimized by selecting a candidate phoneme string having the maximum sum of the matching phoneme symbols as the optimum waveform data. However, according to the study of the inventor of the present invention, it has been found that there is a case where the articulation cannot be sufficiently optimized even in the fourth embodiment. A specific example of such a case will be described with reference to the chart of FIG.

For example, the phoneme string output unit 2 decomposes the text “Oh my good morning” input from text information input means (not shown), and obtains phoneme symbols equal to the phoneme strings stored in the speech database 1. Phoneme string "o",
Consider the case where "ha", "yo-", "go", and "zaimasu" are output to the phoneme string waveform selection unit 3 and the target phoneme string is "go". In this case, the voice database 1 contains
The first to fourth candidate phoneme strings “go” extracted from the text content shown in FIG. 8 are stored, and the respective pre-phoneme strings and candidate phoneme strings are as illustrated.

When the total number N of the preceding and succeeding phoneme symbols in each candidate phoneme string is obtained, first, in the first candidate phoneme string, the preceding phoneme string “yo-” is replaced by the previous phoneme string “y” of the target phoneme string.
o- ”, all three characters match, and then the phoneme sequence“ za ”is followed by“ z ”in the phoneme sequence“ zaimasu ”after the target phoneme sequence.
Since the two characters match with "a", the value of N is 5. Next, in the second candidate phoneme string, the number of characters that match between the preceding phoneme string “no” and the target phoneme string “yo-” becomes zero. That is, the letters "o" and "yo" in "no"
The character "-" in "-" does not match, and the character "n" in "no" does not match the character "o" in "yo-". Further, since the number of characters that match between the phoneme sequence “do-” and the subsequent phoneme sequence “zaimasu” of the target phoneme sequence also becomes zero, the value of N in the second candidate phoneme sequence becomes zero. Hereinafter, similarly, the value of N in the third candidate phoneme string is 3, and the value of N in the fourth candidate phoneme string is 7. Therefore, according to the fourth embodiment, what is selected as the optimum waveform data is “go” of the fourth candidate phoneme sequence where N = 7.

However, when examining the text content relating to each candidate phoneme string, the text used in the first candidate phoneme string is “Oh my good morning”, which is the same as the input text. I have. Therefore, "go" of the first candidate phoneme sequence should be originally selected as the optimal waveform data. Nevertheless,
In the fourth embodiment, since the total number N of the preceding and succeeding matching phoneme symbols is used as a criterion for selection, “go” of the fourth candidate phoneme string is selected. When the content of the value 7 of N of the fourth candidate phoneme string is analyzed, the number of matching characters in the preceding phoneme string is zero, the number of matching characters in the rear phoneme string is 7, and the number of matching characters is It can be seen that it is extremely biased toward the row side. On the other hand, in the first candidate phoneme string, the value of N is 5, which is slightly smaller than the value of N in the fourth candidate phoneme string, but the number of matching characters in the previous phoneme string is Is 3, and the number of matching characters on the rear phoneme string side is 2, which can be regarded as being almost balanced. In other words, as a criterion for selecting the optimum waveform data, it is not sufficient that the value of the number of matching characters N is simply large, and further, the difference between the number of matching characters of the preceding phoneme string and the number of matching characters of the following phoneme string is not enough. It turns out that it is necessary to be balanced.

Therefore, in the fifth embodiment, the selecting means 5 determines whether the number of matching characters in the front phoneme string and the number of matching characters in the rear phoneme string are large to some extent, and the number of matching characters in both is balanced to some extent. The configuration is such that optimal waveform data is selected from the columns. The selecting means 5 in the present embodiment has a fuzzy arithmetic circuit inside. An evaluation value indicating the state of balance between the number of matched characters in the preceding phoneme string and the number of matched characters in the succeeding phoneme string based on the following fuzzy rules set in advance (in this specification, this evaluation value is referred to as a “balance evaluation value ). Y is calculated, and the optimal waveform data is calculated from each candidate phoneme sequence using the balance evaluation value y. In the fuzzy rules described below, N1 is the number of matching characters in the preceding phoneme string, N2 is the number of characters in the succeeding phoneme string, and S, M, and L are language labels.
It means ddle, Large. Then, the selecting means 5 digitizes these language labels and selects the optimum waveform data from candidate phoneme strings having a certain threshold value or more. IF N1 is S AND N2 is S THEN y is S IF N1 is S AND N2 is L THEN y is M IF N1 is L AND N2 is S THEN y is M IF N1 is L AND N2 is L THEN y is L

Using the above fuzzy rule, the balance evaluation value y of each of the first to fourth candidate phoneme strings in FIG.
Are obtained as L, S, M, and M, respectively. Therefore, according to the fifth embodiment, the most originally preferred first
Is selected by the selecting means 5 as the optimal waveform data.

Next, the operation of the fifth embodiment will be described with reference to the flowchart of FIG. 19 by limiting only the selection operation based on the comparison between the preceding phoneme sequence and the rear phoneme sequence described above (FIG. 19). Corresponds to FIG. 17 of the fourth embodiment.) The operation of the other parts is the same as that of the second embodiment or the fourth embodiment, and will not be described.

The selecting means 5 determines whether or not the phoneme sequence of interest is the first one in the text (step 501).
a) Since there is no previous phoneme sequence at the beginning,
The phoneme sequence after the target phoneme sequence and the phoneme sequence after the candidate phoneme sequence are compared, and the number N2 of matching phoneme symbols is stored in its own memory (step 502a). If the phoneme sequence of interest is not the first one, it is determined whether it is the last one (step 503a). If not, both the previous phoneme sequence and the rear phoneme sequence should exist. The pre-phoneme sequence and the post-phoneme sequence of the phoneme sequence of interest are compared with the pre-phoneme sequence and the post-phoneme sequence of the candidate phoneme sequence.
1, and N2 are stored in its own memory (step 504).
a). If the phoneme sequence of interest is the last one, there is no post-phoneme sequence, so the pre-phoneme sequence of the phoneme sequence of interest is compared with the pre-phoneme sequence of the candidate phoneme sequence, and the number N1 of matching phoneme symbols is determined. It is stored in its own memory (step 505a).

As described above, the selection means 5 determines whether or not the storage of the numbers N1 and N2 of the phoneme symbols that match for all the candidate phoneme strings has been completed (step 506a).
If not completed, the process returns to step 501a to obtain and store the total number N for the next candidate phoneme string.
When the selection means 5 finishes storing the number of matching characters N1 and N2 for all the candidate phoneme strings, it calculates the balance evaluation value y using the above fuzzy rule (step 5).
10a). The above fuzzy rule is for the case where the phoneme sequence of interest is neither the first nor the last. If the target phoneme sequence is the first or last, the variable of the condition part is only N2 or only N1, so that a simpler fuzzy rule is obtained. And thereafter, the selection means 5
Candidate phoneme strings other than the candidate phoneme string having the largest balance evaluation value y are removed (step 507a).

As a result of this removal, the selection means 5 determines that the candidate is 1
It is determined whether or not the number has been reduced to one (step 508a).
If it has been narrowed down to one, this is determined as the optimal waveform data (step 509a), and the processing of step 56a (see FIG. 10) is performed. If the number of candidates has not been reduced to one, the processing after step 6a (see FIG. 9) is performed. That is, as in the case of the second embodiment or the fourth embodiment, the selection unit 5 compares the value of the pattern with the value of the candidate phoneme string for the fundamental frequency to narrow down the number of candidate phoneme strings. (Steps 6a to 13a), and if the number of candidates cannot be narrowed down to one, the formant mean square error with the previously determined waveform data is calculated.
The candidate phoneme string having the smallest formant mean square error is selected (steps 50a to 56a).

As described above, in the fifth embodiment, the selecting means 5 determines that the number of matching characters in the front phoneme string and the number of matching characters in the rear phoneme string are somewhat large, and that the number of matched characters is well balanced to some extent. According to the fourth embodiment in which the optimum waveform data is selected from the candidate phoneme sequence, it is possible to optimize the articulation that could not always be optimized.

Although the fifth embodiment is superior to the fourth embodiment from the viewpoint of articulation optimization, the fifth embodiment requires a fuzzy operation to be performed. , The processing time becomes longer. Therefore, when it is desired to reduce the processing time, it is more preferable to adopt the configuration of the fourth embodiment. Further, in the above-described fifth embodiment, a method using fuzzy arithmetic is used to calculate the balance evaluation value y. However, a configuration using another method such as a neural network may be used.

[0093]

As described above, according to the present invention, the phoneme string waveform selecting unit creates a fundamental frequency pattern for each phoneme string having only one accent based on morphological analysis of text information. Selecting means for comparing each basic frequency of the candidate waveform data with the value of the basic frequency pattern and selecting candidate waveform data having a difference between the two within a predetermined range and satisfying a predetermined criterion as optimal waveform data; Therefore, natural and smooth voice output can be obtained without performing complicated prosody processing. As a result, the present invention can reduce the processing time for synthesizing a natural and smooth voice, and can suppress an increase in the memory capacity to achieve a sufficient cost reduction.

First, for phoneme symbols, it is determined whether or not the phoneme strings before and after the target phoneme string in the text information match the phoneme strings before and after each candidate phoneme string stored in the speech database. Is performed, the articulation at the time of text-to-speech synthesis can be optimized, and moreover, the intonation of the speaker recorded in the voice database 1 can be more often extracted. Obtainable.

Further, the amplitude average value over a predetermined section near the end of the waveform data determined as the optimum waveform data by the previous selection operation, and the amplitude average over the predetermined section near the front end of the waveform data of each candidate phoneme string. By adding a process for determining whether or not the difference from the value is within a predetermined range, it is possible to perform a more natural and smooth speech synthesis.

Then, the phoneme strings before and after the phoneme string of interest are compared with the phoneme strings before and after each candidate phoneme string, and the candidate phoneme string with the maximum total number of matching phoneme symbols is selected as optimal waveform data. By doing so, it is possible to more accurately optimize the articulation when synthesizing the text.

Further, a configuration is adopted in which the optimum waveform data is selected from candidate phoneme strings in which the number of matching characters in the front phoneme string and the number of matching characters in the rear phoneme string are somewhat large, and the number of matching characters is well balanced to some extent. By doing so, the accuracy at the time of optimizing the articulation can be further improved.

[Brief description of the drawings]

FIG. 1 is a block diagram showing a configuration of a speech synthesizer according to a first embodiment of the present invention.

FIG. 2 is a flowchart for explaining the operation of a phoneme string waveform selection unit 3 according to the first embodiment.

FIG. 3 is a flowchart for explaining the operation of a phoneme string waveform selection unit 3 according to the first embodiment.

FIG. 4 is a flowchart for explaining the operation of the phoneme string waveform selection unit 3 in the first embodiment.

FIG. 5 is a flowchart for explaining the operation of the phoneme string waveform selection unit 3 in the first embodiment.

FIG. 6 is a flowchart for explaining the operation of a phoneme string waveform selection unit 3 in the second embodiment.

FIG. 7 is a flowchart for explaining the operation of a phoneme string waveform selection unit 3 according to the second embodiment.

FIG. 8 is a flowchart for explaining the operation of a phoneme string waveform selection unit 3 according to the second embodiment.

FIG. 9 is a flowchart for explaining the operation of the phoneme string waveform selection unit 3 according to the second embodiment.

FIG. 10 shows a phoneme string waveform selection unit 3 according to the second embodiment.
4 is a flowchart for explaining the operation of FIG.

FIG. 11 is a diagram illustrating a process performed when obtaining the amplitude of waveform data in the third embodiment.

FIG. 12 is a diagram illustrating a method for obtaining an average amplitude value of waveform data according to the third embodiment.

FIG. 13 shows a phoneme string waveform selection unit 3 according to the third embodiment.
4 is a flowchart for explaining the operation of FIG.

FIG. 14 shows a phoneme string waveform selection unit 3 according to the third embodiment.
4 is a flowchart for explaining the operation of FIG.

FIG. 15 is a chart for explaining a problem to be solved by a fourth embodiment.

FIG. 16 shows a phoneme string waveform selection unit 3 according to the fourth embodiment.
4 is a flowchart for explaining the operation of FIG.

FIG. 17 shows a phoneme string waveform selection unit 3 according to the fourth embodiment.
3 is a flowchart for explaining the operation of FIG.

FIG. 18 is a table for explaining a problem to be solved by a fifth embodiment.

FIG. 19 shows a phoneme string waveform selection unit 3 according to the fifth embodiment.
4 is a flowchart for explaining the operation of FIG.

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 1 Speech database 2 Phoneme string output part 3 Phoneme string waveform selection part 4 Fundamental frequency pattern creation means 5 Selection means 6 Waveform data connection part 7 Voice output part

Claims (7)

[Claims] 1. A speech database in which waveform data and characteristic data of a phoneme string used for speech synthesis are stored, and text information for speech synthesis are input, and the text information is decomposed and stored in the speech database. A phoneme string output unit for sequentially outputting a phoneme string of a phoneme symbol equal to the phoneme string being input, and inputting each phoneme string from the phoneme string output unit, and then inputting a phoneme string of interest to be generated as waveform data. The phoneme sequence having the same phoneme symbol as this phoneme sequence of interest is determined as a candidate phoneme sequence, and the waveform data and feature data of the candidate phoneme sequence are determined. From the audio database, and selecting any one of the extracted waveform data as the waveform data of the optimal phoneme sequence for the phoneme sequence of interest. A phoneme string waveform selection unit that sequentially outputs optimal waveform data for each phoneme string of interest; a waveform data connection unit that connects waveform data of each phoneme string from the phoneme string waveform selection unit; and the waveform data connection unit. And a voice output unit that performs voice output of the text information based on the input of the connection waveform data from the speech information processing device.The phoneme sequence waveform selection unit, based on a morphological analysis of the text information, A basic frequency pattern creating unit that divides the text information in units of a phoneme string having only one, and creates a fundamental frequency pattern for each of the divided phoneme strings; The fundamental frequency of the waveform data included in the feature data of the candidate phoneme sequence is changed to the one accent corresponding to the one target phoneme sequence. Compared with the value of the fundamental frequency at a predetermined position of the fundamental frequency pattern of the phoneme string having, the waveform data of the candidate phoneme string having a difference between the two within a predetermined range and satisfying a predetermined criterion, Selecting means for selecting the optimal waveform data of the text information and sequentially outputting the optimal waveform data for each phoneme sequence of interest in correspondence with the input order of the text information. 2. The method according to claim 1, wherein the selecting means includes: when a difference between the two is within a predetermined range and a plurality of candidate phoneme string waveform data satisfying a predetermined criterion is present,
The formant mean square error between the waveform data already selected as the optimal waveform data by the previous selection operation and the waveform data of each candidate phoneme string is obtained, and the one with the smallest value is selected as the optimal waveform data. The speech synthesizer according to claim 1, wherein: 3. The method according to claim 2, wherein the selecting means determines whether or not the predetermined criterion is satisfied. The phoneme symbol includes a phoneme string before and after the target phoneme string in the text information and the phoneme string stored in the speech database. A determination is made as to whether or not the preceding and following phoneme strings of each candidate phoneme string match, and after the determination of the matching of the preceding and succeeding phoneme strings, the difference between the two values of the fundamental frequency is determined by a predetermined value. 3. The speech synthesizer according to claim 1, wherein a determination is made as to whether or not the speech is within a range. 4. The method according to claim 1, wherein the selecting means determines whether the predetermined criterion is satisfied or not, and determines the phoneme strings before and after the phoneme string of interest in the text information and the candidate phonemes stored in the speech database. By comparing the phoneme sequence before and after the sequence, the number of phoneme symbols that match in the pre-phoneme sequence of the target phoneme sequence and the pre-phoneme sequence of each of the candidate phoneme sequences is determined for each pre-phoneme sequence in each of the candidate phoneme sequences. For each of the candidate phoneme sequences, the number of phoneme symbols that match in the post-phoneme sequence of the target phoneme sequence and the post-phoneme sequence of each of the candidate phoneme sequences is determined. The sum of the number of matching phoneme symbols in the previous phoneme sequence and the number of matching phoneme symbols in the subsequent phoneme sequence is determined, and one candidate phoneme sequence is a candidate phoneme sequence having the maximum total value. It is to determine whether or not there is, After determining whether or not the sum is a candidate phoneme string having the maximum value, whether or not the difference between the two values of the fundamental frequency is within a predetermined range is determined. The speech synthesizer according to claim 1, wherein: 5. The method according to claim 1, wherein the selecting means determines whether or not the predetermined criterion is satisfied. The phoneme strings before and after the phoneme string of interest in the text information and the candidate phonemes stored in the speech database. By comparing the phoneme sequence before and after the sequence, the number of phoneme symbols that match in the pre-phoneme sequence of the target phoneme sequence and the pre-phoneme sequence of each of the candidate phoneme sequences is determined for each pre-phoneme sequence in each of the candidate phoneme sequences. For each of the candidate phoneme sequences, the number of phoneme symbols that match in the post-phoneme sequence of the target phoneme sequence and the post-phoneme sequence of each of the candidate phoneme sequences is determined. An evaluation value for the balance between the number of matching phoneme symbols in the previous phoneme sequence and the number of matching phoneme symbols in the subsequent phoneme sequence is calculated based on a preset rule, and one candidate phoneme sequence is calculated. But the best balance After determining whether or not the candidate phoneme string has the best balance, the above-described basic method is performed after determining whether or not the candidate phoneme string has the best balance. 3. The speech synthesizer according to claim 1, wherein a determination is made as to whether a difference between the two frequency values is within a predetermined range. 6. The voice synthesizing apparatus according to claim 3, wherein said selecting means determines whether or not said predetermined criterion is satisfied, and determines whether said waveform data is optimal waveform data by a previous selecting operation. A determination is made as to whether or not the difference between the average amplitude value over a predetermined section near the end of the determined waveform data and the average amplitude over a predetermined section near the front end of the waveform data of each candidate phoneme string is within a predetermined range. The discrimination is also performed, and the discrimination regarding the difference between the amplitude average values is performed after the discrimination regarding the preceding and following phoneme strings, and the difference between the two values of the fundamental frequency is within a predetermined range. Or after performing the determination as to whether or not the voice synthesis is performed. 7. The speech synthesizer according to claim 6, wherein the amplitude average value of the determined waveform data over a predetermined section near the end portion is weighted such that the amplitude value increases as approaching the end portion. It is calculated based on the waveform data, and the average amplitude value over a predetermined section near the tip in the waveform data of each candidate phoneme string is weighted such that the amplitude value increases as approaching the tip. A speech synthesizer calculated based on waveform data.