JP2001117577A - Voice synthesizing device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a voice synthesizing device which can generate synthesized voice of high quality by retrieving optimum waveform data. SOLUTION: A phoneme symbol string retrieving means (phoneme symbol string candidate retrieval part 3) retrieves a phoneme symbol string which matches at least part of a phoneme symbol string inputted from a phoneme symbol string decomposition part 2 and whose matching length is the longest from a voice database 1 and assigns it to the inputted phoneme symbol string. A voice synthesizing means (voice synthesis part 5) connects waveform data corresponding to the assigned phoneme symbol string and outputs synthesized voice. Thus, the longest phoneme symbol string is retrieved, and the voice can be synthesized with a small number of pieces of waveform data as a result and the synthesized voice of high quality can, therefore, be generated.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

The present invention relates to a speech synthesizer for retrieving waveform data stored in a speech database and outputting a synthesized speech.

[0002]

2. Description of the Related Art In recent years, a technique of recording and synthesizing for generating a synthetic sound from waveform data obtained by recording a human voice has attracted attention.

[0003] In a conventional speech synthesizer utilizing this recording / synthesis technique, first, a human voice is recorded, and waveform data obtained by dividing the speech waveform into predetermined speech units (phonemes) is obtained. Are prepared. Each of these waveform data is provided with a phoneme symbol as a search label, and is stored in advance in a storage device or the like together with the waveform data. The one constructed in this way is called a speech database.

In synthesizing speech, a text input as a command is first decomposed by phoneme decomposition processing using a word dictionary or the like, and is once converted into a plurality of phoneme symbols.

[0005] Subsequently, a phoneme symbol as a search label stored in the speech database is assigned to each converted phoneme symbol.

[0006] Then, waveform data corresponding to these assigned phoneme symbols are read from the speech database and connected, and after D / A conversion, amplified as appropriate and output as synthesized speech.

[0007]

The applicant of the present invention, unlike the above-mentioned conventional speech synthesizer, prepares various long and short waveform data corresponding to a plurality of continuous phonemes.
A large number of these are stored in a speech database, and waveform data searched as appropriate from the speech database is connected to realize a speech synthesis device that generates a synthesized sound.

However, in such a speech synthesizer, a technique for searching for an optimum one from a large number of waveform data in a speech database has not yet been established. For this reason, at present, it is difficult to generate a high-quality synthesized sound, and for example, the characteristics of a recorded voice may be sometimes impaired. Therefore, it is desired to realize a speech synthesizer capable of generating a high-quality synthesized sound by searching for optimal waveform data from a speech database.

SUMMARY OF THE INVENTION The present invention has been made in view of the above problems, and a purpose thereof is to provide a speech synthesizer capable of generating a high-quality synthesized sound by searching for optimum waveform data. To provide.

[0010]

In order to solve the above-mentioned problems, a voice synthesizing apparatus according to claim 1 of the present invention assigns a plurality of waveform data in which voice is recorded and a search label of each waveform data. A plurality of phoneme symbol strings stored are stored in the speech database, and a phoneme symbol string that matches at least a part of the input phoneme symbol string and has the longest length of the matched portion is stored in the speech database. It is characterized by having phoneme symbol string search means for searching and outputting, and speech synthesis means for outputting waveform data corresponding to the phoneme symbol string searched and output by the phoneme symbol string search means.

In the speech synthesizer according to the first aspect of the present invention, a plurality of waveform data in which speech is recorded and a plurality of phoneme symbol strings assigned as search labels for the respective waveform data are stored. An audio database is provided in advance.

For example, when a phoneme symbol string is input from a phoneme decomposition unit for performing phoneme decomposition, the phoneme symbol string search means matches at least a part of the input phoneme symbol string, and The phoneme symbol string having the longest length is retrieved from the speech database and output.

The speech synthesizing means reads out waveform data corresponding to the phoneme symbol string retrieved and output by the phoneme symbol string retrieval means from the speech database, and, for example, retrieves the phoneme symbol string input to the phoneme symbol string retrieval means. Connect and output waveform data in order.

As described above, according to the first aspect of the present invention, the number of connected waveform data and the number of connected portions are reduced by searching for the longest phoneme symbol string.

According to a second aspect of the present invention, in the speech synthesizer according to the first aspect, the phoneme symbol string search means preferentially searches the input phoneme symbol string. A different phoneme symbol string is searched from the speech database depending on whether a target portion is set or not.

In the speech synthesizing apparatus according to the second aspect of the present invention, the phoneme symbol string search means searches the phonetic database for different phoneme symbol strings, and finally obtains one phoneme symbol string from the different phoneme symbol strings. A phoneme symbol string can be selected, and as a result, optimal waveform data can be selected.

According to a third aspect of the present invention, there is provided the speech synthesizer according to the first or second aspect, wherein the phoneme symbol string retrieving means reads the input phoneme from the speech database. A plurality of sets of phoneme symbol strings constituting a symbol string are searched, and when the number of phoneme symbol strings included in each of the searched sets is the same, a set including many phoneme symbol strings terminated by vowel symbols is searched. It is characterized by selecting.

In the speech synthesizer according to the third aspect of the present invention, by selecting a set including a large number of phoneme symbol strings terminated by vowel symbols, in a speech-rich language formed by combining consonants and vowels, In particular, a natural synthesized sound is obtained.

[0019]

BRIEF DESCRIPTION OF THE DRAWINGS FIG. 1 is a block diagram showing an embodiment of a speech synthesizing apparatus according to the present invention.

FIG. 1 shows a first embodiment of a speech synthesizer according to the present invention.
FIG. 3 is a block diagram showing a configuration of the embodiment. The speech synthesizer shown in FIG. 1 includes a speech database 1, a phoneme symbol string decomposition unit 2, a phoneme symbol string candidate search unit 3, a candidate determination unit 4, and a speech synthesis unit 5.

The sound database 1 stores a plurality of waveform data in which sounds are recorded. Each waveform data corresponds to a speech unit or a phoneme sequence (a sequence of a plurality of consecutive phonemes) generally called a phoneme, and one symbol corresponding to the phoneme or two or more symbols corresponding to the phoneme sequence (Hereinafter referred to as “phoneme symbol strings” without distinguishing them)
A search label is assigned to each waveform data, and is stored in the voice database 1 together with the waveform data.

Each waveform data and phoneme symbol string are separately filed, and each phoneme symbol string is associated with waveform data by a file name. Here, it is assumed that the information is stored in the same database for the sake of simplicity.

The phoneme symbol string decomposition unit 2 is a block for decomposing a phoneme symbol string corresponding to a sentence into phrases. Normal,
A phoneme decomposition process is performed on text data corresponding to a sentence, a phoneme symbol string is once generated, and the phoneme symbol string is input to the phoneme symbol string decomposition unit 2 as a speech synthesis command. The phoneme symbol string decomposing unit 2 recognizes a symbol corresponding to a punctuation mark or a reading point in the input phoneme symbol string as a division point, and decomposes the phoneme symbol string into phrases.

The phoneme symbol string candidate search section 3 constitutes a phoneme symbol string search means. When a phoneme symbol string is input from the phoneme symbol string decomposition section 2, the phoneme symbol string corresponding to the waveform data is converted to speech. This is a block searched from the database 1.

By the way, in order to obtain a natural synthesized sound,
It is necessary to perform speech synthesis using waveform data containing as much speech as possible. In other words, a natural synthesized sound can be obtained by composing the synthesized sound with a smaller number of waveform data.

Therefore, the phoneme symbol string candidate search unit 3
The speech database 1 is configured to search the speech database 1 for a phoneme symbol string that matches at least a part of the input phoneme symbol string and has the longest length of the matched part.

As described above, by searching for the longest phoneme symbol strings and connecting the corresponding waveform data, natural synthesized speech can be generated.

The phoneme symbol string candidate search section 3 comprises a first candidate search section 31 and a second candidate search section 32. These blocks are configured so that different phoneme symbol strings can be searched depending on whether a part to be searched is preferentially set or not in the phoneme symbol string input from the phoneme symbol string decomposition unit 2. ing.

More specifically, the first candidate search unit 31 is configured to preferentially search a portion including the leading symbol of the input phoneme symbol string as a search target, while the second candidate search unit 31 Reference numeral 32 does not set such a priority search target, and is configured to set an arbitrary portion as a search target.

The candidate determining section 4 is a block constituting a phoneme symbol string search means. The candidate determining unit 4 uses the first candidate searching unit 31 and the second candidate searching unit 32 to search for a set of phoneme symbol strings (hereinafter, searched by the first candidate searching unit 31 or the second candidate searching unit 32). When one set of phoneme symbol strings is referred to as a “candidate”, one of the candidates is selected and assigned to the input phoneme symbol string, and the assigned candidate (phoneme symbol) is selected. Is output to the speech synthesizer 5.

The voice synthesizing section 5 is a block constituting voice synthesizing means. When the candidate assigned by the candidate deciding unit 4 is input, the speech synthesis unit 5 reads out waveform data corresponding to a phoneme symbol string in the candidate from the speech database 1, connects these waveform data, and connects the synthesized speech data. Is output.

It is also possible to add a phoneme symbol string decomposing unit for decomposing text data into phoneme symbols and a prosody processing unit for assigning a fundamental frequency and a reproduction time length to the converted phoneme symbols. . Further, a plurality of waveform data corresponding to the same phoneme symbol string and having different fundamental frequencies and playback time lengths are stored in the speech database 1 together with information on the fundamental frequencies and playback time lengths. However, a configuration may be adopted in which the plurality of waveform data is searched for and combined according to the fundamental frequency and the playback time length assigned by the prosody processing unit.
According to such a configuration, more appropriate waveform data can be selected and connected, so that a more natural synthesized sound can be generated.

Although not shown in FIG. 1, the speech synthesizing apparatus according to the present embodiment includes an arithmetic unit for executing processing,
A main storage unit capable of storing processing instructions and temporarily storing data is provided, and an external storage device is provided as appropriate. The processing is performed by sequentially reading the instructions stored in the main storage unit or the external storage device and executing them. For data transfer between blocks, a predetermined storage area provided in the main storage unit,
Alternatively, it is performed through a work area appropriately set by an instruction. Specifically, character string variables k1, k2,
The work areas and the like corresponding to k3 and k4 are set as appropriate.

Next, the flow of processing in this embodiment will be described. The phoneme symbol string disassembling unit 2 includes, for example, a phoneme symbol string / hazimemite corresponding to a Japanese text "Hello, thank you."
q yorosikuonegaisimasu qq
/ Is input. In this phoneme symbol sequence, / q /
The reading point “,” and / qq / are phoneme symbols indicating the period “.”, Respectively. The phoneme symbol string decomposing unit 2 recognizes these symbols as division points, and decomposes the input phoneme symbol string into phoneme symbol string units corresponding to phrases. For this reason, it is possible to prevent inconvenience that each phoneme symbol string on both sides of a reading point or a punctuation mark in the input phoneme symbol string is regarded as a continuous phoneme symbol string.

FIG. 2 is a flowchart showing the flow of processing in the first candidate search section 31. Here, the phoneme symbol string decomposing unit 2 outputs a phoneme symbol string / hazimemite /
Will be described.

In step S1, the first candidate search unit 31
However, as initialization processing, a phoneme symbol string / hazimema
site / is set to a character string variable k1. In the following step S2, the first candidate search unit 31 searches the speech database 1 for a phoneme symbol string that includes the leading symbol / h / of the character string variable k1 and has the longest length (column length). read out.

For example, in step S2, a phoneme symbol string corresponding to the entire phrase is searched first, and if it cannot be searched, the speech database 1 is sequentially reduced while reducing the length of the string.
Search inside.

In this manner, the longest phoneme symbol string can be searched from the speech database 1.

Note that the speech database 1 is sequentially searched while the column length is reduced, and when even the first phoneme symbol / h / cannot be finally searched, the first candidate search unit 31
A special symbol / c / is assigned to the first phoneme symbol. This symbol / c / has a special meaning as a search error. Here, it is assumed that the phoneme symbol string / hazime / has been searched without causing a search error.

After the phoneme symbol string is searched in step S2, the process proceeds to step S3, where the first candidate search unit 31 converts the phoneme symbol string read in step S2 from the character string variable k1. cut out. That is, the phoneme symbol string is deleted from the storage area.

In the following step S4, the first candidate search unit 31 counts the number of phoneme symbols included in the character string variable k1, and in the following step S5, the number of phoneme symbols counted is 0 (zero). It is determined whether or not there is.

If it is determined in step S5 that the answer is NO (the number of phoneme symbols is not 0 (zero)), it means that all the searches for the input phoneme symbol string / hazimemite / have not been completed. Returning to S2, after this step, the current character string variable k1,
That is, the search is continued for the currently remaining phoneme symbol string / masite /. Similarly, when the symbol / c / signifying the search error described above is assigned to the first symbol,
A search is performed for the second and subsequent phoneme symbol strings excluding the first symbol.

On the other hand, if it is determined in step S5 that the answer is YES (the number of phoneme symbols is 0 (zero)), it means that the search for the phoneme symbol string / hazimemite / entire has been completed.

Thus, the first candidate search section 31
Is, for example, / hazime /, / masi / and / t
A set of phoneme symbol strings consisting of e / is read from the speech database 1. Hereinafter, the set of phoneme symbol strings read out by the first candidate search unit 31 is referred to as candidate 1.

In step S6, the first candidate search unit 31
Outputs the candidate 1 to the candidate determining unit 4 and ends a series of processes.

FIG. 3 is a flowchart showing the flow of processing in the second candidate search section 32. The phoneme symbol string / ha corresponding to the phrase input from the phoneme symbol string decomposition unit 2
When Zymmassite / is input, step S1
In 1, the second candidate search unit 32 sets this phoneme symbol string as a character string variable k2.

In step S12, the second candidate search unit 3
2 retrieves and reads from the speech database 1 a phoneme symbol string that includes any part of the character string variable k2 and has the longest string length.

Here, similarly to step S2 in FIG.
The phoneme symbol string corresponding to the entire phrase is searched. If the search is not successful, the speech database 1 is searched sequentially while reducing the length of the string. Here, the phoneme symbol string / z
It is assumed that im / has been searched. Note that the phoneme symbol string / h
If none of the phoneme symbol strings included in azimemasite / can be retrieved from the speech database 1, the symbol / c / representing the above-described search error is assigned to each phoneme symbol.

After the phoneme symbol string is retrieved in step S12, the process proceeds to step S13, where the second
Sets the phoneme symbol string / ha / located before the "read phoneme symbol string" in step S12 as the character string variable k3.

In the following step S14, the second candidate search unit 32 performs the same processing as in FIG. 2 for the character string variable k3, and proceeds to step S15.

In step S15, the second candidate search unit 3
2 sets the phoneme symbol string / masit / following the "read phoneme symbol string" in step S12 to the character string variable k4.
And the process proceeds to step S16.

In step S16, the second candidate search unit 3
2 performs the same processing as in FIG. 2 on the character string variable k4.

Thus, the second candidate search section 32
Retrieves a set of phoneme symbol strings consisting of, for example, / ha /, / zyme /, / masi /, and / te /, and reads them from the speech database 1. Hereinafter, the set of phoneme symbol strings read by the second candidate search unit 32 is referred to as candidate 2.

In step S17, the second candidate search unit 3
2 outputs the candidate 2 to the candidate determining unit 4 and ends a series of processing.

The character string variable k3 or the character string variable k4
May be temporarily transferred to a character string variable k1, and then the first candidate search unit 31 may perform a search. Further, the phoneme symbol string searched in step S12 is
Phoneme symbol string / hazimemite / first symbol /
When h / is included, the processing in steps S13 and S14 is omitted, while when h / is included, the processing in steps S15 and S16 is omitted.

FIG. 4 is a flowchart showing the flow of the process performed by the candidate determining unit 4. For the candidate determination unit 4, the first
When the candidates 1 and 2 read by the candidate search unit 31 and the second candidate search unit 32 are input,
First, in step S21, the candidate determination unit 4 compares the number of phoneme symbol strings in each candidate.

If it is determined in step S21 that the number of phoneme symbol strings is different, the process proceeds to step S22, where the candidate determination unit 4 finally determines the candidate having the smaller number of phoneme symbol strings into phoneme symbol strings. Assign to column / hazimemite /. That is, a phoneme symbol string as a search label is assigned to a corresponding part of the input phoneme symbol string. Here, candidate 1 having fewer phoneme symbol strings than candidate 2 is assigned.

Then, the candidate deciding unit 4 outputs the assigned candidates to the speech synthesizing unit 5 and ends the processing.

On the other hand, if it is determined in step S21 that the number of phoneme symbol strings is the same, step S2
Proceeding to 3, the candidate determining unit 4 extracts the last phoneme symbol of the phoneme symbol string in each candidate. In step S24, the candidate determination unit 4 counts the number of symbols (hereinafter, referred to as vowel symbols) corresponding to vowels in the extracted phoneme symbols, and proceeds to step S25.

In step S25, the candidate determination unit 4 finally determines the candidate having the larger number of vowel symbols in the phoneme symbol string / ha
It is assigned to the "Zymmasite /", and the assigned candidate is output to the speech synthesizer 5 to complete the processing.

When the processing by the candidate deciding unit 4 is completed in this way, the speech synthesizing unit 5 reads out the waveform data corresponding to the phoneme symbol string in the input candidate from the speech database 1 and connects these waveform data. , D / A conversion, and signal amplification to generate and output a synthesized sound.

Therefore, according to the speech synthesizer of this embodiment, the phoneme symbol string candidate search unit 3 and the candidate determination unit 4
The phoneme symbol string search means is configured such that when a phoneme symbol string is input, the phoneme symbol string search means matches at least a part of the phoneme symbol string, and the length of the matched part is the longest. Since the phoneme symbol string to be retrieved from the speech database is retrieved and output, and the speech synthesis unit 5 constituting the speech synthesis means outputs the waveform data corresponding to the phoneme symbol string from the phoneme symbol string search means, A synthesized sound can be composed of a small number of waveform data, and therefore, a natural and high-quality synthesized sound can be generated with few connection parts that generate noise.

The phoneme symbol string search means searches the phonetic database for different phoneme symbol strings depending on whether a part to be searched is preferentially set or not in the input phoneme symbol string. So that
Finally, one phoneme symbol string can be selected from the different phoneme symbol strings. As a result, the range of waveform data selection can be expanded, and the optimum waveform data can be used for generating a synthesized sound. Becomes

It should be noted that the portion to be searched preferentially is not limited to the portion including the leading symbol as in the above embodiment. Therefore, a different phoneme symbol string can be searched for by setting the part to be searched in the input phoneme symbol string as appropriate, thereby expanding the range of selection of waveform data. Can be.

The phoneme symbol string search means searches a plurality of sets of phoneme symbol strings constituting the input phoneme symbol string from the speech database, and finds the number of phoneme symbol strings included in each of the searched sets. When the numbers are the same, a set that includes many phoneme symbol strings terminated by vowel symbols is selected and assigned.Therefore, especially in languages such as Japanese with many voices composed of a combination of consonants and vowels, A complex synthesized sound can be generated.

Each block of the speech synthesizer according to the present embodiment is constituted by software and recorded on a computer-readable recording medium so that the speech synthesizing device can be distributed and transported. It is also possible to obtain the effect of.

Further, each of the above blocks can be constituted only by hardware such as an integrated circuit, and a part thereof can be constituted by software. In any case, the same effect as above can be obtained. Furthermore, the blocks may be distributed and deployed not only on a single computer but also on terminals and server machines that configure a network.

[0068]

As described above, according to the first aspect of the present invention,
According to the speech synthesizing apparatus according to the above, since the phoneme symbol string that matches at least a part of the input phoneme symbol string and has the longest length of the matched part is searched, a small number of waveform data Can synthesize speech, and therefore, it is possible to generate a high-quality synthesized sound.

According to the speech synthesizing apparatus according to the second aspect of the present invention, a different phoneme is set in the input phoneme symbol string than when a portion to be searched is preferentially set or not set. Because we decided to search for a symbol string,
As a result, the range of selection of the waveform data can be widened, and therefore, a higher-quality synthesized sound can be generated.

According to the speech synthesizing apparatus of the third aspect of the present invention, a plurality of sets of phoneme symbol strings are searched, and when the number of phoneme symbol strings included in each of the searched sets is the same. Selects a set including many phoneme symbol strings terminated by vowel symbols, so that a particularly natural synthesized sound can be generated in a language with many voices formed by combining consonants and vowels.

[Brief description of the drawings]

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

FIG. 2 is a flowchart showing a flow of processing in a first candidate search unit having the form shown in FIG. 1;

FIG. 3 is a flowchart showing a flow of processing in a second candidate search unit having the form shown in FIG. 1;

FIG. 4 is a flowchart showing a flow of processing in a candidate determining unit having the form shown in FIG. 1;

[Explanation of symbols]

 Reference Signs List 1 voice database 2 phoneme symbol string decomposition section 3 phoneme symbol string candidate search section 31 first candidate search section 32 second candidate search section 4 candidate determination section 5 speech synthesis section

Claims (3)

[Claims] 1. A speech database storing a plurality of waveform data in which speech is recorded and a plurality of phoneme symbol strings assigned as search labels for the respective waveform data, and at least a part of the input phoneme symbol string. And a phoneme symbol string search means for searching and outputting a phoneme symbol string having the longest matching portion from the speech database, and a phoneme symbol string searched and output by the phoneme symbol string search means. And a voice synthesizing unit that outputs waveform data corresponding to. 2. The phoneme symbol string search means according to claim 1, wherein said phoneme symbol string is different from said speech database depending on whether a part to be searched is preferentially set or not in said input phoneme symbol string. 2. The speech synthesizer according to claim 1, wherein a search is made for. 3. The phoneme symbol string search means searches a plurality of sets of phoneme symbol strings constituting the input phoneme symbol string from the speech database, and finds phoneme symbols included in each of the searched sets. When the number of columns is the same,
3. The speech synthesizer according to claim 1, wherein a set including a large number of phoneme symbol strings terminated by vowel symbols is selected.