JP4287664B2 - Speech synthesizer - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a speech synthesizer which is capable of synthesizing read aloud speech of consistently high quality by utilizing the rhythm pattern data formed from actual speech to the maximum possible extent. <P>SOLUTION: A language analysis section 101 which performs analysis of input text and outputs the result of language processing expressed by a phonetic symbol sequence assigns the position where the phonetic symbol sequence is divided in the arbitrary accent phrase of the phonetic symbol sequence and a rhythm retrieval section 102 carries out retrieval in an accent phrase unit or the unit segmented by the accent division position assigned within the accent phrase as the unit for retrieving the rhythm pattern from a rhythm pattern data base 103, thereby carrying out rhythm formation without using means exclusive of the rhythm pattern retrieval. <P>COPYRIGHT: (C)2004,JPO&amp;NCIPI

Description

[0001]
BACKGROUND OF THE INVENTION
The present invention relates to an apparatus for generating synthesized speech. More specifically, the present invention relates to a device that converts an input character string (text) into speech.
[0002]
[Prior art]
A method for accumulating speech feature values from recorded speech data as a database in order to generate natural synthesized speech, and obtaining acoustic parameters required for speech synthesis by combining units of speech feature values in the database Has been proposed. For example, the database uses speech feature values indexed by text strings or phonetic symbol strings, and matches text strings to be synthesized or phonetic symbol strings obtained by text processing. A basic frequency required for synthesis from a method for searching for a unit of a feature value of a portion of a voice (Japanese Patent Laid-Open No. 8-87297; called method 1) or a database of basic frequency patterns indexed as phoneme patterns There is a method of searching and generating a pattern (US PATENT 5,905,972; called method 2). Since these methods directly use feature values of speech extracted from speech, for example, when used as a prosody generation method, synthesized speech with higher naturalness can be generated as compared with a model-based prosody generation method.
[0003]
On the other hand, in a speech synthesizer that generates a prosody based on a rule, it is a general method to generate prosodic information for each accent phrase of a phonetic symbol string output as a result of language processing.
[0004]
[Patent Document 1]
JP-A-8-87297
[Patent Document 2]
US Pat. No. 5,905,972
[0005]
[Problems to be solved by the invention]
However, in the above-described method for retrieving the necessary speech feature values from the database, the prosody is required when synthesizing, that is, when prosody data matching the text to be synthesized or the phonetic symbol string is not found in the database. An alternative means of generating information is required. For example, in method 1, prosody generation based on rules is performed for portions that do not match the prosodic pattern in the database. Similarly, in the method 2, an alternative means of linear interpolation is used for a portion that cannot be generated from the database.
[0006]
Therefore, in these methods, as shown in FIG. 1A, the prosody generated is disadvantageous in that a prosody naturalness gap is generated between the part that matches the prosody data in the database and the part generated by the alternative means. Have In general, when a prosody database is constructed from recorded data of natural utterances, the distribution of prosodic data in the database decreases as the number of mora increases, so the necessary combinations of accent types and phonogram strings should be covered sufficiently. Becomes difficult. In the prior art using alternative means, the above-mentioned disadvantage occurs in a portion with a relatively long mora number (for example, a compound word in which a noun continues, a portion in which an auxiliary verb or a predicate with a particle is converted into a phonogram string). The trend is getting stronger.
[0007]
Also, in a speech synthesizer that generates prosody based on rules, as shown in FIG. 1B, it is difficult to generate a stable prosody without an uncomfortable feeling with respect to an accent phrase having a long mora number. It becomes a cause that becomes difficult to hear.
[0008]
In real speech, the frequency of pose / pitch reconstruction increases as the utterance speed slows down, but the conventional speech synthesizer generates prosody by dividing one accent phrase as a result of language processing. Can not do it.
[0009]
An object of the present invention is to provide a speech synthesizer capable of generating natural synthesized speech.
[0010]
[Means for Solving the Problems]
According to one aspect of the present invention, a speech synthesizer includes a language analysis unit, Prosodic pattern database, A prosody search unit and a waveform generation unit are provided. The language analyzer Output at least accent phrases from input text . The prosodic pattern database stores prosodic patterns extracted from real sounds. The prosody search part Above The prosodic pattern corresponding to the accent phrase From the prosodic pattern database Prosody information is generated by searching. The prosody search part Of the accent phrases, a specific accent phrase composed of one accent phrase is divided, and prosodic information is generated by searching the prosodic pattern in the divided units. . The waveform generator Above Prosodic information From Synthesize speech waveform.
[0011]
Preferably, the specific accent phrase has a corresponding prosodic pattern. Prosodic pattern An accent phrase that does not exist in the database.
[0012]
Preferably, the specific accent phrase is an accent phrase having a mora number exceeding a reference value. The above reference value is Prosodic pattern Stored in the database This is the number of mora of the group with the largest number of mora among the groups with all accent types in the group by number of mora. .
[0013]
Preferably, the specific accent phrase includes a plurality of accent phrases. word Is an accent phrase corresponding to consecutive compound words. The prosodic search unit is included in the compound word Divide into each of the words and into the divided words Prosody information is generated by searching corresponding prosodic patterns.
[0014]
Preferably, The specific accent phrase is an accent phrase that is selected when the set pause frequency exceeds the reference value. .
[0015]
According to another aspect of the present invention, the speech synthesizer includes a language analysis unit, a prosody generation unit, and a waveform generation unit. The language analyzer Output at least accent phrases from input text . The prosody generation part Above Predetermined rules for each accent phrase From Prosody information is generated. The prosody generation part A specific accent phrase composed of one accent phrase is divided among the accent phrases, and prosodic information is generated from the rules in divided units. . The waveform generator Above Prosodic information From Synthesize speech waveform.
[0016]
Preferably, the specific accent phrase is an accent phrase having a mora number exceeding a reference value.
[0017]
Preferably, the specific accent phrase includes a plurality of accent phrases. word Is an accent phrase corresponding to consecutive compound words. The prosody generation unit is included in the compound word word Per the above rules From Prosody information is generated.
[0018]
Preferably, The specific accent phrase is an accent phrase that is selected when the set pause frequency exceeds the reference value. .
[0019]
In the first aspect of the present invention, the input text to be read out is converted into a phonetic symbol string, and the position where the phonetic symbol string is divided in an arbitrary accent phrase of the converted phonetic symbol string is specified. Possible language analysis means, prosodic pattern database storing speech feature values extracted from real speech as prosodic pattern information, and units of accent phrases of phonetic symbol strings output by language analysis means, or division within accent phrases Prosody information that generates prosodic information by searching the prosodic information that matches the units divided by position, and waveform synthesis that synthesizes speech waveforms according to the prosodic information generated by the prosody searching means And a speech synthesizer comprising the means.
[0020]
If a prosodic pattern that matches a given accent phrase cannot be found in the database, the prosody pattern can be re-searched in units of the accent phrase divided according to the division position. Prosody generation with high naturalness using patterns to the maximum is possible.
[0021]
In the second aspect of the present invention, in the first aspect, from the distribution of features that distinguish prosody in prosodic pattern search such as the number of mora of prosodic patterns and accent type stored in the prosodic pattern database, the language analysis means Determine the split position within the phrase.
[0022]
At the language processing stage, the division position in the accent phrase is determined in consideration of the features of the prosodic pattern in the database, so the prosody pattern can be searched from the database at least in the unit divided at the division position. It is possible to generate prosody with high naturalness over the entire sentence without using prosody generation means.
[0023]
In a third aspect of the present invention, in the first aspect, for each morpheme constituting a compound word in an accent phrase of a phonetic symbol string obtained by converting a compound word part in which a noun of an input text is continued by a language processing unit. The division position in the accent phrase is set.
[0024]
By generating prosody by connecting prosodic patterns that are divided and searched for each word, it is possible to generate a synthesized sound in which each word of the compound word is clearly audible.
[0025]
In the fourth aspect of the present invention, in the first aspect, whether or not the division position in the accent phrase set by the language processing means is handled as the division position by the parameter for controlling the pause frequency from the outside in the prosody search means. It was set as the structure which determines.
[0026]
If the parameter that specifies the pause frequency is set to include more pauses, the prosody pattern is searched in units divided by the split position in the accent phrase, and the parameter that specifies the pause frequency includes the pause as much as possible. If it is set to the other side, the prosody pattern can be controlled by ignoring the split position in the accent phrase and searching for the prosody pattern as if there is no split position. Can do.
[0027]
In the fifth aspect of the present invention, the input text to be read out is converted into a phonetic symbol string, and the position where the phonetic symbol string is divided in an arbitrary accent phrase of the converted phonetic symbol string is specified. Prosodic generation means for generating prosodic information according to a rule for each unit of accent phrase of a phonetic symbol string output by the language analyzing means, or for each unit divided at the division position in the accent phrase, The speech synthesizer comprises: a waveform synthesis generation unit that synthesizes a speech waveform according to the prosodic information generated by the generation unit.
[0028]
Even with accent phrases with a large number of mora, prosody generation with a sense of sharpness can be achieved by using the division position as a pitch starting point when generating prosody.
[0029]
According to a sixth aspect of the present invention, in the fifth aspect, for each morpheme constituting a compound word in an accent phrase of a phonetic symbol string obtained by converting a compound word part in which the noun of the input text is continued by the language processing means. The division position in the accent phrase is set.
[0030]
Even in an accent phrase formed by combining accents of compound words, it is possible to control the subtle pitch rise at the division position, so even when generating prosody by rules, pronunciation for each word that makes up the compound word Can produce a clear synthesized sound.
[0031]
In the seventh aspect of the present invention, in the fifth aspect, whether or not the division position in the accent phrase set by the language processing means is treated as a division position by a parameter for controlling the pose frequency from the outside in the prosody generation means. It was set as the structure which determines.
[0032]
When the parameter that specifies the pause frequency is set to include more pauses, the prosody control is performed using the division position in the accent phrase as the starting point of the pitch, and the parameter that specifies the pause frequency is set to pause. Prosody control that does not raise the pitch at the division position in the accent phrase can be performed when the setting is set so as not to allow as much as possible. Control can be performed.
[0033]
DETAILED DESCRIPTION OF THE INVENTION
Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings. In the drawings, the same or corresponding parts are denoted by the same reference numerals, and description thereof will not be repeated.
[0034]
(First embodiment)
<Configuration of text-to-speech synthesizer>
The configuration of the text-to-speech synthesizer according to the first embodiment is shown in FIG. This apparatus includes a language analysis unit 101, a prosody search unit 102, a prosody pattern database 103, and a waveform generation unit 104.
[0035]
The language analysis unit 101 performs language analysis on input text, and includes a phonetic symbol string, accent information, accent phrase delimiter position, and accent phrase division position as information for generating a synthesized sound that reads out the text. Output analysis results.
[0036]
The prosodic pattern database 103 stores acoustic feature quantities extracted from real speech as prosodic patterns that can be searched using at least the number of mora and accent type, and combinations of phonetic symbol strings as search conditions. The prosodic pattern database 103 is a database in which prosodic patterns in units of accent phrases extracted from recorded speech can be searched by the number of accent phrase mora and accent positions (accent type).
[0037]
The prosody search unit 102 uses the phonetic symbol string and accent type of the accent phrase unit output by the language analysis unit 101 as a search key as the search key. A prosodic pattern matching the search key is searched from 103 and extracted. The prosodic search unit 102 extracts prosodic patterns for each accent phrase, connects them, generates prosodic information over one sentence, and outputs it.
[0038]
The waveform generation unit 104 synthesizes a speech waveform according to the prosodic information generated by the prosody search unit 102 (generates synthesized speech).
[0039]
Such a text-to-speech synthesizer is constructed on a computer system as shown in FIG. 3, for example. This computer system includes a main body unit 501, a keyboard 502, a display 503, an input device (mouse) 504, and a speaker 509 and is a system capable of text input and voice output. The language analysis unit 101, the prosody search unit 102, the prosody pattern database 103, and the waveform generation unit 104 in FIG. 1 are stored in the CD-ROM 508 set in the main unit 501, the disc (memory) 505 built in the main unit 501 or And stored in the disk 506 of another system connected by the line 507.
[0040]
<Synthetic voice generation procedure>
A procedure for generating synthesized speech by the text-to-speech synthesizer configured as described above will be described with reference to FIG.
[0041]
[Step ST101]
The text to be read is entered. Text input is performed by an input device such as a keyboard / mouse, reading a text file, or the like. Here, it is assumed that the text “You may think so” is entered.
[0042]
[Step ST102]
The language analysis of the input text is performed by the language analysis unit 101. As information for generating a synthesized sound of the input text, a language analysis result (phonetic symbol string) including reading information, accent information, accent phrase delimiter position, and accent phrase division position is obtained. Here, as shown in FIG. 5, the phonetic symbol string “Saw / Omo'Ukamo | Silemacene” is obtained as the language analysis result for the given text “Maybe I think so”. In this phonetic symbol string, the reading of the text is indicated by “Katakana”, the accent position (accent core) is indicated by the symbol “′”, the accent phrase delimiter is indicated by the symbol “/”, and the delimiter within the accent phrase The position is indicated by the symbol “|”. The accent phrase delimiter position “|” is an accent phrase (accent phrase 2 in this case) combined with one accent phrase in the language analysis, and a position that becomes a sentence boundary in the language analysis (here, the boundary between the phrase 2 and the phrase 3) Inserted into.
[0043]
[Step ST103]
The prosodic search unit 102 selects the first accent phrase (here, accent phrase 1 “saw”).
[0044]
[Step ST104]
The prosodic retrieval unit 102 retrieves corresponding (matching) prosodic patterns from the prosodic pattern database 44 using the number of mora (2 mora) and accent type (0 type) of the selected accent phrase 1 “saw” as a search condition. If the corresponding prosodic pattern exists, the process proceeds to step ST105, and if not, the process proceeds to step ST107. Here, as shown in FIG. 6A, since a 2-mora 0 type prosodic pattern corresponding to the accent phrase 1 “saw” exists in the database 44, the process proceeds to step ST105.
[0045]
[Step ST105]
The prosody search unit 102 determines whether all accent phrases have been searched. If it is determined that all the accent phrases have been searched, the process proceeds to step ST108, and if not, the process proceeds to step ST106. Here, since the accent phrase 2 has not been searched yet, the process proceeds to step ST106.
[0046]
[Step ST106]
The prosodic search unit 102 selects the next accent phrase (in this case, accent phrase 2 “Omo'Ukamo | Sillemasen”).
[0047]
[Step ST104]
The prosodic search unit 102 searches the prosodic pattern database 44 for a corresponding prosodic pattern using the number of mora (10 mora) and accent type (type 2) of the selected accent phrase 2 “Omo'Ukamo | Here, as shown in FIG. 6B, the 10 mora type 2 prosodic pattern corresponding to the accent phrase 2 “Omo'Ukamo | Silemacene” does not exist in the database 44. Therefore, the process proceeds to step ST107.
[0048]
[Step ST107]
The prosodic search unit 102 searches for a prosodic pattern corresponding to the unit 1 “Omo'ucamo” and the unit 2 “Siremasen” obtained by dividing the accent phrase 2 by the break position “|”. As shown in FIG. 7A, the prosody search unit 102 corresponds from the prosody pattern database 44 using the number of mora (5 mora) and the accent type (type 2) of the unit 1 “Omo'ukamo” as a search condition (match). Search for prosodic patterns. Here, a 5-mora type 2 prosodic pattern corresponding to the unit 1 “Omo'Ukamo” exists in the database 44. Further, as shown in FIG. 7B, the prosody search unit 102 corresponds (matches) from the prosodic pattern database 44 using the number of mora (5 mora) and the accent type (0 type) of the unit 2 “Sillemasen” as a search condition. ) Search for prosodic patterns. Here, a 5-mora 0 type prosodic pattern corresponding to the unit 2 “Shiramasen” exists in the database 44. Then, the process proceeds to step ST105.
[0049]
[Step ST105]
Since all the accent phrases have been searched, the process proceeds to step ST108.
[0050]
[Step ST108]
As shown in FIG. 8, the prosodic search unit 102 connects the prosodic patterns searched in steps ST104 and ST107 and outputs prosodic information. Since all the prosodic information is a prosodic pattern obtained from the prosodic database 44, the auditory gap at the connection portion is smaller than that of the conventional example.
[0051]
[Step ST109]
The waveform generation unit 104 synthesizes a speech waveform (generates synthesized speech) according to the prosodic information generated by the prosody search unit 102.
[0052]
<Effect>
In the text-to-speech synthesizer according to the first embodiment, the language analysis unit 101 sets a break position in the accent phrase. Therefore, the prosody search unit 102 can search for prosodic patterns in units with a shorter number of mora than the accent phrase. As a result, it is possible to increase the ratio that matches the prosodic pattern stored in the prosodic pattern database 103. As a result, it is possible to generate a synthesized sound having a high-quality prosody using the most natural prosodic information extracted from real speech.
[0053]
(Second Embodiment)
The configuration of the text-to-speech synthesizer according to the second embodiment is the same as the configuration of the apparatus shown in FIG. The difference from the apparatus according to the first embodiment is that a reference number of mora is set in advance and a prosodic pattern is searched according to this number of reference mora.
[0054]
<Setting the number of reference mora>
A procedure for setting the number of reference mora will be described with reference to FIG.
[0055]
[Step ST201]
The prosodic patterns stored in the prosodic pattern database 103 are divided into groups for each number of mora. Here, as shown in FIG. 10, it is assumed that the group is divided into a group 2 of 2 mora GR2, a group GR3 of 3 mora, a group GR4 of 4 mora, and a group GR5 of 5 mora.
[0056]
[Step ST202]
Determine whether the prosodic pattern for each group covers all possible accent types. As shown in FIG. 10, the 2-mora group GR2 includes a 0-type prosodic pattern, a 1-type prosodic pattern, and a 2-type prosodic pattern. That is, the prosodic patterns included in the group 2 of 2 mora cover all the accent types that can be taken in the number 2 of mora. Similarly, the prosodic pattern included in the group 3 of 3 mora covers all the accent types that can be taken in the number 3 of mora, and the prosodic pattern included in the group 4 of 4 mora has the accent type that can be taken in the number 4 of mora. It covers everything. On the other hand, the 5-mora group GR5 includes a prosodic pattern with an accent type of 0 and a prosodic pattern of type 2. Prosodic patterns with accent types of 1, 3, 4, and 5 are not included. That is, the prosodic patterns included in the 5-mora group GR5 do not cover all of the accent types that can be taken in the number 5 of mora.
[0057]
[Step ST203]
The maximum value of the number of mora of the group determined to cover all the accent types in step ST202 is set as the reference number of mora for setting the division position in the accent phrase. Here, the reference number of mora is set to 4 mora.
[0058]
<Synthetic voice generation procedure>
Next, a synthetic speech generation procedure will be described with reference to FIG.
[0059]
[Step ST101]
The text to be read (here, the text “information processing apparatus”) is input (see FIG. 12A).
[0060]
[Step ST102]
The language analysis of the input text is performed by the language analysis unit 101. Here, as shown in FIGS. 12A and 12B, the phonetic symbol string “Joe Hoshori Sochi” is obtained as the language analysis result for the given text “information processing apparatus”.
[0061]
[Step ST301]
The first accent phrase is selected from the phonetic symbol string obtained in step ST102. Here, the accent phrase “Joe Hoshori Sochi” is selected.
[0062]
[Step ST302]
It is determined whether or not the number of mora of the selected accent phrase exceeds the reference number of mora. When it exceeds, the process proceeds to step ST303, and when it does not exceed, the process proceeds to step ST304. Here, the number of mora of the accent phrase “Joe Hoshori Sochi” is 9 mora, which exceeds the standard mora number of 4 mora. Therefore, the process proceeds to step ST303.
[0063]
[Step ST303]
Set the break position within the selected accent phrase. The delimiter position is set so that the number of mora in each unit (division unit) obtained by dividing the accent phrase by the delimiter position is equal to or less than the reference mora number. The break position is inserted at the boundary position of the word. Here, as shown in FIG. 12C, a separation position is set between the word “Joe Ho” and the word “Shori”, and between the word “Shori” and the word “Sochi”. The unit 1 “Joho”, unit 2 “Shori”, and unit 3 “Sochi” obtained by dividing the accent phrase at this break position have 4 mora, 2 mora, and 3 mora. It is less than the number of mora (4 mora).
[0064]
[Step ST304]
It is determined whether or not the determination in step ST302 has been made for all accent phrases of the phonetic symbol string obtained in step ST102. If the determination in step ST302 is made for all accent phrases, the process proceeds to step ST306. Otherwise, the process proceeds to step ST305, and after the next accent phrase is selected, the process returns to step ST302. Here, since all the accent phrases are determined in step ST302, the process proceeds to step ST306.
[0065]
[Step ST306]
The first accent phrase is selected from the phonetic symbol string obtained in step ST102. If a break position is set in the accent phrase in step ST303, the first division unit is selected. Here, the (division) unit 1 “Joe Ho” is selected.
[0066]
[Step ST307]
The prosodic search unit 102 searches for the corresponding prosodic pattern from the prosodic pattern database 44 using the selected accent phrase or the number of mora of the division unit and the accent type as search conditions. Here, as shown in FIG. 13, a 4-mora 0 type prosodic pattern corresponding to the unit 1 “Joe Ho” is searched.
[0067]
[Step ST308]
It is determined whether or not corresponding prosodic patterns have been searched for all accent phrases / dividing units. If all the items have been searched, the process proceeds to step ST108, and if not, the process proceeds to step ST309. Here, since the search for the unit 2 “Shori” and the unit 3 “Sochi” has not been performed yet, the process proceeds to Step ST309.
[0068]
[Step ST309]
The next accent phrase or division unit is selected, and the process returns to step ST307. Here, the (division) unit 2 “Shori” is selected.
[0069]
[Step ST307]
As shown in FIG. 13, a 2-mora 0 type prosodic pattern corresponding to the unit 2 “Shori” is searched.
[0070]
[Step ST308]
Since the search for the unit 3 “Sochi” has not been performed yet, the process proceeds to Step ST309.
[0071]
[Step ST309]
(Division) Unit 3 “Sochi” is selected.
[0072]
[Step ST307]
As shown in FIG. 13, a 3 mora type 1 prosodic pattern corresponding to the unit 3 “Sochi” is searched.
[0073]
[Step ST308]
Since prosodic patterns corresponding to all accent phrases and division units have been searched, the process proceeds to step ST108.
[0074]
[Step ST108]
As shown in FIG. 13, the prosodic search unit 102 connects the prosodic patterns searched in step ST307 and outputs prosodic information.
[0075]
[Step ST109]
The waveform generation unit 104 synthesizes a speech waveform (generates synthesized speech) according to the prosodic information generated by the prosody search unit 102.
[0076]
<Effect>
In the second embodiment, for accent phrases that exceed the number of reference mora, a delimiter position is set, and prosodic patterns are searched in units of division. Therefore, a matching prosodic pattern is always found in the prosodic database 44. This eliminates the need to generate prosody by alternative means. As a result, a synthesized sound having a high-quality prosody over the entire text to be read can be generated.
[0077]
(Third embodiment)
In the text-to-speech synthesizer according to the first embodiment, consider a case where a compound word that becomes one accent phrase is given as input text by accent combination. Here, as shown in FIG. 14, the text “European Front Victory Ceremony” is given. This text is a compound word composed of four words “Europe”, “front line”, “war” and “ceremony”. By accent concatenation, the phonetic symbol string of this text becomes one accent phrase of 16 mora 14 type. Then, as shown in FIG. 14, the 16-mora 14-type prosodic pattern is retrieved from the prosodic pattern database 44 to generate prosodic information. However, this prosodic information is not always a prosodic pattern in which the pronunciation of each word constituting the compound word of the text is easy to understand, such as the pronunciation of “aushoe”, “sensen”, “sensho”, and “shikiten”. Also, there are three consecutive “sen” sounds, and it is difficult to understand the word just by listening to the sound. If the prosodic patterns are classified according to the number of mora of the constituent words, the number of combinations increases in the order of the factorial of the number of mora, so the scale of the prosodic pattern database 44 becomes too large.
[0078]
Therefore, in the third embodiment, a table corresponding to a morpheme constituting a compound word when each word constituting the compound word becomes one accent phrase as a result of the accent combination with respect to the compound word part in the input text. The division position is set so that the phonetic symbol is a unit.
[0079]
The configuration of the text-to-speech synthesizer according to the third embodiment is the same as that of the apparatus according to the first embodiment. The procedure for generating synthesized speech by the text-to-speech synthesizer according to the third embodiment is performed according to the flowchart shown in FIG. Hereinafter, differences from the synthetic speech generation procedure according to the first embodiment will be described.
[0080]
[Step ST101]
The text to be read is entered. Here, as shown in FIG. 16A, the text “European Front Victory Ceremony” is given. This text is a compound word composed of four words “Europe”, “front line”, “war” and “ceremony”.
[0081]
[Step ST102]
The language analysis of the input text is performed by the language analysis unit 101. For accent phrases obtained by combining accents of words that make up a compound word, an accent zone delimiter is set at the boundary between the words that make up the compound word. That is, the accent phrase delimiter position is set for each pronunciation of each word constituting the compound word. Here, as shown in FIG. 16 (b), the four words "Europe", "front line", "war victory", and "ceremony" that compose the compound word are accent-coupled to form one accent phrase "Ousen Sensen Shoshi 'Kitten". Is formed. Therefore, the accent position delimiter positions are set between the words “Oshu” and “Sensen”, between the words “Sensen” and “Sensho”, and between the words “Sensho” and “Shikiten”.
[0082]
[Step ST103]
The prosodic search unit 102 selects the first accent phrase (in this case, the accent phrase “Oshu | sensen | sensho | shi'kiten”).
[0083]
[Step ST110]
It is determined whether or not the selected accent phrase is an accent phrase obtained by combining accents of words constituting the compound word. If the selected accent phrase is an accent phrase obtained by the accent combination of the words constituting the compound word, the process proceeds to step ST107, and if not, the process proceeds to step ST104. Since the accent phrase “Ohsue | Sensen | Sensho | Shikiten” is an accent phrase obtained by combining accents of words constituting a compound word, the process proceeds to step ST107.
[0084]
[Step ST107]
As shown in FIG. 16 (b), the prosody searching unit 102 obtains units (each word constituting the compound word) “Osho”, “Sensen”, “Sensho” obtained by dividing the accent phrase by the delimiter position “|”. , Search for a prosodic pattern corresponding to “Shikiten”.
[0085]
[Step ST108]
As shown in FIG. 16C, the prosodic search unit 102 connects the prosodic patterns searched in step ST107 and outputs prosodic information.
[0086]
<Effect>
In the third embodiment, prosodic patterns are searched and connected for each pronunciation of the compound word, and therefore the prosody that clearly shows the unit of pronunciation of each word even in the compound word part having a relatively long mora number as a result of the accent combination. It is possible to generate a synthesized sound with high intelligibility.
[0087]
(Fourth embodiment)
The configuration of the text-to-speech synthesizer according to the fourth embodiment is the same as that of the apparatus according to the first embodiment. The difference is that when searching for prosodic patterns in the prosodic search unit 102, whether or not the accent phrase is to be divided and searched at the division position in the accent phrase set by the language analysis unit 101 is given from the outside. It is a point determined by parameters. The pause frequency parameter is, for example, a value of 0 or more that changes continuously, and it means that the larger the value is, the more actively the pause is entered. When the pause frequency parameter is below a certain positive threshold, the prosody pattern is searched for by accent phrase, ignoring the division position in the accent phrase, and when the pause frequency parameter exceeds the threshold, it is divided by the division position within the accent phrase. Search for prosodic patterns.
[0088]
The procedure for generating synthesized speech by the text-to-speech synthesizer according to the fourth embodiment is performed according to the flowchart shown in FIG. Hereinafter, differences from the synthetic speech generation procedure according to the first embodiment will be described.
[0089]
The text “I think so” may be input (ST101). The phonetic symbol string “Saw / Omo'Ukamo | Silemasen” is obtained as a linguistic analysis result for the given text “Maybe” (ST102).
[0090]
In step ST111, the pause frequency parameter is compared with the threshold value. When (pause frequency parameter) ≦ (threshold value), the process proceeds to step ST104. When (pause frequency parameter)> (threshold value), the process proceeds to step ST111.
[0091]
1. When (pause frequency parameter) ≤ (threshold)
As shown in FIG. 18, a 2-mora 0 type prosodic pattern corresponding to accent phrase 1 “saw” is searched from prosodic pattern database 44 (ST104). The 10 mora type 2 prosodic pattern corresponding to the next accent phrase 2 “Omo'Ukamo | Sillemasen” is retrieved from the prosodic pattern database 44 (ST105, ST106, ST104). Prosodic information is generated by connecting these prosodic patterns (ST108). In this way, a prosody that utters an accent phrase in as long a unit as possible and does not pose much is generated.
[0092]
2. When (pause frequency parameter)> (threshold)
As shown in FIG. 19, a 2-mora 0 type prosodic pattern corresponding to accent phrase 1 “saw” is searched from prosodic pattern database 44 (ST107). For the next accent phrase 2 “Omo'Ukamo | Sillemasen”, it corresponds to the unit 5 “Siremasen”, a 5-mora type 2 prosodic pattern corresponding to unit 1 “Omo'Ukamo” obtained by dividing at the delimiter position. The 5-mora 0 type prosodic pattern is retrieved from the prosodic pattern database 44 (ST105, ST106, ST107). Prosodic information is generated by connecting these prosodic patterns (ST108). In this way, a prosody that utters accent phrases in small units and frequently poses is generated.
[0093]
<Effect>
According to the fourth embodiment, when reading at a normal speed or fast, the pause frequency parameter is reduced, the prosody generation unit is made as long as possible, and the prosody control is performed with less pitch rise. When reading out slowly, prosodic control with pitch rise can be performed even in an accent phrase. That is, appropriate prosody control can be performed according to the desired reading speed.
[0094]
(Fifth embodiment)
<Configuration of text-to-speech synthesizer>
The configuration of a text-to-speech synthesizer according to the fifth embodiment is shown in FIG. This apparatus includes a language analysis unit 101, a prosody generation unit 301, and a waveform generation unit 104.
[0095]
The prosody generation unit 301 generates prosody information according to rules using the mora number / accent type information for each accent phrase of the language analysis output output by the language analysis unit 101 or for each unit obtained by dividing the accent phrase. . Prosody generation according to the rules of the prosody generation unit 301 can be realized by, for example, the Fujisaki model.
[0096]
Such a speech synthesizer is constructed on a computer system as shown in FIG. 3 as in the first embodiment. The language analysis unit 101, prosody generation unit 301, and waveform generation unit 104 are connected to each other in a CD-ROM 508 set in the main unit 501, a disk (memory) 505 built in the main unit 501, or connected via a line 507. Stored in the system disk 506.
[0097]
<Synthetic voice generation procedure>
A procedure for generating synthesized speech by the text-to-speech synthesizer configured as described above will be described with reference to FIG.
[0098]
As shown in FIG. 22, the phonetic symbol string “Saw / Omo'Ukamo | Silemacene” is obtained as a language analysis result for the input text “Maybe” (ST101 to ST102).
[0099]
Prosody generation section 301 selects the first accent phrase (here, accent phrase 1 “saw”) (ST103). The prosody generation unit 301 selects the number of mora (2 mora) of the selected accent phrase 1 “saw” and the threshold (here, 5 Let's call it Mora. ) Is compared (ST301). Since the number of mora of the selected accent phrase 1 is less than or equal to the threshold value, the prosody generation unit 301 generates a prosody of the accent phrase 1 “saw” (2 mora type 0) based on the rules (ST302, see FIG. 22). .
[0100]
Prosody generation section 301 determines whether or not prosody has been generated for all accent phrases (ST304). Here, since the prosody for the accent phrase 2 “Omo'Ukamo | Sillemasen” has not yet been generated, the process proceeds to Step ST106.
[0101]
The prosodic search unit 102 selects the next accent phrase 2 “Omo'Ukamo | Sillemasen” (ST106). The prosody generation unit 301 selects the number of mora (10 mora) and the threshold value (10 mora) of the selected accent phrase 2 “Omo'Ukamo | Sillemasen”. 5 (Mora) is compared (ST301). Since the number of mora of the selected accent phrase 2 is larger than the threshold value, the prosody generation unit 301 obtains the unit 1 “Omo'Ukamo” (5 mora type 2) obtained by dividing the accent phrase 2 by the delimiter position “|”. And the prosody of the unit 2 “Shiramasen” (5 mora type 0) is generated based on the rules (see ST303, FIG. 22).
[0102]
When the prosody is generated for all accent phrases, as shown in FIG. 22, the prosody generation unit 301 connects the prosody generated in steps ST302 and ST303 and outputs prosodic information.
[0103]
<Effect>
According to the fifth embodiment, even in an accent phrase with a large number of mora, the division position is used as a pitch starting point when generating a prosody, so that it is possible to generate a prosody with a sense of clarity and a little uncomfortable feeling.
[0104]
(Sixth embodiment)
In the text-to-speech synthesizer according to the fifth embodiment, consider a case where a compound word that becomes one accent phrase is given as input text by accent combination. Here, as shown in FIG. 23, the text “European Front Victory Ceremony” is given. This text is a compound word composed of four words “Europe”, “front line”, “war” and “ceremony”. By accent concatenation, the phonetic symbol string of this text becomes one accent phrase of 16 mora 14 type. Then, as shown in FIG. 23, a 16-mora 14-type prosody is generated based on the rules. However, this prosody is not always a prosodic pattern in which the pronunciation of each word constituting the compound word of the text is easy to understand, such as “Ohsoo” “Sensen” “Sensho” “Shikiten”. Also, there are three consecutive “sen” sounds, and it is difficult to understand the word just by listening to the sound.
[0105]
Therefore, in the sixth embodiment, a table corresponding to a morpheme constituting a compound word when each word constituting the compound word becomes one accent phrase as a result of the accent combination with respect to the compound word part in the input text. The division position is set so that the phonetic symbol is a unit.
[0106]
The configuration of the text-to-speech synthesizer according to the sixth embodiment is the same as that of the apparatus according to the fifth embodiment. The procedure for generating synthesized speech by the text-to-speech synthesizer according to the sixth embodiment is performed according to the flowchart shown in FIG. Hereinafter, differences from the synthetic speech generation procedure according to the fifth embodiment will be described.
[0107]
[Step ST101]
The text to be read is entered. Here, as shown in FIG. 25, the text “European Front Victory Ceremony” is given. This text is a compound word composed of four words “Europe”, “front line”, “war” and “ceremony”.
[0108]
[Step ST102]
The language analysis of the input text is performed by the language analysis unit 101. For accent phrases obtained by combining accents of words that make up a compound word, an accent zone delimiter is set at the boundary between the words that make up the compound word. That is, the accent phrase delimiter position is set for each pronunciation of each word constituting the compound word. In this case, as shown in FIG. 25, the four words “Europe”, “front line”, “war” and “ceremony” constituting the compound word are accent-joined to form one accent phrase “Ousen Sensen Shoshi Kiten”. Is done. Therefore, the accent position delimiter positions are set between the words “Oshu” and “Sensen”, between the words “Sensen” and “Sensho”, and between the words “Sensho” and “Shikiten”.
[0109]
[Step ST103]
The prosody generation unit 301 selects the first accent phrase (in this example, the accent phrase “Oshu | sensen | sensho | shi'kiten”).
[0110]
[Step ST110]
It is determined whether or not the selected accent phrase is an accent phrase obtained by combining accents of words constituting the compound word. If the selected accent phrase is an accent phrase obtained by accent concatenation of the words constituting the compound word, the process proceeds to step ST303; otherwise, the process proceeds to step ST302. Since the accent phrase “Ohsue | Sensen | Sensho | Shikiten” is an accent phrase obtained by combining accents of words constituting a compound word, the process proceeds to step ST303 here.
[0111]
[Step ST303]
As shown in FIG. 25, the prosody generation unit 301 is a unit (each word constituting a compound word) obtained by dividing an accent phrase by a delimiter position “|”, “Ohshou”, “Sensen”, “Sensho”, “Shisho” Prosody of 'Kiten' is generated based on the rules.
[0112]
[Step ST108]
As shown in FIG. 25, the prosody generation unit 301 connects the prosody generated in step ST303 and outputs prosodic information.
[0113]
<Effect>
In the sixth embodiment, a prosody is generated and connected for each pronunciation of a compound word. As a result, even in an accent phrase formed by combining accents of compound words, it is possible to control the subtle pitch start at the division position, so that synthesized sounds with clear pronunciation for each word constituting the compound word can be obtained. Can be generated.
[0114]
(Seventh embodiment)
The configuration of the text-to-speech synthesizer according to the seventh embodiment is the same as that of the device according to the fifth embodiment. The difference is that, when the prosody generation unit 301 generates a prosody, whether or not to generate a prosody by dividing the accent phrase at the division position in the accent phrase set by the language analysis unit 101 is a pose frequency parameter given from the outside. It is a point determined by. The pause frequency parameter is, for example, a value of 0 or more that changes continuously, and it means that the larger the value is, the more actively the pause is entered. When the pause frequency parameter is less than a certain positive threshold, the prosody is generated in the accent phrase unit ignoring the division position in the accent phrase, and when it exceeds the threshold, the prosody is divided in the unit divided at the accent phrase. Is generated.
[0115]
The procedure for generating synthesized speech by the text-to-speech synthesizer according to the seventh embodiment is performed according to the flowchart shown in FIG. Hereinafter, differences from the synthetic speech generation procedure according to the fifth embodiment will be described.
[0116]
The text “I think so” may be input (ST101). The phonetic symbol string “Saw / Omo'Ukamo | Silemasen” is obtained as a linguistic analysis result for the given text “Maybe” (ST102).
[0117]
In step ST111, the pause frequency parameter is compared with the threshold value. When (pause frequency parameter) ≦ (threshold value), the process proceeds to step ST302. If (pause frequency parameter)> (threshold value), the process proceeds to step ST303.
[0118]
1. When (pause frequency parameter) ≤ (threshold)
As shown in FIG. 27, the prosody of the accent phrase 1 “saw” (2 mora 0 type) is generated by the prosody generation unit 301 (ST302). The prosody of the next accent phrase 2 “Omo'umo | Sillemacene” (10 mora type 2) is generated by the prosody generation unit 301 (ST304, ST106, ST111, ST302). The prosody generation unit 301 connects these prosody and generates prosodic information (ST305). In this way, a prosody that utters an accent phrase in as long a unit as possible and does not pose much is generated.
[0119]
2. When (pause frequency parameter)> (threshold)
As shown in FIG. 28, the prosody of accent phrase 1 “saw” (2 mora 0 type) is generated by the prosody generation unit 301 (ST303). For the next accent phrase 2 “Omo'Umo | Sillemasen”, the prosody of unit 1 “Omo'Umomo” (5 mora type 2) obtained by dividing at the delimiter position, unit 2 “Siremasen” (5 mora 0 type) ) Is generated by the prosody generation unit 301 (ST304, ST106, ST111, ST303). The prosody generation unit 301 connects these prosody and generates prosodic information (ST305). In this way, a prosody that utters accent phrases in small units and frequently poses is generated.
[0120]
<Effect>
According to the seventh embodiment, when reading at a normal speed or faster, the pause frequency parameter is reduced, the prosody generation unit is lengthened as much as possible, and the prosody control is performed with less pitch rise. When reading out slowly, prosodic control with pitch rise can be performed even in an accent phrase. That is, appropriate prosody control can be performed according to the desired reading speed.
[0121]
【The invention's effect】
In the accent phrase of the language analysis result, the division position in the accent phrase is set so that the prosodic pattern search unit can be appropriately selected according to the condition. As a result, it is possible to generate a certain high-quality synthesized sound by making maximum use of the prosodic patterns in the prosodic pattern database and without requiring prosody generation by an alternative means of prosodic pattern search. In addition, even when prosody generation is performed according to rules, it is possible to generate clear and easy-to-understand synthesized sounds even for accent phrases with a long mora number. Prosodic control can be appropriately performed according to the speed of reading.
[Brief description of the drawings]
FIG. 1A is a diagram for explaining inconveniences in a conventional method.
FIG. 1B is a diagram for explaining inconveniences in a conventional method.
FIG. 2 is a block diagram showing a configuration of a text-to-speech synthesizer according to the first embodiment.
3 is a diagram showing a configuration of a computer system that realizes the text-to-speech synthesizer shown in FIG. 1. FIG.
FIG. 4 is a flowchart showing a procedure for generating synthesized speech by the text-to-speech synthesizer shown in FIG. 1;
FIG. 5 is a diagram illustrating an example of an input text and a phonetic symbol string.
FIGS. 6A and 6B are diagrams showing a state in which prosodic patterns are searched for each accent phrase. FIGS.
FIGS. 7A and 7B are diagrams showing a state in which prosodic patterns are searched for each unit divided at an accent phrase delimiter position; FIGS.
FIG. 8 is a diagram illustrating a state in which prosodic information is generated by connecting searched prosodic patterns.
FIG. 9 is a flowchart showing a procedure for setting the number of reference mora.
FIG. 10 is a diagram showing an example in which prosodic patterns in a prosodic pattern database are grouped.
FIG. 11 is a flowchart showing a procedure for generating synthesized speech according to the second embodiment.
FIG. 12A shows an example of input text. (B) shows an example of a phonetic symbol string. (C) shows an example of setting of the position within the accent.
FIG. 13 is a diagram illustrating a state in which prosodic information is generated by connecting searched prosodic patterns.
FIG. 14 is a diagram showing a procedure for generating prosodic pattern information for an input text of a compound word.
FIG. 15 is a flowchart showing a procedure for generating synthesized speech according to the third embodiment.
FIG. 16A shows an example of input text. (B) shows an example of a phonetic symbol string obtained as a result of language analysis. (C) shows an example of the prosodic information to be generated.
FIG. 17 is a flowchart showing a synthetic speech generation procedure according to the fourth embodiment.
FIG. 18 is a diagram showing how prosodic information is generated when (pause frequency parameter) ≦ (threshold value).
FIG. 19 is a diagram showing how prosodic information is generated when (pause frequency parameter)> (threshold value).
FIG. 20 is a block diagram showing a configuration of a text-to-speech synthesizer according to a fifth embodiment.
FIG. 21 is a flowchart showing a procedure for generating synthesized speech by the text-to-speech synthesizer shown in FIG. 20;
FIG. 22 is a diagram showing examples of text, phonetic symbol strings, prosody, and prosody information.
FIG. 23 is a diagram showing a procedure for generating a prosody of an input text of a compound word.
FIG. 24 is a flowchart showing a synthetic speech generation procedure according to the sixth embodiment.
FIG. 25 is a diagram showing examples of text, phonetic symbol strings, prosody, and prosody information.
FIG. 26 is a flowchart showing a synthetic speech generation procedure according to the seventh embodiment.
FIG. 27 is a diagram showing how prosodic information is generated when (pause frequency parameter) ≦ (threshold value).
FIG. 28 is a diagram showing how prosodic information is generated when (pause frequency parameter)> (threshold value).

Claims (5)

A language analysis unit for outputting an accent phrase having at least one clause boundary or word boundary from the input text;
A prosody search unit for searching for a prosodic pattern corresponding to the accent phrase from a prosodic pattern database storing prosodic patterns extracted from real sounds and generating prosodic information;
A waveform generation unit that synthesizes a speech waveform from the prosodic information,
The prosody search part
Search prosodic patterns corresponding to the accent phrase from the prosodic pattern database, and for accent phrases that do not have a corresponding prosodic pattern, divide the accent phrase at the clause boundary or word boundary, and search for the prosodic pattern in divided units. To generate prosodic information,
A speech synthesizer characterized by the above. In claim 1,
The prosody search part
If the number of mora of the accent phrase output by the language analysis unit exceeds the reference value, the accent phrase is divided at the clause boundary or word boundary, and prosodic information is generated by searching the prosodic pattern in the divided unit. And
The reference value is
Among the groups for each number of mora stored in the prosodic pattern database, the number of mora of the group having the largest number of mora among the groups in which all accent types exist.
A speech synthesizer characterized by the above. In claim 1,
The word boundary is
A word boundary in a compound word in which a plurality of words are consecutive,
The prosody search part
Dividing into each of the words included in the compound word, and searching for prosodic patterns corresponding to the divided words to generate prosodic information,
A speech synthesizer characterized by the above. In claim 1,
Before Kia accent clause,
An accent phrase that is selected when the set pause frequency exceeds the reference value.
A speech synthesizer characterized by the above. A language analysis step of outputting an accent phrase having at least one clause boundary or word boundary from the input text;
A search step for searching a prosodic pattern corresponding to the accent phrase from a prosodic pattern database storing prosodic patterns extracted from real sounds;
For the accent phrase in which the corresponding prosodic pattern exists in the search step, the prosody is generated by the corresponding prosodic pattern, and for the accent phrase in which the corresponding prosodic pattern does not exist in the search step, the accent phrase at the phrase boundary or word boundary A prosody generation step of generating prosodic information by searching prosodic patterns in the divided units,
A waveform generation step of synthesizing a speech waveform from the prosodic information,
A speech synthesis method characterized by the above.

Images