JP2012042974A - Voice synthesizer - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a high-quality voice synthesizer which generates voice data of a sentence including a fixed portion and a variable portion by combining a recorded voice and a regularly synthesized voice, the voice synthesizer solving the problem in which, when a recorded voice is connected with a synthesized voice, the discontinuity of sound quality and rhythm is perceived.SOLUTION: A voice synthesizer comprises: a recorded voice storing unit 5 for pre-storing recorded voice data including a fixed portion; a regularly synthesizing unit 7 for generating regularly synthesized voice data including a variable portion and at least part of the fixed portion, from an accepted text; a connection boundary calculating unit 8 for calculating a connection boundary position in a section where the recorded voice data is overlapped with the regularly synthesized voice data, based on audio feature information of the recorded voice data and the regularly synthesized data corresponding to the text; and a connection synthesizing unit 9 for generating synthesized voice data corresponding to the text by connecting the recorded voice data divided at the connection boundary position and extracted with the regularly synthesized voice data.

Description

  The present invention relates to an apparatus for synthesizing speech, and more particularly, to a speech synthesis technique for synthesizing speech data of a sentence composed of a fixed part and a variable part by combining recorded speech and rule-synthesized speech.

  In general, a recorded voice refers to a voice created based on a recorded voice, and a rule-synthesized voice refers to a voice synthesized from a character or a code string representing pronunciation. In speech synthesis, the input text is subjected to language processing, an intermediate symbol string indicating reading and accent information is generated, and then the fundamental frequency pattern (vibration frequency corresponding to the pitch of the voice) and A prosody parameter such as a phoneme duration (the length of each phoneme corresponding to the utterance speed) is determined, and a speech waveform matching the prosody parameter is generated by waveform generation processing. As a method for generating a speech waveform from prosodic parameters, waveform-connected speech synthesis, in which speech segments corresponding to phonemes and syllables are combined, is widely used.

  The general rule composition flow is as follows. First of all, in language processing, a sequence of phonemes (minimum unit for discriminating the meaning of speech) and syllables (a kind of speech audible unit consisting of a combination of 1 to 3 phonemes) is arranged from the input text. Accent information that expresses reading information to be expressed, accent (information specifying the strength of pronunciation) and inflection (information indicating question sentence or speaker's emotion) is generated, and this is used as an intermediate symbol string. To generate the intermediate symbol string, language processing using a dictionary or morpheme analysis processing is applied. Next, prosodic parameters such as a fundamental frequency pattern and a phoneme duration are determined so as to correspond to the accent information of the intermediate symbol string. The prosodic parameters are generated based on a prosodic model learned in advance using the real voice and heuristics (control rules determined heuristically). Finally, a speech waveform that matches the prosodic parameters is generated by waveform generation processing.

  In the rule synthesis, any input text can be output as speech, so that a highly flexible speech guidance system can be constructed compared to the case of using recorded speech. However, the quality is not sufficient compared to the real voice, and there has been a problem in terms of quality in order to introduce rule-synthesized voice into a voice guidance system such as an in-vehicle car navigation system that uses recorded voice.

  Therefore, in order to realize a voice guidance system using rule-synthesized voice, the recorded voice recorded in advance is used for the standard part and the rule-synthesized voice is used for the variable part. A method that combines the flexibility of speech is used.

  However, the sound that is output by combining the recorded voice and the regular synthesized voice is perceived as a discontinuity in sound quality and prosody between the recorded voice and the regular synthesized voice. There was a problem of not being high quality.

  As a method for eliminating discontinuity of prosody, a method is disclosed that uses the characteristics of recorded speech when setting parameters for rule-synthesized speech (see, for example, Patent Document 1). In addition, a method of extending the rule-synthesized speech part in consideration of the continuity of the prosody of the fixed part and the variable part is disclosed (for example, see Patent Document 2).

JP-A-11-249677 JP 2005-321520 A

  According to the prior art, the prosody of the regular synthesized speech part becomes natural, but on the other hand, the difference in the sound quality between the regular synthesized speech and the recorded speech may become large, and it is difficult to obtain a natural speech as a whole. It has a problem that it cannot be done.

  An object of the present invention is to solve the above problems, and to provide a high-quality speech synthesizer in which discontinuity of sound quality and prosody is not perceived when connecting recorded speech and synthesized speech.

  To achieve the above object, according to the present invention, there is provided a speech synthesizer that synthesizes text composed of a fixed part and a variable part, and is first speech data including a fixed part created based on recorded speech. A recording voice storing means for previously storing voice data (recorded voice data), and a rule for generating second voice data (rule synthesis voice data) including a variable part and at least a part of a fixed part from the received text Based on acoustic feature information of a section where the first voice data and the second voice data corresponding to the text overlap, the recorded voice data and voice data generated by rule synthesis corresponding to the text A connection boundary calculation means for selecting a connection boundary position, third audio data obtained by dividing the first audio data by the connection boundary, and the second audio data. Connects the fourth audio data cut out, separated by serial connection boundary, characterized in that it comprises a connection synthesizing means for synthesizing speech data of the text. Here, as an example, the standard part can be defined as a part having a part corresponding to audio data, and the variable part can be defined as a part having no part corresponding to audio data.

  In this configuration, the rule-synthesized voice data is generated to include a part of the standard part in addition to the variable part, and the recorded voice and the rule synthesis are created by creating an overlapping section between the rule-synthesized voice data and the recorded voice data. The voice connection position can be made variable. By calculating the optimum connection position using the acoustic feature information of the recorded voice and the rule-synthesized voice in the overlapping section, a natural synthesized voice is generated as compared with the conventional technique.

  In another configuration of the present invention, rule synthesis means for generating rule-synthesized voice data that matches the recorded voice data using the acoustic feature information of the recorded voice data in the overlapping section is provided.

  In this configuration, the prosody discontinuity can be eliminated by matching the prosody in the overlapping section, and the variable synthesized speech data of the variable part preceding or following the overlapping section is also matched at the same time. Thus, a synthesized speech in which not only the connection boundary but also the entire match is obtained is generated.

  Further, in another configuration of the present invention, rule synthesis means for processing the rule synthesized voice data based on the acoustic feature information of the recorded voice data and the rule synthesized voice data at the position of the connection boundary obtained from the connection boundary calculation means. Prepare.

  In this configuration, after determining the connection boundary, the discontinuity of prosody and sound quality is further inconspicuous by processing the characteristics of the regularly synthesized voice data so that the acoustic characteristics near the connection boundary are closer to the recorded voice. A synthesized speech is generated.

  By using a phoneme category as acoustic feature information in the present invention, a suitable connection boundary can be obtained. The phoneme category is information defining the classification of phonemes, such as voiced / unvoiced / plosive / frictional sounds. It goes without saying that the connection distortion becomes inconspicuous by connecting in the pause (silence) section, but the connection distortion is not conspicuous because there is also a short silence section at the head of the unvoiced plosive sound. In addition, connection in a voiced sound section is desirable because connection may be noticeable due to differences in fundamental frequency and phase differences before and after the connection boundary. Further, by using power as acoustic feature information, a connection boundary having a small power can be selected, and connection distortion can be made inconspicuous.

  Further, when the fundamental frequency is used as the acoustic feature information, a connection boundary with a smooth prosodic connection can be obtained. By selecting a phoneme boundary where the difference between the fundamental frequencies of the recorded speech and the regular synthesized speech is small, the discontinuity of the fundamental frequency becomes difficult to perceive. Further, when the phoneme duration is used, it is possible to select a connection boundary so that the speech speed does not change suddenly before and after the connection boundary.

  Further, if a spectrum (information indicating the frequency component of speech) is used as the acoustic feature information, it is possible to avoid a sudden change in sound quality near the connection boundary. This is particularly effective in the case of a configuration in which the characteristics of rule-synthesized speech data are processed using acoustic feature information in the vicinity of the connection boundary after the connection boundary is determined. It can be processed to be closer.

  In the present invention, the range for generating the regularly synthesized voice data is to include a part of the fixed part in addition to the variable part, but this range is one expiratory paragraph (one unit divided by a pause for breathing). It is desirable to define it in one of a sentence (one unit divided by a punctuation mark) or a whole fixed part. In particular, in order to match the prosody of the recorded voice and the rule-synthesized voice, it is preferable that the overlapping section is large. However, when a prosody matching method by another means can be used or when there is a problem from the viewpoint of calculation amount, it may be defined to be within a range of one exhalation paragraph.

  In the connection boundary calculation means of the present invention, the connection boundary candidate positions are all sample points in the overlapping section. However, if a connection boundary is selected only for the phoneme boundary, an effective connection boundary can be obtained. . By adopting such a configuration, it is sufficient that the acoustic feature information of the recorded voice and the rule-synthesized voice is calculated only at the phoneme boundary, which is advantageous in terms of storage capacity and calculation amount.

  In the recorded voice storage means of the present invention, by storing voice data recorded in advance in one breath paragraph or one sentence unit, including a fixed part and a part other than the fixed part, a section other than the fixed part in the recorded voice Can also be used effectively. If the recorded text is determined according to the text of the variable part when the text of the standard part is set in advance, if the recorded voice can be used for a part of the variable part, the variable part is used as the overlapping section. Can be included. In this way, many parts of the recorded voice can be utilized, and a higher quality synthesized voice can be generated.

  Furthermore, the speech synthesizer of the present invention is a speech synthesizer that synthesizes text composed of a fixed part and a variable part, and has received a recorded voice storage unit that stores recorded voice data including the fixed part in advance. From the text, a rule synthesis unit that generates the rule synthesis voice data including the variable part and at least a part of the fixed part, and acoustic feature information of the recorded voice data and the rule synthesis voice data corresponding to the text A connection boundary calculation unit for calculating a connection boundary position in a section in which the recorded voice data and the rule-synthesized voice data overlap; and the recorded voice data and the rule-synthesized voice data cut out by being separated at the connection boundary position And a connection synthesis unit that generates synthesized speech data corresponding to the text.

  Furthermore, the speech synthesizer of the present invention is a speech synthesizer that synthesizes text composed of a fixed part and a variable part, and a recorded voice storage unit that stores recorded voice data including the fixed part in advance. A rule synthesis parameter calculating unit that calculates a rule synthesis parameter including the variable part and at least a part of the fixed part from the text, and generates acoustic feature information of the rule-synthesized speech; and acoustic feature information of the recorded voice; A connection boundary calculation unit that calculates a connection boundary position in a section where the recorded voice data and the rule synthesis parameter overlap using the acoustic characteristic information of the rule synthesis, the acoustic feature information of the recorded voice, and the rule Using the acoustic feature information of synthesized speech and the connection boundary position, a rule-synthesized speech data unit that generates rule-synthesized speech data, and A connection synthesizer for connecting the recorded voice data and the rule synthesized voice data cut out by dividing at a continuous boundary position and outputting synthesized voice data corresponding to the text; and means for outputting the synthesized voice data It is characterized by providing.

  Furthermore, the speech synthesizer of the present invention is a device that creates a synthesized speech by connecting a pre-recorded sound piece including a variable part and a sound piece including a fixed part, from the pre-recorded sound piece. A voice recording unit that stores voice data, an input analysis unit that creates an intermediate symbol string of the variable piece and an intermediate symbol string of the fixed part from the received input text, and the variable A recording voice selection unit that selects appropriate recording voice data from among a plurality of recording voice data having the same fixed part in accordance with the input of the part; and an intermediate symbol string of the variable pieces obtained by the input analysis unit; A rule synthesizing unit for determining a range in which the rule synthesized voice data is generated using an intermediate symbol string of the fixed-part sound pieces; acoustic feature information of the recorded voice data and the rule synthesized voice data Using acoustic feature information, a connection boundary calculation unit that calculates a connection boundary position in an overlapping section between the recorded voice data and the rule-synthesized voice data, and using the connection boundary position obtained from the connection boundary calculation unit Cutting the recorded voice data and the rule synthesized voice data, and connecting the cut recorded voice data and the rule synthesized voice data to obtain synthesized voice data corresponding to the sound piece including the variable portion. It has a connection composition part to create and a sound piece connection part which connects sound pieces and generates output speech based on the order of sound pieces obtained from input text.

  In the speech synthesis method of the present invention, the recorded voice data and the first intermediate symbol string corresponding to the recorded voice data are stored in advance, the first step of preparing the input text, and the input text as the second text. A second step of converting to an intermediate symbol string; a third step of referring to the first intermediate symbol string and discriminating the second intermediate symbol string into variable parts not corresponding to the fixed part corresponding to the first intermediate symbol string The fourth step of acquiring from the recorded voice data a portion where the first intermediate symbol string corresponds to the fixed portion, and all the portions corresponding to the variable portion using the second intermediate symbol sequence and the fixed portion A fifth step of generating at least a part of the synthesized speech data of a part corresponding to the above, a sixth step of combining a part of the acquired recorded voice data and a part of the generated ruled synthesized voice data; Have

  Here, the acquired recorded voice data and the generated rule-synthesized voice data can each be one continuous phrase, and the two phrases have overlapping parts, so the degree of freedom of the parts to be joined is large. , Can be combined with natural connections. That is, since the two audio data have a section that overlaps with the fixed portion, a portion where the two audio data matches in this section may be selected as a connection boundary and connected. As an evaluation criterion for which part is matched, for example, a portion having a small difference in feature quantity such as a fundamental frequency, a spectrum, and a duration of two audio data can be selected. Further, if necessary, one of the two data can be corrected (processed) and connected. For example, the acoustic features can be matched by correcting the parameters when generating the regularly synthesized speech data so that the difference between the feature amounts of the recorded speech data and the regularly synthesized speech data becomes small.

  According to the present invention, it is possible to realize a high-quality speech synthesizer in which sound quality and prosody discontinuity are not perceived when recording speech and synthesized speech are connected.

It is a block diagram which shows the structure of the speech synthesizer in 1st Example of this invention. It is a flowchart which shows operation | movement of the speech synthesizer in a 1st Example. It is a figure which shows the information memorize | stored in the sound recording audio | voice storage part in a 1st Example. It is a figure which shows the specific example of the information memorize | stored in the sound recording audio | voice storage part in a 1st Example. It is explanatory drawing for demonstrating the production | generation method of the rule synthetic speech in a 1st Example. It is explanatory drawing for demonstrating the selection method of the connection boundary position in a 1st Example. It is a block diagram which shows the structure of the speech synthesizer in 2nd Example of this invention. It is a flowchart which shows operation | movement of the speech synthesizer in a 2nd Example. It is a block diagram which shows the structure of the speech synthesizer in the 3rd Example of this invention. It is a flowchart which shows the operation | movement of the speech synthesizer in a 3rd Example. It is a figure which shows the structure of the input screen in a 3rd Example. It is a figure which shows the information memorize | stored in the sound piece information storage part in a 3rd Example. It is a figure which shows the specific example of the information memorize | stored in the sound piece information storage part in a 3rd Example. It is explanatory drawing for demonstrating the selection method of the connection boundary position in a 3rd Example.

  Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings.

Example 1
FIG. 1 is a block diagram showing a speech synthesizer of the present invention configured for a car navigation system according to a first embodiment of the present invention.
As shown in the figure, this embodiment includes a speech synthesizer 1 and a navigation control device 2. The speech synthesizer 1 of the present invention uses an input analysis unit 4 that analyzes text input from the navigation control unit 3 and a recorded voice from the recorded speech storage unit 5 using an intermediate symbol string of a fixed part obtained by the input analysis unit 4. Recorded sound selection unit 6 for searching for data, variable part intermediate symbol string obtained by input analysis unit 4 and part of fixed part intermediate symbol string, and acoustic feature information of recorded sound obtained by recorded voice selection unit 6 Using the rule synthesis unit 7 that generates the rule synthesis voice data using the, and the sound feature information of the recorded voice obtained by the recording voice selection unit 6 and the sound feature information of the rule synthesis voice obtained by the rule synthesis unit 7, A connection boundary calculation unit 8 that calculates a connection boundary between the recorded voice data and the rule synthesized voice data, and a connection that cuts out and connects the recorded voice data and the rule synthesized voice data using the connection boundary obtained by the connection boundary calculation unit And a generation unit 9.

  Next, the operation of the speech synthesizer 1 according to the first embodiment of the present invention will be described using FIG. 1 and FIG. FIG. 2 is a flowchart showing the operation of the speech synthesizer 1 according to the first embodiment of the present invention.

  First, the navigation control unit 2 determines an input text to be passed to the speech synthesizer 1.

  The navigation control unit 3 receives various information such as weather forecasts and traffic information from the information receiving unit 10 and combines it with the current position information obtained from the GPS 11 and the map information of the navigation data storage unit 12. An input text to be passed to the speech synthesizer 1 is created (step 101).

  Next, the input analysis unit 4 receives the input text for voice output from the navigation control unit 2 and converts it into an intermediate symbol string (step 102). The input text is, for example, a character string mixed with kanji characters such as “Kokubunji tomorrow's weather”. Here, the input analysis unit 4 performs language processing and converts it into an intermediate symbol string for speech synthesis such as “Kokubunzino Asitano Tenchides”.

  Next, the input analysis unit 4 refers to the intermediate symbol string 402 stored in the recorded voice storage unit 5 in association with the recorded voice data 401 shown in FIG. An intermediate symbol string to be determined is determined, and a portion that cannot be associated with the speech waveform data 401 is determined as a variable portion (step 103).

  As described above, a plurality of sets of intermediate symbol strings 402 associated with the recorded voice data 401 are stored in the recorded voice storage unit 5 with the configuration shown in FIG. Here, as shown in FIG. 4, the operation of step 103 will be described by taking as an example the case where the intermediate symbol string “Shinjukuno Ashitano Tenchides” is stored in the recorded voice storage unit 5.

When the intermediate symbol string “Kokubunji no Ashitano Tenchides” obtained from the input analysis unit 4 and the intermediate symbol sequence 402 stored in the recorded voice storage unit 5 are sequentially compared, the “No. Ashitano Tenchides” Therefore, the recorded voice data 401 can be used with the corresponding part as a fixed part. Therefore, “Noashitan Tenkides” is determined as the standard part, and “Kokubunji” that cannot be associated with the recorded voice data is determined as the variable part.

  Next, the recorded voice selection unit 6 acquires the recorded voice data 401 and the acoustic feature information 403 of the recorded voice (step 104).

The recorded voice selection unit 6 acquires the recorded voice data 401 from the recorded voice storage unit 5 using the intermediate symbol string of the standard part obtained by the input analysis unit 4. Here, even if the intermediate symbol string of the standard part is “Noashitan Tenchides”, at least one of the recorded voice data before and after the intermediate symbol string is acquired together. Here, as an example, it is assumed that the entire recorded audio data corresponding to “Shinjukuno Ashitano Tenchides” is acquired. The process of cutting out only the part corresponding to the fixed part is not performed here.

  Also, acoustic feature information 403 stored in the recorded voice storage unit 5 in association with the recorded voice data 401 is acquired. The acoustic feature information is stored in a configuration as shown in the example of FIG. 4, and for each phoneme of the recorded voice, the phoneme category, the start / end time, and the fundamental frequency are described.

  The rule synthesizing unit 7 determines the range for generating the rule synthesized speech using the intermediate symbol string of the variable part and the intermediate symbol string of the fixed part obtained by the input analyzing unit 4 (step 105). Here, if the range for creating a rule-synthesized speech is defined as a sentence containing a variable part, a rule-synthesized speech is generated that includes the fixed part “Noashitta Tenchides” in addition to the variable part “Kokubunji”. .

  Next, the rule synthesizing unit 7 refers to the acoustic feature information 403 of the recorded voice and generates rule synthesized voice data (step 106). Here, the rule synthesis parameters such as the fundamental frequency and the phoneme duration are calculated using the prosodic model 13 stored in advance by the rule synthesis unit 7. At this time, referring to the acoustic feature information of the recorded voice By modifying the rule synthesis parameters, it is possible to generate rule synthesis voice data that can be easily connected to the recorded voice.

  FIG. 5 shows how the rule synthesis parameters are determined using the fundamental frequency information in the acoustic feature information 403 of the recorded voice. As shown in FIG. 5, in a section 501 where the recorded voice data and the generated rule synthesized voice data overlap, the error from the acoustic feature information (basic frequency pattern of the recorded voice) 502 of the recorded voice data is reduced. The rule synthesis parameter (basic frequency pattern set by the prosody model) 503 calculated using the prosody model 13 is modified to generate acoustic feature information (modified fundamental frequency pattern) 504 of the rule synthesized speech data. As a correction method, operations such as parallel movement and expansion or reduction of the dynamic range are used.

  As described above, the operation for matching the acoustic features is performed in the section 501 where the recorded voice data and the rule-synthesized voice data overlap, and the same operation is performed on the variable portion 505 that does not overlap with the recorded voice data. It becomes possible to match the prosody of the part and the fixed part.

  The acoustic feature information is used not only for the fundamental frequency but also for the long phoneme duration, thereby eliminating the rhythm inconsistency between the recorded voice data and the rule-synthesized voice data. Further, the spectrum information of the recorded voice can be used as the acoustic feature information, and the discontinuity between the recorded voice data and the rule-synthesized voice data can be eliminated in terms of sound quality.

  Next, the connection boundary calculation unit 8 uses the acoustic feature information 502 of the recorded voice data and the acoustic feature information 504 of the regular synthesized voice data in the overlapping section 501 of the recorded voice data and the regular synthesized voice data in FIG. Is calculated (step 107). A calculation method when a fundamental frequency is given as acoustic feature information in the overlapping section 501 between the recorded voice data and the rule synthesized voice data will be described with reference to FIG.

  First, using phoneme category information, an unvoiced sound segment in a word, such as the head of an unvoiced plosive, is selected as a connection boundary candidate. Subsequently, a difference between the fundamental frequencies of the recorded speech and the rule-synthesized speech in the phoneme boundary candidate is calculated, and a connection boundary candidate is selected if the difference is small. At this time, when there are a plurality of calculated equivalent candidates, the connection boundary position 601 is determined in consideration of shortening the section of the regularly synthesized speech data.

  When obtaining connection boundary candidates using phoneme category information, the head position of the unvoiced plosive sound is effective, but other unvoiced sounds can be smoothly connected as compared to voiced sounds. However, when a cross fading can be used as the connection method in the connection synthesis unit 9, there is a possibility that a smooth connection can be made even in voiced sound. It is not limited to.

  In addition to using the fundamental frequency difference as the acoustic feature information for calculating the connection position, it is possible to calculate a smoother position at the time of connection by using the phoneme duration difference and the spectrum difference together. It becomes possible.

The connection boundary calculation unit 8 not only calculates the connection boundary in the order of calculating the fundamental frequency difference after narrowing down the candidates with the phoneme category information as in the above example, but also shows the following (Equation 1 It is also possible to define and calculate a cost function as shown in the example.
C (b) = Wp × Cp (b) + Wf × Cf (b) + Wd × Cd (b) + Ws × Cs (b) + Wl × Cl (b)
・ ・ ・ ・ ・ ・ ・ (Equation 1)
Here, the difficulty of connection determined from the phoneme category information is defined as the phoneme category cost Cp (b), and the weight is defined as Wp. Also, differences in acoustic feature information are also defined as fundamental frequency cost Cf (b), phoneme duration cost Cd (b), and spectrum cost Cs (b), respectively, and their weights are Wf, Wd, Ws, respectively. And Further, the time difference between the boundary between the variable part and the fixed part is obtained from each phoneme boundary position, defined as the rule synthesis speech length cost Cl (b), and the weight is defined as Wl. It is also possible to calculate the cost C (b) related to the connection boundary position as a weighted sum of the respective costs, and to set the phoneme boundary having the smallest cost as the connection boundary position.

  Next, the connection synthesis unit 9 cuts the recorded voice data and the rule synthesized voice data using the connection boundary position obtained from the connection boundary calculation unit 8, and connects the cut recorded voice data and the rule synthesized voice data. Thus, synthesized speech data corresponding to the input text is output (step 108). Here, the connection boundary position is calculated as the time in the recorded voice data and the time in the rule synthesized voice data, and the voice data is disconnected and connected using the calculated time.

  When connecting the disconnected voice, the connection synthesis unit 9 can not only make the connection but also make the connection portion inconspicuous by using the cross-fade process. In particular, when a connection is made in the middle of a voiced part, the crossfade process is performed only for one fundamental period of the speech waveform at the connection boundary position in synchronization with the fundamental frequency to eliminate abnormal noise during connection. Can do. However, since there is a possibility that the signal is deteriorated by using the cross-fade processing, it is desirable to determine the connection boundary position so as to avoid the connection in the middle of the voiced portion.

  In the above-described embodiment, the range in which the rule-synthesized voice data is created is defined as one sentence including a variable part. However, the range may be generated in one unit of one breath paragraph or one sentence.

  As described above, in the first embodiment, in the speech synthesizer configured for an in-vehicle car navigation system and connecting the recorded voice data and the rule synthesized voice data, the sound quality and prosody of the rule synthesized voice data are recorded voice data. It is possible to generate a natural synthesized speech by calculating a suitable connection boundary.

(Example 2)
Next, a second embodiment of the present invention will be described.

  In the first embodiment, the recorded voice data and the rule synthesized voice data are connected using the connection boundary position determined after the rule synthesized voice data is generated. After the connection boundary position is determined, the rule synthesis is performed. It is good also as a structure which produces | generates audio | voice data.

  FIG. 7 is a block diagram showing a second embodiment of the present invention. In the second embodiment, a rule synthesis parameter calculation unit 21 and a rule synthesis voice data generation unit 22 are provided instead of the rule synthesis unit 7 in the first embodiment. FIG. 8 is a flowchart showing the operation of the speech synthesizer 20 according to the second embodiment. The operation of the speech synthesizer 20 according to the second embodiment will be described with reference to FIGS.

  First, the navigation control unit 3 determines an input text to be passed to the speech synthesizer 20 (step 201).

  Next, the input analysis unit 4 determines the intermediate symbol string of the fixed part and the intermediate symbol string of the variable part (steps 202 to 203), and the recorded voice data and the acoustic feature information of the recorded voice are obtained by the recorded voice selection unit 6. Is obtained (step 204). Subsequently, a creation range of the rule synthesized speech is determined (step 205). The processing so far is performed by the same method as in the first embodiment.

  Next, the rule synthesis parameter calculation unit 21 calculates a rule synthesis parameter and generates acoustic feature information of the rule synthesis voice (step 206). Here, in the first embodiment, the rule synthesis voice data is created in the rule synthesis unit 7, but in the second embodiment, the rule synthesis voice data is not created.

  Next, the connection boundary calculation unit 8 calculates the connection boundary position in the overlapping section between the recorded voice data and the rule synthesis parameter using the acoustic feature information of the recorded voice and the acoustic feature information of the rule synthesized voice (step) 207). This step is performed in the same manner as in the first embodiment.

  Next, the rule-synthesized voice data generation unit 22 generates rule-synthesized voice data using the acoustic feature information of the recorded voice, the acoustic feature information of the rule-synthesized voice, and the connection boundary position obtained by the connection boundary calculation unit 8. (Step 208). This step refers to the acoustic feature information of the recorded voice at the connection boundary position, corrects the rule synthesis parameter obtained in step 206, and generates rule synthesis voice data.

  For example, when the rule synthesis parameter is modified so that the difference in acoustic characteristics is small with respect to the phoneme at the connection boundary position, a synthesized speech with less connection distortion is generated.

  In the first embodiment, the rule synthesis parameter is created using the range of the rule synthesis voice data defined as one sentence including a variable part and the acoustic feature information of the overlapping section with the recorded voice data. In the second embodiment, however, the rule synthesis parameters are generated again using the acoustic feature information of the recorded voice at the connection boundary position obtained by the connection boundary calculation unit 8, and the rule synthesis voice data is generated. It is. Thereby, a smoother connection in consideration of the connection boundary position is performed.

  Next, using the connection boundary position obtained from the connection boundary calculation unit 8, the connection synthesis unit 9 cuts the recorded voice data and the rule synthesized voice data, and combines the cut recorded voice data and the rule synthesized voice data. By connecting, synthesized speech data corresponding to the input text is output (step 209).

  As described above, in the second embodiment, unlike the first embodiment, the rule synthesis parameter is set in two stages. In the first stage, rule synthesis parameters are set in consideration of the smooth connection of the entire sentence. In the second stage, the rule synthesis parameters are corrected in consideration of the connection boundary position obtained by the connection boundary calculation unit 8. In this way, by modifying the rule synthesis parameter, a more natural connection between the recorded voice data and the rule synthesized voice data is enabled.

(Example 3)
Next, a third embodiment of the present invention will be described.

  FIG. 9 is a block diagram showing a configuration in which the present invention is applied to a railway broadcast system according to a third embodiment of the present invention. FIG. 10 is a flowchart showing the operation of the speech synthesizer 30 according to the second embodiment.

  The present embodiment is a device that creates a synthesized speech by connecting prerecorded sound pieces, and has a function of generating a sound piece including a variable part by implementing the present invention.

  As shown in FIG. 11, the input unit 31 includes a display unit 33 for selecting a sentence example, a display unit 34 having a sequence configuration of sound pieces according to the selected sentence example, and a sound piece including a variable part. While viewing the input screen 32 having the display means 35 that can recognize the fixed part and the variable part of the text, and looking at the input screen 32, the user selects a sentence example that the user wants to output, and the order composition of the sound pieces is selected. An input device 36 is provided for editing and inputting variable text with a keyboard or the like.

  The sound piece information storage unit 35 is configured as shown in FIG. 12 and classifies the sound data recorded in advance in the recorded sound storage unit 5 as shown in the example of FIG. It is configured so that it can be expressed by a combination of the piece classification codes 701. Also, the sound piece information storage unit 35 stores a sound piece code 702 uniquely determined for each recorded sound data as shown in FIG. At this time, the sound piece classification code 701 is determined from the sound piece code 702. For example, in FIG. 13, the highest digit of the sound piece code 702 is configured to match the highest digit of the sound piece classification code 701.

  The operation of the third embodiment will be described below.

  The input unit 31 determines the composition of the sound piece by selecting a sentence example (step 301). Here, in the order configuration of the sound pieces, when a sound piece code is specified, a fixed sound piece is used, and when a sound piece classification code is specified, the corresponding sound piece is It can be generated by a speech synthesis method. For example, when the sound piece information shown in the example of FIG. 13 is stored and the sound piece classification code “200” is set in the input unit, the input screen includes an area for inputting text of a variable part. As the display data 703, the standard part “I will be going” is displayed.

  Subsequently, the variable part text is input from the keyboard, and the variable part text is determined (step 302). For example, if “Harajuku” is entered as the text of the variable part, “I will go to Harajuku” combined with the fixed part is generated as a sound piece.

  The input analysis unit 4 acquires an intermediate symbol string 704 of a fixed part corresponding to the sound piece classification code 701 in order to create a sound piece including the variable part specified by the input unit 31. Further, the variable part text obtained from the input unit is converted into an intermediate symbol string by language processing, and the variable part intermediate symbol string is determined (step 303). By this step, when the text of the variable part is “Harajuku”, the intermediate symbol string “Harajuk” of the variable part is obtained.

  Next, the recorded voice selection unit 6 selects an appropriate recorded voice from a plurality of recorded voices having the same fixed part according to the input of the variable part. Here, the intermediate symbol string including the fixed part and the variable part is compared with the intermediate symbol string corresponding to the recorded voice, and the one that matches the longest intermediate symbol string is selected (step 304). In this way, the connection boundary position between the recorded voice and the regular synthesized voice can be determined not only in the fixed part but also in the variable part depending on the case, so that a higher quality synthesized voice can be obtained. Can be generated.

  Next, the rule synthesis unit 7 determines a range for generating the rule synthesis speech using the intermediate symbol string of the variable part and the intermediate symbol string of the fixed part obtained by the input analysis unit 4 (step 305). Here, if the range for creating a rule-synthesized speech is defined as a single sound piece that includes a variable part, in addition to the variable part “Harajuk”, a regular-synthesized part “Yukiga Miraimas” will be generated. To do.

  Next, the connection boundary calculation unit 8 uses the acoustic feature information of the recorded voice and the acoustic feature information of the rule-synthesized voice to calculate the connection boundary position in the overlapping section of the recorded voice data and the rule-synthesized voice data (step 306). ).

  This step 306 is the same as step 106 of the first embodiment, but the connection boundary position between the recorded voice and the rule-synthesized voice is not only determined in the fixed part, but in some cases in the variable part. It is also possible to decide. An example in which the connection boundary position is determined in the variable part is shown in FIG. When the recorded sound corresponding to the sound piece information as shown in FIG. 13 is stored in the recorded sound storage unit 5 and the sound piece classification code “200” is designated as the standard part, the sound piece code “201”, Recorded voices “202” and “203” are selected. Here, when the intermediate symbol string of the variable part is “Harajuk”, comparing the intermediate symbol string “Harajuk Yukiga Mairimus” combined with the fixed part and the intermediate symbol string of each recorded voice, the speech piece code “201” “Shinjuku Yukiga My Limas” is selected.

  In this way, the overlapping section 801 of the recorded voice and the rule-synthesized voice is a section corresponding to “Jukuno Ashitano Tenchides”, which is not only the pre-designated fixed part 803 but also a part of the variable part 802 “ The recorded voice can be used also for the “juku” portion, and the connection boundary position 804 can be determined in the variable portion 802.

  Next, the connection synthesis unit 9 cuts the recorded voice data and the rule synthesized voice data using the connection boundary position obtained from the connection boundary calculation unit 8, and connects the cut recorded voice data and the rule synthesized voice data. Thus, synthesized speech data corresponding to the sound piece including the variable portion is created (step 307). Here, the connection boundary position is calculated as the time in the recorded voice data and the time in the rule synthesized voice data, and the voice data is disconnected and connected using the calculated time. This step is the same as step 107 in the first embodiment, but the process of outputting audio data from the speaker is performed by the next sound piece connecting unit 36.

  The sound piece connecting unit 36 connects the sound pieces based on the order of the sound pieces obtained from the input unit, and generates output sound (step 308). Here, the synthesized speech obtained from the connection synthesis unit is used as the sound piece including the variable part.

  In this way, in a device that creates a synthesized speech by connecting recorded sound pieces, a synthesized speech with a natural connection can be output using the sound pieces using the regular synthesized speech.

  As described above, in the third embodiment, when the present invention is applied to a railway broadcasting system, a function for generating a sound piece including a variable part in an apparatus for creating a synthesized voice by connecting sound pieces recorded in advance. It is possible to output high-quality sound.

  As described above in detail, according to the present invention, based on the acoustic feature information of the section where the recorded voice data stored in advance and the voice data generated by the rule synthesis overlap, the recorded voice and the rule synthesis are synthesized. It is possible to select a connection boundary in consideration of sound quality and prosody continuity with speech, and generate a natural synthesized speech. In addition, the rule synthesis creation means creates rule synthesis speech with the acoustic feature information of overlapping sections as a target, so that the sound quality and prosody of the rule synthesis speech approaches that of the recorded speech, and natural synthesis speech can be generated It becomes.

  The present invention is preferably applied to an in-vehicle car navigation system and a railroad broadcasting system, but can be applied to a general voice guidance system that outputs text as a voice.

  DESCRIPTION OF SYMBOLS 1 ... Speech synthesizer, 2 ... Navigation control device, 3 ... Navigation control part, 4 ... Input analysis part, 5 ... Recording voice storage part, 6 ... Recording voice selection part, 7 ... Rule synthesis part, 8 ... Connection boundary calculation part , 9 ... connection synthesis unit, 10 ... information reception unit, 11 ... GPS, 12 ... navigation data storage unit, 13 ... prosody model, 20 ... speech synthesis device, 21 ... rule synthesis parameter calculation unit, 22 ... rule synthesis voice data Generating unit, 30 ... speech synthesizer, 31 ... input unit, 35 ... sound piece information storage unit, 36 ... sound piece connecting unit.

Claims (17)

In a speech synthesizer that synthesizes text consisting of a fixed part and a variable part,
Recorded sound storage means for storing in advance first sound data which is created based on the recorded sound and which is the sound data including the fixed portion;
Rule synthesis means for generating, from the received text, second speech data including the variable part and at least a part of the fixed part;
The acoustic feature information corresponding to the text is at least one of a phoneme category, a fundamental frequency, a phoneme duration, a power, and a spectrum in a section where the first voice data and the second voice data overlap. Based on the connection boundary calculation means for selecting the position of the connection boundary between the recorded voice data and the voice data generated by rule synthesis,
Audio data of the text by connecting third audio data obtained by dividing the first audio data at the connection boundary and fourth audio data obtained by dividing the second audio data at the connection boundary. Connection synthesis means for synthesizing
The speech synthesizer characterized in that the connection boundary calculation means selects a position of the connection boundary from a plurality of phoneme boundaries included in a section where the first speech data and the second speech data overlap. The rule synthesis means uses the acoustic feature information of the first voice data in a section corresponding to the text where the first voice data and the second voice data overlap. The speech synthesizer according to claim 1, wherein the second speech data that matches speech data is generated. The rule synthesizing unit processes the second audio data based on acoustic feature information of the first audio data and the second audio data at a connection boundary position obtained from the connection boundary calculating unit. The speech synthesizer according to claim 1. The rule synthesizing unit generates the second audio data in a unit of one of the entire fixed part, one expiratory paragraph, and one sentence among the variable part and the fixed part preceding or following the variable part. The speech synthesizer according to claim 1. The recorded voice storage means stores, as the first voice data, voice data recorded in advance in one breath paragraph or one sentence unit, including a fixed part and at least a part other than the fixed part. The speech synthesizer according to claim 1 or 2. 2. The connection boundary position is calculated as a time in the first audio data and a time in the second audio data, and the audio data is disconnected and connected using the calculated time. Or the speech synthesizer of 2. 3. The speech synthesis apparatus according to claim 1, further comprising means for outputting the voice data synthesized by the connection synthesis means. In a speech synthesizer that synthesizes text consisting of a fixed part and a variable part,
A recorded voice storage unit that stores the recorded voice data including the fixed portion recorded in advance;
A rule synthesizing unit that generates rule-synthesized speech data including the variable part and at least a part of the fixed part from the received text;
Based on acoustic feature information that is at least one of phoneme category, fundamental frequency, phoneme duration, power, spectrum of the recorded voice data and the rule-synthesized voice data corresponding to the text, the recorded voice data and A connection boundary calculation unit for calculating a connection boundary position in a section where the rule-synthesized speech data overlaps;
A connection synthesis unit that connects the recorded voice data and the rule synthesized voice data cut out by dividing at the connection boundary position, and generates synthesized voice data corresponding to the text;
The speech synthesis apparatus, wherein the connection boundary calculation unit selects the connection boundary position from a plurality of phoneme boundaries included in a section where the recorded speech data and the rule-synthesized speech data overlap. The rule synthesizing means uses the acoustic feature information of the recorded voice data in a section corresponding to the text where the recorded voice data and the rule synthesized voice data overlap, and matches the rule with the recorded voice data 9. The speech synthesizer according to claim 8, wherein the synthesized speech data is generated. The rule synthesizing means processes the rule synthesized voice data based on acoustic feature information of the recorded voice data and the rule synthesized voice data at a connection boundary position obtained from the connection boundary calculating means. The speech synthesizer according to claim 8. The rule synthesizing unit generates the second audio data in units of one of the entire fixed part, one expiratory paragraph, and one sentence among the variable part and the fixed part preceding or following the variable part. The speech synthesizer according to claim 8. The recorded voice storage means stores the voice data recorded in advance in one breath paragraph or one sentence unit including the fixed part and at least a part other than the fixed part as the recorded voice data. Item 10. The speech synthesizer according to Item 8 or 9. 10. The connection boundary position is calculated as a time in the recorded voice data and a time in the rule synthesized voice data, and the voice data is disconnected and connected using the calculated time. The speech synthesizer described. 10. The speech synthesizer according to claim 8, further comprising means for outputting the synthesized speech data generated by the connection synthesis means. In a speech synthesizer that synthesizes text consisting of a fixed part and a variable part,
A recorded voice storage unit that stores the recorded voice data including the fixed portion recorded in advance;
A rule synthesis parameter calculation unit that calculates a rule synthesis parameter including the variable part and at least a part of the standard part from the received text, and generates acoustic feature information of the rule synthesis speech;
A connection boundary calculation unit that calculates a connection boundary position in a section where the recorded voice data and the rule synthesis parameter overlap using the acoustic feature information of the recorded voice and the acoustic feature information of the rule synthesis;
Using the acoustic feature information of the recorded speech, the acoustic feature information of the rule-synthesized speech, and the connection boundary position, a rule-synthesized speech data unit that generates rule-synthesized speech data;
A connection synthesizer that connects the recorded voice data and the rule-synthesized voice data cut out at the connection boundary position and outputs synthesized voice data corresponding to the text;
Means for outputting the synthesized voice data,
The connection boundary calculation unit selects the connection boundary position from a plurality of phoneme boundaries included in a section where the recorded voice data and the rule synthesis parameter overlap,
The speech synthesizer characterized in that the acoustic feature information is at least one of a phoneme category, a fundamental frequency, a phoneme duration, power, and a spectrum. In an apparatus for creating a synthesized speech by connecting a pre-recorded sound piece including a variable part and a sound piece including a fixed part,
A recorded voice storage unit for storing voice data composed of the previously recorded pieces;
An input analysis unit for creating an intermediate symbol string of the variable piece and an intermediate symbol string of the fixed part from the received input text;
According to the input of the variable part, a recording voice selection unit for selecting appropriate recording voice data from a plurality of recording voice data having the same fixed part,
A rule synthesizing unit that determines a range for generating rule-synthesized speech data, using an intermediate symbol string of the variable piece sound pieces obtained by the input analysis unit, and an intermediate symbol string of the fixed part sound pieces;
A connection boundary calculation unit that calculates a connection boundary position in an overlapping section between the recorded voice data and the rule-synthesized voice data, using the acoustic feature information of the recorded voice data and the acoustic feature information of the rule-synthesized voice data;
By using the connection boundary position obtained from the connection boundary calculation unit, the recorded voice data and the rule synthesized voice data are disconnected, and the disconnected recorded voice data and the rule synthesized voice data are connected. A connection synthesizer for creating synthesized voice data corresponding to the sound piece including the variable part;
Based on the order of the sound pieces obtained from the input text, and having a sound piece connection unit for connecting the sound pieces to generate output speech,
The connection boundary calculation unit selects the connection boundary position from a plurality of phoneme boundaries included in an overlapping section of the recorded voice data and the rule synthesized voice data,
The speech synthesizer characterized in that the acoustic feature information is at least one of a phoneme category, a fundamental frequency, a phoneme duration, power, and a spectrum. In a speech synthesis method that synthesizes text consisting of a fixed part and a variable part,
Recorded recording voice data including the standard part is stored in advance,
Generating rule-synthesized speech data including the variable part and at least a part of the fixed part from the received text,
Based on acoustic feature information that is at least one of phoneme category, fundamental frequency, phoneme duration, power, spectrum of the recorded voice data and the rule-synthesized voice data corresponding to the text, the recorded voice data and Selecting a connection boundary position from a plurality of phoneme boundaries included in a section where the rule-synthesized speech data overlaps;
A speech synthesizing method comprising: connecting the recorded speech data segmented and cut out at the connection boundary position and the rule synthesized speech data to generate synthesized speech data corresponding to the text.

Images