JP5218971B2 - Voice message creation apparatus and method - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To automatically create a voice message with uniform vocal quality. <P>SOLUTION: A real voice data corresponding to an input text is specified, and difference between a sound feature quantity of the specified real voice data and a reference sound feature quantity defined for the specified real voice data is calculated. On the basis of the calculated difference and a value indicating importance of the specified real voice data, it is determined whether or not, the sound feature quantity of the specified real voice data is converted. When it is determined that the sound feature quantity of the specified real voice data is converted, the sound feature quantity of the specified real voice data is converted, and the voice message based on the real voice data in which the sound feature quantity is converted is created. When it is determined that the sound feature quantity of the specified real voice data is not converted, the voice message based on the specified real voice data is created. <P>COPYRIGHT: (C)2010,JPO&amp;INPIT

Description

  The technology disclosed in this specification relates to an apparatus for creating a voice message, and more particularly to an apparatus for creating a voice message by connecting a plurality of sound pieces.

  Systems that create a voice message by connecting a plurality of sound pieces, such as on-site broadcast voice at a railway station and route guidance voice of a car navigation system, are widely used. The sound piece here is configured with a unit of a word, a syllable, or a phrase in which a plurality of them are connected. There are a case where the recorded real voice is used as a sound piece as it is, and a case where a synthesized voice generated by the speech synthesis technique is used as a sound piece.

  In such a system, the following two factors exist as quality degradation factors of the created voice message.

  The first quality deterioration factor is that a difference in voice quality occurs between sound pieces when the recorded real voice is directly used as a sound piece. In general, since voice recording work for creating a piece often involves a long period of time, it is extremely difficult to keep the voice quality of a speaker constant over the entire recording period. Therefore, there is a variation in voice quality in the created sound piece database, which becomes a quality deterioration factor when connected as a voice message.

  A second quality deterioration factor is that a difference occurs between the voice quality of the recorded real voice sound piece and the voice quality of the synthesized voice sound piece. In recent years, the progress of speech synthesis technology has been remarkable, and the quality of synthesized speech has been greatly improved compared to the past. However, when the synthesized voice is compared with the real voice, there is still a voice quality difference, which becomes a quality deterioration factor when connected as a voice message.

  As a countermeasure against such a quality deterioration problem due to a difference in sound quality between connected sound pieces, for example, a technique as shown in Patent Document 1 exists. According to Patent Document 1, the sound quality difference between the real voice sound piece and the synthetic sound piece is reduced by lowering the bit rate of the real voice sound piece.

  Patent Document 2 discloses a technique for creating a synthesized speech having an arbitrary prosody.

Non-Patent Document 1 and Non-Patent Document 2 disclose techniques for converting the voice quality of sound pieces.
JP-A-62-215299 JP-A-11-249677 Kazuo Nakata, “Speech” (first revised edition), Corona, 1995, pp. 126-128 Tamura et al., “Voice quality conversion for multiple speech unit selection / fusion speech synthesis”, Proceedings of the Acoustical Society of Japan, March 2006, pp. 237-238

  According to the technique disclosed in Patent Document 1, the difference in sound quality between the two is reduced by lowering the sound quality of the real voice sound piece on the premise that the sound quality of the synthesized sound piece is lower than the sound quality of the real voice sound piece. For this reason, the problem of quality deterioration resulting from the non-uniform sound quality of the entire voice message is improved. However, there is a problem that the sound quality of the entire created voice message is greatly deteriorated due to a decrease in the bit rate. In addition, although the sound quality can be reduced by lowering the bit rate, the voice quality cannot be adjusted. For this reason, according to the technique disclosed in Patent Document 1, it is not possible to eliminate quality deterioration due to a voice quality difference between real voice sound pieces. Furthermore, it is impossible to eliminate the voice quality difference between the synthesized speech piece and the real voice piece.

A representative invention disclosed in the present application is a voice message creation device for creating a voice message using real voice data, a storage device in which the real voice data is stored in advance, and a processor connected to the storage device If, and an input device and an output device coupled to the processor, the input information specifying the text to identify the human voice sound data corresponding to the specified text by the input information, the the synthesized speech data corresponding to the specified text, the prosody of the synthesized speech data is created to equal the prosody of human voice audio data the specific, the acoustic features of human voice audio data the specified, the It calculates a difference between the reference acoustic features an acoustic feature quantity of synthesized speech data generated, and the calculated difference, which is the specific Determining whether to convert the acoustic feature quantity of the identified real voice data based on a value indicating the importance of the voice voice data, and converting the acoustic feature quantity of the identified real voice data If determined, the acoustic feature amount of the identified real voice data is converted so that a difference from the reference acoustic feature amount is small , and a voice message based on the converted real voice data is converted into a voice message. If it is determined that the acoustic feature amount of the identified real voice data is not to be converted, a voice message based on the identified real voice data is created.

  According to an embodiment of the present invention, the voice quality conversion cost is calculated for each real voice sound piece, and the voice quality conversion is performed only on the sound piece whose voice quality conversion cost is equal to or less than the threshold value. The voice quality in the voice message can be made uniform while minimizing the quality degradation of the voice message. As a result, a high-quality voice message can be created.

  Hereinafter, an embodiment of the present invention will be described with reference to the drawings.

  FIG. 1 is a block diagram showing a schematic configuration of a voice message creating apparatus according to an embodiment of the present invention.

  The voice message creation device of this embodiment includes an input unit 1, an output unit 2, a storage unit 3, a voice message creation unit 4, a voice synthesis unit 5, a voice quality difference amount calculation unit 6, a voice quality conversion sound piece determination unit 7, and a voice quality conversion. Part 8 is provided.

  The input unit 1 accepts input of text or sound piece string information for creating a voice message.

  In addition to outputting the created voice message, the output unit 2 also outputs operation progress and operation results.

  The storage unit 3 stores programs and data necessary for carrying out the present invention.

  The voice message creation unit 4 determines a voice piece necessary for creating a voice message based on the text or the voice piece string information input to the input unit 1 and connects the voice piece whose voice quality has been changed to connect the voice piece. Create a message.

  The voice synthesizing unit 5 synthesizes sound pieces necessary for creating a voice message based on the input text.

  The voice quality difference amount calculation unit 6 calculates, for each real voice sound piece, a voice quality difference amount between the real voice sound piece and a synthesized sound piece created so that the prosody is equal to the real voice sound piece.

  The voice quality conversion sound piece determination unit 7 calculates a voice quality conversion cost based on the voice quality difference amount calculated by the voice quality difference amount calculation unit 6 and the sound piece information for each real voice sound piece, and based on the voice quality conversion cost. Then, the target voice conversion piece is determined.

  The voice quality conversion unit 8 converts the voice quality of the real voice sound piece determined by the voice quality conversion sound piece determination unit 7.

  Here, the concepts of “prosody” and “voice quality” will be described. A person who listens to the voice can recognize various characteristics of the audibility of the voice. Among the features of the voice, the pitch, the loudness, and the speaking speed are described as “prosody” in the present embodiment. On the other hand, among the features of the voice, those not included in the prosody are described as “voice quality” in the present embodiment. The voice quality is recognized from the viewpoint of hearing, for example, the thickness of the voice, the degree of voice blur, and the like. Of the voice features, those digitized by measurement are called acoustic features. Typical parameters treated as acoustic features are, for example, spectrum, cepstrum, Δ cepstrum, mel cepstrum, and the like.

  FIG. 2 is a block diagram showing a hardware configuration of the voice message creation device according to the embodiment of the present invention.

  The voice message creation device shown in FIG. 1 is realized by the hardware shown in FIG.

  The voice message creation device according to the present embodiment includes a control device 110, a storage device 120, an input device 130, and an output device 140 that are communicably connected to each other.

  The control device 110 controls the operation of this embodiment. The control device 110 includes a CPU 111 and a memory 112. The CPU 111 is a processor that executes a program stored in the memory 112. The memory 112 is a semiconductor memory, for example, and stores a program executed by the CPU 111 and data referred to by the CPU 111. The program and data stored in the memory 112 may be stored in the storage device 120 and copied from the storage device 120 to the memory 112 as necessary. The CPU 111 controls the input / output of data and various other processes in the storage device 120, the input device 130, and the output device 140 by executing a program stored in the memory 112.

  The storage device 120 corresponds to the storage unit 3 in FIG. The storage device 120 stores a program executed by the CPU 111 and data referred to by the CPU 111. The storage device 120 may be, for example, a disk device such as a hard disk drive (HDD) or a semiconductor memory such as a flash memory. The storage device 120 of the present embodiment stores a voice message creation unit 121, a voice synthesis unit 122, a voice quality difference amount calculation unit 123, a voice quality conversion sound piece determination unit 124, and a voice quality conversion unit 125. When the CPU 111 executes these, the voice message creation unit 4, the voice synthesis unit 5, the voice quality difference amount calculation unit 6, the voice quality converted sound piece determination unit 7 and the voice quality conversion unit 8 shown in FIG. 1 are realized.

  The storage device 120 further stores an audio database 126. The speech database 126 includes sound piece data corresponding to various texts (that is, data obtained by A / D converting real voice sound pieces obtained when a speaker speaks various texts). .

  The input device 130 corresponds to the input unit 1 in FIG. The input device 130 includes a keyboard 133 and a mouse 134. The keyboard 133 and the mouse 134 are interfaces that receive an instruction from the operator and transmit the instruction to the control device 110. The input device 130 may include any type of interface instead of (or in addition to) the keyboard 133 and mouse 134. The operator can input text or speech string information to the voice message creation device by operating the input device.

  The output device 140 corresponds to the output unit 2 in FIG. The output device 140 includes a digital / analog (D / A) converter 141, a speaker 142, and a display 143. The D / A converter 141 converts audio data into an analog electric signal. The speaker 142 converts the analog electrical signal output from the D / A converter 141 into sound. The display 143 is an interface for displaying various information to the operator.

  Note that the voice message creation unit 4, the voice synthesis unit 5, the voice quality difference amount calculation unit 6, the voice quality conversion sound piece determination unit 7 and the voice quality conversion unit 8 according to the present embodiment are stored in the memory 112 as shown in FIG. The program is realized by the CPU 111 executing the program. However, these may be realized by dedicated hardware (for example, a dedicated processor) provided in the voice message creating apparatus.

  FIG. 3 is a flowchart showing the overall operation of the voice message creation apparatus according to the embodiment of the present invention.

  In the voice message creation device of FIG. 1, first, the voice message creation unit 4 selects a speech piece in a database used for creating a voice message based on the text or the speech piece string information input to the input unit 1. Determine (step S201). An example of the result determined as the sound piece to be used is shown in FIG. Text 302 is the input text. The sound piece ID 301 is an ID assigned to a real voice sound piece corresponding to the input text. This ID is assigned to identify the data of the real voice fragment in the voice database 126. When the data of the real voice unit corresponding to the input text 302 does not exist in the speech database 126, the speech unit ID 301 corresponding to the text 302 is blank. The sound piece corresponding to the text 302 needs to be generated by synthesized speech.

  FIG. 4A shows, as an example, the result determined in step S <b> 201 when the text “soon to the right after Nakano Sakaue intersection” is input for a while. In this example, the data of the real voice component corresponding to the text “Soon”, “Right”, “After that”, and “For a while” is stored in the speech database 126. On the other hand, the data of the real voice segment corresponding to the text “Nakano Sakagami Intersection” is not stored in the speech database 126. For this reason, the sound piece ID 301 corresponding to the text “Nakano Sakaue Intersection” is blank.

  Next, the speech synthesizer 5 generates a synthesized speech corresponding to a text in which no real voice segment exists (that is, a synthesized speech that reads out the text) (step S202). The generation of the synthesized speech may be performed by any conventionally known method. An example of the use sound piece determination result after the synthesized speech is generated is shown in FIG. In the example of FIG. 4B, the sound piece ID 301 whose head is an alphabet identifies a sound piece generated by synthesized speech.

  Specifically, in FIG. 4B, a sound piece is generated by a synthesized speech corresponding to the text “Nakano Sakagami intersection” for which no data of the real voice sound piece exists, and the sound piece ID “A001” is assigned to the sound piece. Is given. Other parts are the same as those in FIG.

  Next, the voice quality difference amount calculation unit 6 performs a voice quality difference amount calculation process (A) for each used real voice sound piece determined by the voice message creation unit 4 (step S204).

  FIG. 5 is a flowchart showing voice quality difference amount calculation processing (A) executed in the embodiment of the present invention.

  First, the voice quality difference amount calculation unit 6 creates a real voice prosody synthesized speech having the same prosody as the real voice sound piece existing in the speech database 126 among the sound pieces determined in step S201 (step S401). As a technique for creating a real voice prosody synthesis voice, for example, a technique described in Patent Document 2 cited as a background technique may be used.

  Next, the voice quality difference amount calculation unit 6 calculates the difference in acoustic feature amount between the created real voice prosody synthesized speech and the real voice sound piece as a voice quality difference amount (step S402). Since the prosody of the real voice prosody synthesis speech and the real voice voice fragment are the same, the difference in acoustic feature values can be treated as a difference in voice quality. A spectrum, a cepstrum, a Δ cepstrum, a mel cepstrum, or the like may be used as the acoustic feature amount.

  When the voice quality difference amount calculation processing for the real voice sound piece is completed, the voice quality difference amount calculation unit 6 determines whether or not the voice quality difference sentence amount calculation processing is completed for all the real voice sound pieces determined by the voice message creation unit 4 ( Step S403). When there is a real voice sound piece that has not yet been subjected to the voice quality difference amount calculation process, the voice quality difference amount calculation unit 6 repeats the processes of steps S401 and S402 until the voice quality difference amount calculation process is completed for all the real voice sound pieces. .

  For example, if in step S201 real voice sound pieces corresponding to the texts “Soon”, “Right”, “After that”, and “It's been a while” are determined, Steps S401 and S402 are executed for each of them. The

  For example, the voice quality difference amount calculation unit 6 creates a synthetic speech piece corresponding to the text “Soon” so that its prosody is equal to the prosody of the real voice sound piece “Soon” (step S 401). Then, the voice quality difference amount calculation unit 6 calculates a difference amount between the acoustic feature amount of the synthesized speech piece “Soon” and the acoustic feature amount of the real voice sound piece “Soon” created in Step S401 (Step S402). Steps S401 and S402 are repeated until the same processing is completed for the texts “rightward”, “after that”, and “after a while” (step S403).

  FIG. 6 is an explanatory diagram showing a voice quality difference amount calculation processing result in the embodiment of the present invention.

  FIG. 6 shows an example of the voice quality difference amount calculated for each real voice sound piece determined as shown in FIG. In FIG. 6, the sound piece ID 501 and the text 502 correspond to the sound piece ID 301 and the text 302 in FIG. 4, respectively.

  The voice quality difference amount 503 is a voice quality difference amount calculated for each voice voice piece. In the example of FIG. 4, “1.2345”, “0.5467”, “3.210” are respectively used as the voice quality difference amounts 503 corresponding to the texts “soon”, “right”, “after that”, and “for a while”. And “0.3322” are calculated. Since the voice voice sound piece corresponding to the text “Nakano Sakagami Intersection” does not exist in the voice database 126, the voice quality difference amount 503 corresponding to this text is not calculated.

  According to a conventional synthesized speech creation technique (for example, one described in Patent Document 2), the voice quality of the created synthesized speech is substantially constant. That is, the value of the voice quality difference amount 503 corresponding to each real voice sound piece calculated in step S204 is obtained by combining each real voice sound piece and a synthesized sound piece (in the example of FIG. This is considered to be equivalent to the difference in voice quality between the Nakano-Sakaue intersection and the synthesized speech piece corresponding to “ Therefore, by converting the voice quality of each real voice sound piece so that the value of the voice quality difference amount 503 corresponding to each real voice sound piece becomes smaller (preferably 0), It can be expected that the difference in voice quality between the two will be eliminated, and further, the difference in voice quality between the real voice segments will also be eliminated.

  However, generally, by converting the voice quality of a real voice sound piece, the sound quality (that is, ease of hearing of the sound piece) is lowered. The greater the voice quality conversion amount (that is, the greater the value of the voice quality difference amount 503 to be eliminated), the greater the sound quality reduction amount. The lower the sound quality of a sound piece, the more likely it is that the sound piece cannot be heard. Therefore, whether or not to actually convert the voice quality of the real voice sound piece needs to be determined for each real voice sound piece based on various factors.

  Therefore, next, the voice quality conversion sound piece determination unit 7 performs voice quality conversion sound piece determination processing (B) in step S205 of the overall process flowchart (FIG. 3).

  FIG. 7 is a flowchart showing voice quality conversion sound piece determination processing (B) executed in the embodiment of the present invention.

  First, the voice quality conversion sound piece determination unit 7 calculates sound piece information of each real voice sound piece (step S601). An example of the sound piece information calculation result is shown in FIG.

  The table in FIG. 8 includes columns of a speech piece ID 701, text 702, voice quality difference amount 703, sound piece information 704, voice quality conversion cost 705, and voice quality conversion availability 706.

  The sound piece ID 701, text 702, and voice quality difference amount 703 are the same as the sound piece ID 501, text 502, and voice quality difference amount 503, respectively.

  The sound piece information 704 is determined based on the importance of each real voice sound piece. The importance level of a real voice sound piece is an index indicating the magnitude of a disadvantage that occurs when a real voice sound piece cannot be heard, and may be referred to as necessity or usefulness. In the example of FIG. 8, a predetermined arbitrarily set importance and the length of a sound piece are used as an index indicating the importance of each real voice sound piece.

  The arbitrarily set importance level is an index that is arbitrarily set in advance for each real voice segment, which represents the importance level of the real voice segment in the created voice message. For example, in the case of the voice message of the car navigation system as shown in FIG. 8, the importance of the real voice sound piece indicating the destination, the distance, the direction, etc. is relatively high, and the importance of the other real voice sound pieces is relatively low. Note that the arbitrarily set importance may be set for each sound piece in advance by the manufacturer or user of the voice message creation device, or is set to a predetermined standard when the voice message creation unit 4 determines the use voice piece. Therefore, it may be calculated.

  In the example of FIG. 8, the reciprocal of the arbitrarily set importance (column 704 </ b> A) is calculated as part of the sound piece information 704. Specifically, FIG. 8 shows an example in which it is determined that the sound piece corresponding to the text “rightward” is the most important (that is, the disadvantage is the greatest when it cannot be heard). Therefore, a value “0.05” smaller than that corresponding to the other text is stored in the column 704A corresponding to the text “rightward”. On the other hand, in FIG. 8, it is determined that the importance level of the sound pieces corresponding to the text “soon” and “after” is relatively low. For this reason, relatively large values (“3.00” and “2.50”, respectively) are stored in the column 704A corresponding to the texts.

  On the other hand, it is considered that the longer the sound piece is, the higher the possibility that an important message is included therein (that is, the importance is high). In the example of FIG. 8, the number of syllables of the text corresponding to the sound piece is defined as the length of the sound piece, and the reciprocal number (column 704 </ b> B) is calculated as a part of the sound piece information 704. For example, since the number of syllables of the text “Soon” is “4”, “0.25” which is the reciprocal of “4” is stored in the column 704B corresponding to the text.

  Note that the piece information 704 may be further calculated based on the part of speech of the word included in the text corresponding to the real voice piece. For example, the arbitrarily set importance may be determined based on the part of speech. In the case of the voice message of the car navigation system as shown in FIG. 8, the proper noun may be the place name of the destination or the point on the route to reach the destination. Considered high importance. For this reason, a value larger than the arbitrarily set importance corresponding to other parts of speech may be set as the arbitrarily set importance corresponding to the proper noun.

  When the text corresponding to one real voice sound piece includes a plurality of words and the parts of speech of these words are different from each other, for example, the reciprocal of the total value of the arbitrarily set importance of the parts of speech of the plurality of words is the sound piece information 704. Or the reciprocal of the largest value among the arbitrarily set importance levels of the parts of speech of the plurality of words may be used as the piece information 704.

  Next, the voice quality conversion sound piece determination unit 7 calculates a voice quality conversion cost 705 based on the calculated sound piece information 704 and the voice quality difference amount calculated by the voice quality difference amount calculation processing of FIG. S602). Specifically, the voice quality conversion sound piece determination unit 7 calculates the voice quality conversion cost 705 by weighting the voice quality difference amount based on the sound piece information 704.

  In the example of FIG. 8, the sum of the calculated sound piece information 704 and the calculated voice quality difference amount is calculated as the voice quality conversion cost 705. For example, as shown in FIG. 8, when the voice quality difference amount 703 corresponding to the text “Soon” is “1.234” and the corresponding sound piece information 704 is “3.00” and “0.25”, A total value “4.484” of these values is calculated as the voice quality conversion cost 705. However, the voice quality conversion cost 705 may be calculated by a method other than the above (for example, by multiplying the voice quality difference amount by the value of the speech piece information 704).

  Next, the voice quality conversion sound piece determination unit 7 determines whether or not the value of the voice quality conversion cost 705 exceeds a predetermined threshold value for each real voice sound piece (step S603).

  In the example of FIG. 7, since the threshold value is 1.0, the voice quality conversion cost 705 for the sound piece whose sound piece ID 701 is “0001” and the sound piece whose sound piece ID 701 is “0015” in FIG. 4.484 "and" 5.960 ") exceed the threshold. Information (“1” in the example of FIG. 8) indicating that the sound piece is a voice quality conversion target is stored as voice quality conversion availability 706 corresponding to the sound piece exceeding the threshold (step S604).

  On the other hand, information (“0” in the example of FIG. 8) indicating that the sound piece is not subject to voice quality conversion is stored as voice quality conversion availability 706 corresponding to the sound piece not exceeding the threshold (step S605).

  In FIG. 7, a sound piece determined to be a voice quality conversion target is described as a voice quality conversion sound piece, and a sound piece determined to be not a voice quality conversion target is described as a voice quality non-converted sound piece.

  According to the above processing, the voice quality conversion cost 705 increases as the value of the voice quality difference amount calculated for the real voice sound piece increases and the importance of the real voice sound piece decreases. A voiced voice piece having a higher voice quality conversion cost 705 is more likely to be subject to voice quality conversion.

  The greater the value of the voice quality difference calculated for the real voice sound piece, the farther the voice quality of the real voice sound piece is from the voice quality of the uniform voice message to be created. That is, in order to create a voice message with a uniform voice quality, it can be said that the voice quality of a voice voice piece having a larger voice quality difference value is more likely to be converted. For this reason, according to said process, when importance is the same, the more the voice quality sound piece with a larger value of voice quality difference amount, the easier it is to be subject to voice quality conversion.

  However, since the sound quality of the sound piece deteriorates due to the voice quality conversion, it is better not to execute the voice quality conversion in order to ensure the ease of hearing of the sound piece with high importance. For this reason, according to said process, when the amount difference of voice quality is the same, the more important the voice voice piece, the less the voice quality conversion target.

  Next, the voice quality conversion unit 8 performs voice quality conversion processing of the real voice sound piece determined to be the voice quality converted sound piece in step S205 in step S206 of the overall processing flowchart (FIG. 3). This voice quality conversion process may be executed using the technique disclosed in Non-Patent Document 1 or Non-Patent Document 2 cited as the background art of the present invention. This voice quality conversion is executed with the target difference between the acoustic feature quantities calculated in step S402 of the voice quality difference amount calculation process of FIG. 5 between the real voice sound piece and the real voice prosody synthesized speech. In other words, the difference between the acoustic feature quantities of the real voice sound piece after the voice quality is converted by the voice quality conversion process and the real voice prosody synthesized speech corresponding to the real voice sound piece is the difference calculated in step S402 in FIG. It becomes smaller (preferably 0).

  Next, the voice message creation unit 4 creates a voice message by connecting each sound piece, and outputs the created voice message (step S207).

  As described above, according to the present invention, the voice quality conversion cost is calculated for each used voice voice sound piece, and the voice quality conversion is performed only for the voice quality conversion cost exceeding the threshold value, thereby reducing the sound quality deterioration more than necessary. It is possible to create a voice message with uniform voice quality while preventing it.

  As mentioned above, although embodiment of this invention was described, this invention is not limited to these embodiment, In the range which does not deviate from the summary, it can implement in various deformation | transformation.

It is a block diagram which shows schematic structure of the voice message preparation apparatus of embodiment of this invention. It is a block diagram which shows the hardware constitutions of the voice message preparation apparatus of embodiment of this invention. It is a flowchart which shows the operation | movement of the whole voice message preparation apparatus of embodiment of this invention. It is explanatory drawing which shows an example of the utilization sound piece determination result in embodiment of this invention. It is a flowchart which shows the voice quality difference amount calculation process performed in embodiment of this invention. It is explanatory drawing which shows the voice quality difference amount calculation process result in embodiment of this invention. It is a flowchart which shows the voice quality conversion sound piece determination process performed in embodiment of this invention. It is explanatory drawing which shows an example of the determination result of the voice quality conversion sound piece in embodiment of this invention.

Explanation of symbols

DESCRIPTION OF SYMBOLS 1 Input part 2 Output part 3 Storage part 4 Voice message preparation part 5 Speech synthesis part 6 Voice quality difference amount calculation part 7 Voice quality conversion sound piece determination part 8 Voice quality conversion part

Claims (16)

A voice message creation device that creates a voice message using real voice data,
A storage device in which the real voice data is stored in advance, a processor connected to the storage device, an input device and an output device connected to the processor,
When the information specifying the text is input, the voice data corresponding to the text specified by the input information is identified,
Creating synthesized speech data corresponding to the specified text so that the prosody of the synthesized speech data is equal to the prosody of the identified real voice data;
Calculating a difference between the acoustic feature amount of the identified real voice data and a reference acoustic feature amount which is an acoustic feature amount of the created synthesized speech data ;
Based on the calculated difference and a value indicating the importance level of the identified real voice data, it is determined whether or not to convert an acoustic feature amount of the identified real voice data,
When it is determined to convert the acoustic feature amount of the identified real voice data, the acoustic feature amount of the identified real voice data is converted so that a difference from the reference acoustic feature amount is small , Create a voice message based on the real voice data with converted features,
A voice message creation device that creates a voice message based on the identified real voice data when it is determined not to convert an acoustic feature amount of the identified real voice data. 2. The voice message creation device according to claim 1, wherein the prosody of the voice data includes one corresponding to a pitch, a loudness, and a speaking speed among acoustic feature quantities of the voice data. 2. The voice message creation device according to claim 1, wherein the acoustic feature amount of the voice data includes at least one of a spectrum, a cepstrum, a Δ cepstrum, and a mel cepstrum calculated from the voice data. The voice message creation device comprises:
  By calculating the conversion cost by weighting the calculated difference using a value indicating the importance level of the identified real voice data,
  2. The voice message creation device according to claim 1, wherein when the calculated conversion cost exceeds a predetermined threshold, it is determined to convert an acoustic feature amount of the identified real voice data. The voice message creation device is configured such that the conversion cost increases as the calculated difference increases, and the conversion cost increases as the value indicating the importance of the identified real voice data indicates a lower importance. The voice message creation device according to claim 4, wherein the conversion cost is calculated so as to increase. The value indicating the importance level of the real voice data is a value given in advance to the real voice data, the length of the voice corresponding to the real voice data, and the word included in the text corresponding to the real voice data. 6. The voice message creating apparatus according to claim 5, wherein the voice message creating apparatus is determined based on at least one of the parts of speech. 7. The voice message creation device according to claim 6, wherein the value indicating the importance level of the real voice data is determined so as to indicate a higher importance level as the length of the voice corresponding to the real voice data is longer. . The value indicating the importance of the real voice data when a proper noun is included in the text corresponding to the real voice data, the value of the real voice data when the proper noun is not included in the text corresponding to the real voice data The voice message creation device according to claim 6, wherein the voice message creation device is determined so as to indicate an importance level higher than a value indicating the importance level. A voice message creation device creates a voice message using real voice data,
  The voice message creation device includes a storage device in which the real voice data is stored in advance, a processor connected to the storage device, and an input device and an output device connected to the processor,
  The method
  When information specifying a text is input, a first procedure for specifying real voice data corresponding to the text specified by the input information;
  A second step of creating synthesized speech data corresponding to the designated text so that the prosody of the synthesized speech data is equal to the prosody of the identified real voice data;
  A third procedure for calculating a difference between the acoustic feature quantity of the identified real voice data and a reference acoustic feature quantity that is an acoustic feature quantity of the synthesized voice data created in the second procedure;
  A fourth procedure for determining whether or not to convert an acoustic feature amount of the identified real voice data, based on the calculated difference and a value indicating the importance of the identified real voice data; ,
  When it is determined to convert the acoustic feature amount of the identified real voice data, the acoustic feature amount of the identified real voice data is converted so that a difference from the reference acoustic feature amount is small, A fifth procedure for creating a voice message based on the real voice data into which the feature amount is converted;
  And a sixth step of creating a voice message based on the identified real voice data when it is determined not to convert the acoustic feature quantity of the identified real voice data. The method according to claim 9, wherein the prosody of the voice data includes one corresponding to a voice pitch, a loudness, and a speaking speed among acoustic feature quantities of the voice data. The method according to claim 9, wherein the acoustic feature amount of audio data includes at least one of a spectrum, a cepstrum, a Δ cepstrum, and a mel cepstrum calculated from the audio data. The fourth procedure includes
  A procedure for calculating a conversion cost by weighting the calculated difference using a value indicating the importance of the identified real voice data;
  The method according to claim 9, further comprising a step of determining to convert an acoustic feature amount of the identified real voice data when the calculated conversion cost exceeds a predetermined threshold. The procedure for calculating the conversion cost is such that the conversion cost increases as the calculated difference increases, and the value indicating the importance of the identified real voice data indicates a lower importance. The method according to claim 12, further comprising a step of calculating the conversion cost such that the conversion cost is increased. The value indicating the importance level of the real voice data is a value given in advance to the real voice data, the length of the voice corresponding to the real voice data, and the word included in the text corresponding to the real voice data. 14. The method of claim 13, wherein the method is defined based on at least one of the parts of speech. 15. The method according to claim 14, wherein the value indicating the importance level of the real voice data is determined such that the higher the voice length corresponding to the real voice data is, the higher the importance level is. The value indicating the importance of the real voice data when a proper noun is included in the text corresponding to the real voice data, the value of the real voice data when the proper noun is not included in the text corresponding to the real voice data 15. The method of claim 14, wherein the method is defined to indicate a higher importance than a value indicating the importance.

Images