JP4580317B2 - Speech synthesis apparatus and speech synthesis program - Google Patents

Description

  The present invention relates to a speech synthesizer and a speech synthesis program for performing speech synthesis using accented phonemes.

  2. Description of the Related Art Conventionally, a speech synthesizer that outputs speech synthesis data based on input text data is known (see, for example, Patent Document 1). The speech synthesizer disclosed in Patent Document 1 is provided with a speech database in which phonemes and utterance times of the phonemes are recorded. The speech synthesizer disassembles input text data into phonemes. The speech database is searched for phonemes as search units for the decomposed phonemes, and the search result that minimizes the sum of the concatenation cost and the phoneme prosody cost is output as speech synthesis data.

Conventionally, in speech recognition, a method of clustering as phonemes (triphones) considering phonemes (front and back phoneme environments) arranged before and after the phoneme is known (see Non-Patent Document 1, for example).
JP 10-49193 (paragraphs 0014 to 0018, FIG. 1) SJ Young, JJ Odell, and PC Woodland, “Tree-based state tying for high accuracy acoustics modeling”, Proc. Of ARPA Human Language Technology Workshop, 1994, vol.1, p.307-312

  However, since the clustering method disclosed in Non-Patent Document 1 is a clustering that considers only the phoneme environment before and after, that is, the characteristics of the spectral envelope, the influence of accents due to changes in the fundamental frequency is ignored. Therefore, in the speech synthesizer using this clustering method, the synthesized speech (speech synthesis data) to be synthesized feels unnatural (the connection of the phonemes before and after becomes unnatural), that is, synthesized speech (speech synthesis data) ) Is deteriorated in naturalness.

  The present invention has been made in view of the above problems, and a speech synthesizer and a speech synthesizer program capable of preventing deterioration of naturalness that a synthesized speech (speech synthesized data) sounds unnaturally. The purpose is to provide.

  In order to achieve the above object, a speech synthesizer according to claim 1 of the present invention includes a text data analysis means, a phoneme clustering means, a phoneme accent clustering means, a speech database, a speech data search means, and a speech. And data linking means.

  According to such a configuration, the speech synthesizer performs morphological analysis on the input text data by the text data analysis means, and converts the text data into accented phonemes. Here, the morpheme refers to the minimum character string that has no meaning if it is made finer than this, and the morpheme analysis is to analyze the sentence by breaking it down to the morpheme level. In addition, the accent includes at least one of a pitch, strength, length, and rhythm. Then, the speech synthesizer clusters the phonemes with accents converted by the text data analysis unit by the phoneme clustering unit with the phonemes arranged before and after the phonemes with the accents, and the phoneme clustering unit performs the phoneme clustering unit. The phonemes with accents clustered by the means are clustered with phoneme accents arranged before and after the phonemes with accents. Here, clustering refers to classifying a set of phonemes to be classified into a predetermined cluster (group) by paying attention to some attribute. In addition, the clustering method using the phoneme accents arranged before and after the accented phonemes uses at least one of the pitches of the front and back sounds, strength, shortness, and rhythm.

  Then, the speech synthesizer calculates a concatenation score of speech data strings generated by combining speech data corresponding to accented phonemes clustered by the phoneme accent clustering means by means of a Viterbi search. The voice data string that maximizes the connection score is searched from the voice database. This speech database stores phonemes arranged before and after the accented phonemes and speech data corresponding to accented phonemes clustered by the accents of the phonemes. Here, the Viterbi search is a method of leaving only the history of a hypothesis (a combination of speech data corresponding to accented phonemes) that gives the best (maximum) score. Then, the speech synthesizer concatenates the speech data strings searched by the speech data search means by the speech data connection means. By combining the speech data strings, speech synthesis data is generated and output from the speech synthesizer.

  Further, the speech synthesizer according to claim 2 is a text data analysis means, a phoneme clustering means, a phoneme accent clustering means, a phoneme string storage means, a speech database, a phoneme string dividing means, and a speech data search means. And voice data connecting means.

  According to such a configuration, the speech synthesizer performs morphological analysis on the input text data by the text data analysis means, and converts the text data into accented phonemes. Then, the speech synthesizer clusters the phonemes with accents converted by the text data analysis unit by the phoneme clustering unit with the phonemes arranged before and after the phonemes with the accents, and the phoneme clustering unit performs the phoneme clustering unit. The phonemes with accents clustered by the means are clustered with phoneme accents arranged before and after the phonemes with accents. In this speech synthesizer, a phoneme sequence storage unit stores in advance a phoneme arranged before and after the accented phoneme and a sequence of accented phonemes clustered with the accent of the phoneme, and is stored in the phoneme sequence storage unit. Voice data corresponding to the accented phoneme string is stored in the voice database. Then, the speech synthesizer divides the text data converted into the phonemes clustered by the phoneme accent clustering means by the phoneme string dividing means into accented phoneme strings stored in the phoneme string storage means. Therefore, the input text data is divided into a plurality of phoneme strings registered in advance.

  The speech synthesizer then performs a Viterbi search on a concatenation score of the speech data sequence generated by combining speech data corresponding to the accented phoneme sequence divided by the phoneme sequence division unit by the speech data search unit. The voice data string having the maximum connection score is searched from the voice database. Then, the speech synthesizer concatenates the speech data strings searched by the speech data search means by the speech data connection means. By combining the speech data strings, speech synthesis data is generated and output from the speech synthesizer.

  The speech synthesis program according to claim 3, in order to synthesize speech corresponding to the text data, comprises a computer, text data analysis means, phoneme clustering means, phoneme accent clustering means, speech data search means, speech data It is made to function as a connection means.

  According to such a configuration, the speech synthesis program performs morphological analysis on the input text data by the text data analysis means, and converts the text data into accented phonemes. The speech synthesis program clusters the phonemes with accents converted by the text data analysis unit with the phoneme clustering unit using the phonemes arranged before and after the phonemes with the accents, and the phoneme clustering unit with the phoneme clustering unit. The phonemes with accents clustered by the means are clustered with phoneme accents arranged before and after the phonemes with accents. Then, the speech synthesis program calculates a concatenation score of speech data strings generated by combining speech data corresponding to accented phonemes clustered by the phoneme accent clustering means by means of a Viterbi search. The voice data string that maximizes the connection score is searched. Then, the speech synthesis program connects the speech data strings searched by the speech data search means by the speech data connection means.

  According to a fourth aspect of the present invention, there is provided a speech synthesis program comprising: a computer for text data analysis means, phoneme clustering means, phoneme accent clustering means, phoneme string dividing means, speech data search, in order to synthesize speech corresponding to text data. And functioning as voice data connection means.

  According to such a configuration, the speech synthesis program performs morphological analysis on the input text data by the text data analysis means, and converts the text data into accented phonemes. The speech synthesis program clusters the phonemes with accents converted by the text data analysis unit with the phoneme clustering unit using the phonemes arranged before and after the phonemes with the accents, and the phoneme clustering unit with the phoneme clustering unit. The phonemes with accents clustered by the means are clustered with phoneme accents arranged before and after the phonemes with accents. Then, the speech synthesis program divides the text data converted into the phonemes clustered by the phoneme accent clustering means into phoneme strings with accents by the phoneme string dividing means. Then, the speech synthesis program uses a speech data search means to generate a Viterbi search for a concatenation score of speech data strings generated by combining speech data corresponding to the accented phoneme strings divided by the phoneme string splitting means. And the voice data string having the maximum connection score is searched. Then, the speech synthesis program connects the speech data strings searched by the speech data search means by the speech data connection means.

  According to the first or third aspect of the invention, since the accented phoneme environment is taken into consideration in addition to the preceding and following phoneme environment, it is possible to create a synthesized speech with a correct accent. Become. As a result, it is possible to prevent the deterioration of naturalness that the synthesized speech is heard unnaturally.

  According to the second or fourth aspect of the invention, the input text data is divided into pre-registered accented phoneme strings, and the phoneme strings are used as search units for searching the speech database. Therefore, it is possible to prevent searching for phonemes having different phoneme environments before and after searching, and to shorten the time required for speech synthesis processing. As a result, it is possible to prevent deterioration in sound quality of the synthesized speech synthesis data.

Hereinafter, embodiments of the present invention will be described with reference to the drawings.
(First embodiment)
[Configuration of speech synthesizer]
FIG. 1 is a functional block diagram showing the configuration of the speech synthesizer according to the first embodiment. The speech synthesizer 1 synthesizes speech to be output with a speech data string based on input text data, and includes a CPU (Central Processing Unit), a ROM (Read Only Memory), and a RAM (Random Access). Memory (HDD), hard disk drive (HDD), input / output interface and the like (not shown), and the CPU implements various functions to be described later by expanding a program stored in the HDD or the like into the RAM. is there. As shown in FIG. 1, the speech synthesizer 1 includes an input unit 2, a storage unit 3, and a speech synthesis control unit 4.

The input unit 2 inputs text data to the speech synthesis control unit 4 and is composed of, for example, a keyboard and a mouse. The input unit 2 may be configured by a communication interface that receives text data via a communication circuit network (not shown) such as the Internet.
The storage unit 3 includes pronunciation word storage unit 31, connection probability storage unit 32, speech database 33, and memory 34.

The pronunciation word storage means 31 stores accented phonemes for each word, and is a general recording medium such as an HDD.
The connection probability storage means 32 stores word connection probabilities and is a general recording medium such as an HDD.
The audio database 33 stores unit audio (audio data) and is a general recording medium such as an HDD. The unit speech stored in the speech database 33 is based on phonemes in which accents are added to all vowels. The phonemes are phonemes arranged before and after the accented phonemes and accented phonemes clustered by the accents of the phonemes, and are classified into a plurality of clusters. Here, the accent includes at least one of a pitch, strength, length, and rhythm. For example, a “sentence” made up of a set of these plural accented phonemes is a structural unit of the database. Note that “sentence number” may be assigned to each sentence, or the utterance time (utterance start time and utterance end time; utterance time) of each sentence may be recorded.
The memory 34 is a rewritable storage unit such as a semiconductor memory or a magnetic memory, and is used for processing by the speech synthesis control unit 4.

  The speech synthesis control means 4 synthesizes speech data for accented phonemes in consideration of the front and back phoneme environment and the front and back accent environment. As shown in FIG. 1, the text synthesis analysis means 41 and the phoneme Clustering means 42, phoneme accent clustering means 43, speech data searching means 44, speech data correcting means 45, and speech data connecting means 46 are provided.

  The text data analysis unit 41 refers to dictionary data (not shown) stored in the storage unit 3, morphologically analyzes the input text data, converts it into accented phonemes, and outputs them to the phoneme clustering unit 42. is there. Here, the morpheme refers to the minimum character string that has no meaning if it is made finer than this, and the morpheme analysis is to analyze the sentence by breaking it down to the morpheme level. For example, the sentence “Today's weather is sunny” in Japanese can be separated as “Today's weather. In this embodiment, the text data analysis means 41 refers to the pronunciation word storage means 31 and the connection probability storage means 32 and converts them into accented phonemes. The text data analysis means 41 may use the connection probability of “part of speech” instead of “word”. Further, the text data analysis means 41 may convert into phonemes with accents based on “pronunciation rules determined in advance regarding word connection” instead of “connection probability”.

  The phoneme clustering means 42 clusters (classifies) the accented phonemes converted by the text data analysis means 41 with the phonemes arranged before and after the accented phonemes, and outputs them to the phoneme accent clustering means 43. The phoneme clustering means 42 performs clustering by a known method described in Non-Patent Document 1, for example. Here, clustering refers to classifying a set of phonemes to be classified into a predetermined cluster (group) by paying attention to some attribute. That is, the attribute attracting attention by the phoneme clustering means 42 is the type of phoneme arranged before and after the accented phoneme (phoneme environment before and after).

  The phoneme accent clustering means 43 clusters the phonemes with accents clustered by the phoneme clustering means 42 with the accents of phonemes arranged before and after the accented phonemes. That is, the phoneme accent clustering means 43 performs clustering by paying attention to the attribute of phoneme accents (accent environment before and after) arranged before and after the accented phoneme. The phoneme accent clustering means 43 uses the pitch of the front and rear sounds as the front and rear accent environment.

  Specifically, the phoneme accent clustering means 43 is low when, for example, the focused phoneme (central phoneme) is a Japanese vowel (the vowel), the accents of the preceding and following vowels are high (high), and low. Since there are cases (low), it is classified into the following seven types (cluster “1” to cluster “7”). Note that high and low indicate whether the fundamental frequency is higher or lower than a predetermined value.

(Cluster “1”) = (Low, Low, Low) or (High, High, High)
(Cluster “2”) = (Low, Low, High)
(Cluster “3”) = (Low, High, Low)
(Cluster “4”) = (Low, High, High)
(Cluster “5”) = (High, Low, Low)
(Cluster “6”) = (High, Low, High)
(Cluster “7”) = (High, High, Low)
Here, for example, cluster “3” (low, high, low) is a cluster composed of phonemes with low accent of the previous vowel, high accent of the vowel (center phoneme), and low accent of the back vowel. Means.

If the accent of the vowel is “low” and there is no previous vowel, the accent of the previous vowel is regarded as “low” and classified as cluster “1” or cluster “2”. Further, when the accent of the vowel is “low” and there is no back vowel, the back vowel accent is regarded as “low” and is classified into cluster “1” or cluster “5”.
Similarly, when the accent of the vowel is “high” and there is no previous vowel, the accent of the previous vowel is regarded as “high” and is classified into cluster “1” or cluster “7”. Further, when the accent of the vowel is “high” and there is no back vowel, the back vowel accent is regarded as “high” and is classified into cluster “1” or cluster “4”.

The phoneme accent clustering means 43 indicates, for example, the cluster (the level of the accent of the previous vowel and the level of the accent of the back vowel) if the target phoneme (central phoneme) is a Japanese consonant. In this case, it is classified into the following three types (cluster “8” to cluster “10”). If there is no vowel before or after, it is classified into cluster “8”.
(Cluster “8”) = (Low, Low) or (High, High)
(Cluster “9”) = (Low, High)
(Cluster “10”) = (High, Low)

  The voice data search means 44 calculates the connection score of the voice data string generated by combining the voice data corresponding to the phonemes with accents clustered by the phoneme accent clustering means 43 by viterbi search, and the connection score is the maximum. Is searched from the voice database 33 and output to the voice data correction means 45. The voice data search means 44 uses phonemes as search units. The Viterbi search is a method of leaving only the history of a hypothesis (combination of speech data corresponding to accented phonemes) that gives the best (maximum) score.

  Here, how to calculate the connection score will be described. Assume that two phonemes with accents are phoneme A and phoneme B, respectively, and phoneme B is connected behind phoneme A. The connection score Sc (A, B) in this case can be obtained by, for example, the following mathematical formula (1).

In Equation (1), p ^E _A represents the fundamental frequency at the end (termination) of phoneme A, p ^I _B represents the fundamental frequency at the beginning (tip) of phoneme B, and c ^E _jA represents j The feature amount at the end (end) of the phoneme A in the dimension is represented, and c ^I _jB represents the feature amount at the beginning (tip) of the phoneme B in the jth dimension.

In Equation (1), (a) and (b) are clusters T ^E _A , T including a triphone at the end of phoneme A in the j-th dimension (or triphone if phoneme A is a triphone). hidden Markov models _{^{I B (HMM: hidden Markov model}} ) represents a variance value of, (c) and (d) cluster c is included the triphone end of phonemes a in the j-th dimension (a) HMM D represents the total number of dimensions of the feature quantity, ω ₇ and ω ₈ represent positive weights, and a represents a positive constant. Note that δ _AB is 0 when the phoneme A and the phoneme B are continuously in the speech database 33, and 1 when there is no phoneme B.

  The voice data correction unit 45 corrects the voice data string searched by the voice data search means 44 with voice data strings before and after the voice data string. Specifically, the audio data correction unit 45 averages the fundamental frequency of the audio data sequence searched by the audio data search unit 44, and calculates an audio data sequence in which a deviation between the average value and individual audio data is corrected. The data is output to the audio data connecting means 46. For this correction, the method described in JP-A-2-47700 is applied. The voice data correction unit 45 is not an essential component and may not be provided in the voice synthesizer 1.

  The audio data connecting unit 46 connects (connects) the audio data included in the audio data sequence corrected by the audio data correcting unit 45 and outputs the audio data to the audio output device SP. When the speech synthesizer 1 does not include the speech data correction unit 45 described above, the speech data connection unit 46 connects the speech data strings searched by the speech data search unit 44 and outputs them to the speech output device SP. . The sound output device SP may be anything as long as it outputs sound, and is, for example, a speaker or a display device (liquid crystal display, CRT (Cathode Ray Tube), etc.) including the speaker. Note that the voice synthesizer 1 may output to the voice output device SP via a communication circuit network (not shown) such as the Internet by a communication interface (not shown).

[Operation of speech synthesizer]
Next, the operation of the speech synthesizer will be described with reference to FIG. FIG. 2 is a flowchart showing the operation of the speech synthesizer shown in FIG. First, the speech synthesizer 1 inputs text data by the input unit 2 (step S1), and the text data analysis unit 41 converts the input text data into a phoneme with accent by performing morphological analysis (step S2). Then, the speech synthesizer 1 performs clustering with the phoneme clustering means 42 (step S3), and the phoneme accent clustering means 43 performs clustering with the phoneme accents (step S4).

  Subsequently, the speech synthesizer 1 searches the speech database 33 for a speech data string having the maximum connection score based on the mathematical formula (1) by the speech data search means 44 (step S5). Then, the voice synthesizer 1 corrects the searched voice data string by the voice data correction unit 45 (step S6). Thereby, the audio data string can be smoothly connected to the preceding and succeeding audio data strings. Then, the speech synthesizer 1 connects the corrected audio data strings by the audio data connecting means 46 (step S7). Thereby, the connected speech synthesis data is output as synthesized speech from the speech output device SP. If the speech synthesizer 1 does not include the speech data correction unit 45, the process proceeds to step S7 following step S5.

[Specific speech synthesis example]
Next, a specific speech synthesis example of the speech synthesizer 1 will be described with reference to FIG. 3 (refer to FIG. 1 as appropriate). FIG. 3 is an explanatory diagram showing a specific speech synthesis example of the speech synthesizer shown in FIG. 1, where (a) is an output example of text data analysis means, and (b) is phoneme clustering means and phoneme accent clustering. An output example of the means is shown.

  As an example, it is assumed that “<Sentence> Good morning <End of sentence>” is input as text data (input Japanese text) to the input means 2 of the speech synthesizer 1. Then, the speech synthesizer 1 converts the input Japanese text into accented phonemes by the text data analysis means 41, and as shown in FIG. 3A, the phoneme string “o (low) ha (high)”. yo: (high) go (high) za (high) i (high) ma (high) su (low) "is output.

  Then, as shown in FIG. 3B, the speech synthesizer 1 uses the phoneme clustering means 42 to convert each phoneme 301 constituting the continuous phoneme sequence shown in FIG. A string of accented phonemes (triphones, phoneme clustering 302) clustered in consideration of phonemes arranged before and after (phoneme environments before and after) is output. That is, “o (low) + h o−h + a ha− (high) + y a−y + o: yo− (high) + go = −g + o go− (high) + z o−z + a z−a (high) + I a−i (high) + m i−m + a m−a (high) + s a−s + u su− (low) ”is output. Here, for example, in the triphone “oh−a + a”, the central phoneme is “h”, and the phoneme (preceding phoneme) arranged in front of the central phoneme is “o”, and is arranged behind the central phoneme. This means that the phoneme (following phoneme) is “a”.

  Further, the speech synthesizer 1 uses the phoneme accent clustering unit 43 to convert the phonemes with accents (phoneme clustering 302) output from the phoneme clustering unit 42 into phonemes (preceding phonemes) arranged before and after the accented phoneme (central phoneme). , The subsequent phoneme) is output to the speech data search means 44 as a string of accented phonemes (phoneme accent clustering 303) clustered in consideration of the accent (accent environment before and after). That is, “o2 + h o−h ^ 9 + a h−a ^ 4 + y a−y ^ 8 + o: yo − ^^ 1 + go: −g ^ 8 + o g−o ^ 1 + z o−z ^ 8 + a z−a ^ 1 + ia -I ^ 1 + m i-m ^ 8 + a m-a ^ 6 + s a-s ^ 10 + u s-u ^ 5 "is output. Here, for example, “o−h ^ 9 + a” is a phoneme “h” whose central phoneme belongs to the above-mentioned (cluster “9”), a previous phoneme (preceding phoneme) is “o”, and a rear This means that the phoneme (subsequent phoneme) is “a”.

  Subsequently, the speech synthesizer 1 searches the speech database 33 for speech data combinations belonging to the corresponding cluster by the speech data search means 44, and outputs a speech data string that maximizes the connection score. At this time, a phoneme (for example, “o−h ^ 9 + a” which is a set of the preceding phoneme, the classified central phoneme and the subsequent phoneme, and “h ^ 9” which is the classified central phoneme) is connected as a search unit. A combination of voice data that maximizes the score is searched from the voice database 33.

  Specifically, the voice data search means 44 first calculates mathematical expressions for all combinations of voice data strings connected from “o ^ 3 + h” in the voice database 33 to “oh−9 + a” in the voice database 33. Calculate the connection score determined using (1).

  As a result of the calculation, in the voice database 33, the voice data having the largest concatenation score in the voice data string (output candidate) of “o ^ 3 + h” connected to the first “oh9 + a” is voice data. It is recorded in the memory 34 by the search means 44. Then, the calculation of the connection score obtained using Equation (1) and the recording operation are executed for all “oh−9 ^ a” in the speech database 33.

  Similarly, “h−a ^ 4 + y” is also included in the voice data string (output candidate) of “o−h ^ 9 + a” connected to the first “h−a ^ 4 + y” in the voice database 33. The voice data search means 44 records the one with the largest connection score. Then, the calculation of the connection score obtained using Expression (1) and the recording operation are executed for all “ha−4 + y” in the voice database 33.

  In the same manner, a concatenated score for connecting each accented phoneme is obtained, and finally, the combination of speech data strings (output candidates) corresponding to the input Japanese text has the largest concatenated score. Will be searched. The searched audio data string is smoothly connected by the audio data correcting unit 45, connected by the audio data connecting unit 46, and finally becomes one audio data, which is output from the audio output device SP as synthesized speech. .

  According to the first embodiment, the speech synthesizer 1 searches the speech database 33 for accented phonemes in consideration of the front and back accent environments in addition to the front and back phoneme environments. It becomes possible to create. As a result, it is possible to prevent the deterioration of naturalness that the synthesized speech is heard unnaturally.

(Second Embodiment)
[Configuration of speech synthesizer]
FIG. 4 is a functional block diagram showing the configuration of the speech synthesizer according to the second embodiment.
As shown in FIG. 4, the speech synthesizer 1A has the same configuration as the speech synthesizer 1 shown in FIG. 1 except that the storage unit 3A and the speech synthesis control unit 4A have different functions. Are denoted by the same reference numerals, and description thereof is omitted.

The storage means 3A includes a phoneme string list (phoneme string storage means) 35.
The phoneme string list 35 stores phonemes arranged before and after accented phonemes, and strings of accented phonemes clustered by the accents of the phonemes, and is a general recording medium such as an HDD. As a method for creating the phoneme string list 35, a method described in Japanese Patent Application Laid-Open No. 2005-70164 can be applied.
Note that the unit speech stored in the speech database 33 is based on a phoneme or phoneme sequence (phoneme sequence division hypothesis data) in which accents are assigned to all vowels.

  The voice synthesis control unit 4A includes a phoneme string division unit 51 and a voice data search unit 44A. The phoneme string dividing means 51 divides the input text data converted into the phonemes clustered by the phoneme accent clustering means 43 into accented phoneme strings (phoneme string dividing hypothesis data) stored in the phoneme string list 35, This is output to the voice data search means 44A. The phoneme string dividing means 51 outputs phoneme string dividing hypothesis data based on the number of connection points and the number of phonemes, for example. Note that the method described in Japanese Patent Laid-Open No. 2005-70165 can be applied to the division of accented phonemes into columns.

The speech data search means 44A calculates a concatenation score of speech data strings generated by combining speech data corresponding to accented phoneme strings divided by the phoneme string split means 51 by viterbi search, and the concatenation score is The maximum audio data string is searched from the audio database 33 and output to the audio data correcting means 45. If the audio data correcting means 45 is not provided, the audio data connecting means 46 is output.
The voice data search means 44A uses a string of phonemes with accents (phoneme string) as a search unit. This sequence of phonemes with accents is represented by the example shown by the phoneme accent clustering 303 in FIG. 3B, and corresponds to the concatenation of the phoneme accent clustering 303. It is expressed by “(o ^ 2 h ^ 9) + a” or “o− (h ^ 9 a ^ 4 y ^ 8) + o”, which is a set of phonemes and / or subsequent phonemes.

[Operation of speech synthesizer]
Next, the operation of the speech synthesizer 1A will be described with reference to FIG. FIG. 5 is a flowchart showing the operation of the speech synthesizer shown in FIG. The speech synthesizer 1A sequentially processes steps S21 to S24. These steps S21 to S24 are the same as steps S1 to S4 shown in FIG. Subsequent to step S24, the speech synthesizer 1A divides the input text data converted into phonemes clustered by the phoneme accent clustering unit 43 by the phoneme sequence dividing unit 51 into accented phoneme sequences (step S25). Then, the speech synthesizer 1A sequentially processes steps S26 to S28. These steps S26 to S28 are the same as steps S5 to S7 shown in FIG. However, in step S26, the speech data search means 44A does not target accented phonemes (for example, “o−h ^ 9 + a” or “h ^ 9”), but instead arranges a sequence of accented phonemes. A speech data string having the maximum connection score is searched for (for example, “(o ^ 2 h ^ 9) + a”, “o− (h ^ 9 a ^ 4 y ^ 8) + o”).

  Next, the process of step S25 described above will be specifically described. As an example, it is assumed that “<start of sentence> good morning <end of sentence>” is input as text data (input Japanese text) to the input unit 2 of the speech synthesizer 1A. In this case, as shown in FIG. 3A, the speech synthesizer 1A uses the text data analysis unit 41 to generate a continuous phoneme string “o (low) ha (high) yo: (high) go (high). ) Za (high) i (high) ma (high) su (low) ". Then, as shown in FIG. 3B, the phoneme string dividing unit 51 of the speech synthesizer 1A receives the accent clustered phoneme string “o ^ 2 + h o−h ^ 9 + a h−a ^ 4 + y a−y ^. 8 + o: yo: ^ 1 + go: -g ^ 8 + o g-o ^ 1 + z o-z ^ 8 + a z-a ^ 1 + i a-i ^ 1 + m i-m ^ 8 + a m-a ^ 6 + s a-s ^ 10 + u s -U ^ 5 "is input. The phoneme string dividing means 51 converts the input phoneme string consisting of the 14 phonemes (central phonemes) into “(o ^ 2 h ^ 9) + a”, “h− (a ^ 4 y ^ 8 o: ^ 1 g ^ 8) + o "," g- (o ^ 1 z ^ 8 a ^ 1 i ^ 1) + m ", and" i- (m ^ 8 a ^ 6 s ^ 10 u ^ 5) " This is divided into four phoneme string division hypothesis data (accented phoneme string) and output as a final output result.

  According to the second embodiment, the speech synthesizer 1A divides the input text data into accented phoneme strings registered in the phoneme string list 35 in advance, and the accented phoneme strings are converted into a speech database. Since 33 is used as a search unit for searching, it is possible to prevent searching for phonemes having different phoneme environments before and after searching, and to shorten the time required for speech synthesis processing. As a result, it is possible to prevent deterioration in sound quality of the synthesized speech synthesis data.

  As mentioned above, although this invention was demonstrated based on each embodiment, this invention is not limited to these. For example, it is possible to regard each configuration of the speech synthesizer 1 (1A) as a speech synthesis method in which each configuration is regarded as one process, or as a speech synthesis program in which the processing of each configuration is described in a general-purpose computer language. is there. In this case, the same effect as the speech synthesizer 1 (1A) can be obtained.

In each embodiment, the voice database 33 has been described as having accents added to the base phoneme vowels, but accents may be added to the vowels. For devoicing, it may be possible to give high or low or strong or weak as vowels.
Further, in each embodiment, the phoneme accent clustering unit 43 uses the pitch of the front and rear sounds as the front and rear accent environment, but at least one of the pitch of the front and rear sounds, strength, shortness, and rhythm is used. Use it.

  Next, examples in which the effects of the present invention have been confirmed will be described. Clustering considering synthesized speech (example) synthesized by clustering in consideration of up to and including the accent environment using the speech synthesizer 1A of the second embodiment, and the conventional phoneme environment as in the past. The synthesized speech (comparative example) synthesized by doing so was compared in terms of naturalness (whether it sounds more natural).

[Comparison experiment]
The data stored in the speech database 33 in advance is composed of 25484 sentences spoken by the Morita announcer in the NHK news database from June 3, 1996 to June 22, 2001, and the balance sentence read by the Morita announcer (phoneme environment). A total of 79.5 hours of 99 sentences, which is a sentence created by balancing the two), the total number of triphones is 3.56 million, and the number of different triphones is 8452. Further, 96 sentences (phoneme number 13267) uttered by the announcer Morita in the program “NHK News 10” from June 25 to June 29, 2001 were used as the evaluation text.

  First, a comparative experiment will be described. This paired comparison experiment is performed using speakers in a soundproof room, and the subjects of the experiment are three women who have experience in voice evaluation. In this comparative comparison experiment, all 96 sentences as evaluation texts were listened to for the example and the comparative example, and each listening was limited to once. In each trial of the comparative experiment, the example and the comparative example were presented in pairs in a random order, and the subject was instructed to select the subject that felt more natural. This comparative experiment was conducted while having each subject take a break at an appropriate time interval.

  As a result of this comparison experiment, with respect to a total of 56.0% speech, the synthesized speech synthesized by the speech synthesizer 1A (example) does not consider the accent environment before and after as in the past ( Compared to Comparative Example), it was evaluated as natural. Using the binomial test, the risk is 5%, and this difference is significant.

[5-level quality evaluation experiment]
Next, a five-stage quality evaluation experiment will be described. In this five-step quality evaluation experiment, the synthesized speech (Example) created using the speech database 33 considering the front and back accent environments and the synthesized speech created using the speech database not considering the front and back accent environments ( A comparative example) and natural speech were evaluated for quality in five stages.

  This five-step quality evaluation experiment is performed using a speaker in a soundproof room as in the comparative comparison experiment, and the subjects are three women who have experience in voice evaluation. In each trial, the evaluation data is presented to the subject in a random order, and the subject evaluates the difference in naturalness. The evaluation of this naturalness is “5” (natural), “4” (unnatural part but not bothered), “3” (somewhat worried), “2” (worried) , "1" (very worrisome) was decided to perform quality evaluation in five stages. Prior to the quality evaluation, the subjects were given 3 sentences of natural speech and given an evaluation standard (instruction) as to how much speech should be heard naturally. In addition, since the news sentence actually broadcasted is used as the evaluation text, the average length of one sentence is about 10 seconds. Therefore, listening is limited to one time and breaks are made at appropriate intervals. I went while holding it.

  Here, an average opinion score (MOS) will be described with reference to FIG. FIG. 6 is a graph showing the results of a five-stage quality evaluation experiment using the speech synthesizer shown in FIG. From the graph shown in FIG. 6, the MOS in the example is “4.19”, and the MOS in the comparative example is “3.95”. It can be said that the MOS “4.19” of the example has a naturalness between “natural” and “not natural but there is an unnatural part”, and the MOS “3.95” of the comparative example It can be said that it is a good evaluation. The natural voice MOS was "4.99".

  Further, as shown in FIG. 6, in the embodiment, 49% of the total synthesized speech is evaluated as “5” (natural). For this reason, in the embodiment, it can be said that voice data having the same quality as natural voice is synthesized with high frequency. In addition, 8% of the total received evaluations of “2” and “1”.

It is a functional block diagram which shows the structure of the speech synthesizer which concerns on 1st Embodiment. It is a flowchart which shows operation | movement of the speech synthesizer shown in FIG. It is explanatory drawing which shows the specific speech synthesis example of the speech synthesizer shown in FIG. 1, (a) is an example of an output of a text data analysis means, (b) is an output example of a phoneme clustering means and a phoneme accent clustering means. Is shown. It is a functional block diagram which shows the structure of the speech synthesizer of 2nd Embodiment. It is a flowchart which shows operation | movement of the speech synthesizer shown in FIG. It is a graph which shows the result of the 5-step quality evaluation experiment using the speech synthesizer shown in FIG.

Explanation of symbols

DESCRIPTION OF SYMBOLS 1,1A Speech synthesizer 2 Input means 3,3A Storage means 31 Pronunciation word storage means 32 Connection probability storage means 33 Speech database 34 Memory 35 Phoneme string list (phoneme string storage means)
4, 4A Speech synthesis control means 41 Text data analysis means 42 Phoneme clustering means 43 Phoneme accent clustering means 44, 44A Speech data searching means 45 Speech data correcting means 46 Speech data connecting means 51 Phoneme sequence dividing means SP Speech output device

Claims (4)

Text data analysis means for converting input text data into accented phonemes by performing morphological analysis,
Phoneme clustering means for clustering accented phonemes converted by the text data analyzing means with phonemes arranged before and after the accented phonemes;
Phoneme accent clustering means for clustering accented phonemes clustered by the phoneme clustering means with phoneme accents arranged before and after the accented phonemes;
A speech database for storing speech data corresponding to phonemes arranged before and after the accented phonemes and accented phonemes clustered by the accents of the phonemes;
A concatenation score of a speech data sequence generated by combining speech data corresponding to accented phonemes clustered by the phoneme accent clustering means is calculated by viterbi search, and a speech data sequence that maximizes the concatenation score A voice data search means for searching from a voice database;
Voice data connection means for connecting the voice data strings searched by the voice data search means;
A speech synthesizer comprising: Text data analysis means for converting input text data into accented phonemes by performing morphological analysis,
Phoneme clustering means for clustering accented phonemes converted by the text data analyzing means with phonemes arranged before and after the accented phonemes;
Phoneme accent clustering means for clustering accented phonemes clustered by the phoneme clustering means with phoneme accents arranged before and after the accented phonemes;
Phoneme string storage means for storing a phoneme arranged before and after the accented phoneme and a string of accented phonemes clustered by the accent of the phoneme;
A speech database for storing speech data corresponding to accented phonemes stored in the phoneme string storage means;
Phoneme string dividing means for dividing the text data converted into phonemes clustered by the phoneme accent clustering means into accented phoneme strings stored in the phoneme string storage means;
A speech data sequence generated by combining speech data corresponding to the accented phoneme sequence divided by the phoneme sequence partitioning unit is calculated by viterbi search, and the speech data that maximizes the connection score Speech data search means for searching for a sequence from the speech database;
Voice data connection means for connecting the voice data strings searched by the voice data search means;
A speech synthesizer comprising: In order to synthesize speech corresponding to text data,
Text data analysis means for analyzing input morphological data and outputting accented phonemes,
Phoneme clustering means for clustering accented phonemes converted by the text data analyzing means with phonemes arranged before and after the accented phonemes;
Phoneme accent clustering means for clustering accented phonemes clustered by the phoneme clustering means with phoneme accents arranged before and after the accented phonemes;
A concatenation score of speech data sequences generated by combining speech data corresponding to accented phonemes clustered by the phoneme accent clustering means is calculated by Viterbi search, and a speech data sequence having the maximum concatenation score is searched. Voice data search means to
Voice data connecting means for connecting voice data strings searched by the voice data searching means;
A speech synthesis program characterized by functioning as In order to synthesize speech corresponding to text data,
Text data analysis means for converting input text data into accented phonemes by morphological analysis,
Phoneme clustering means for clustering accented phonemes converted by the text data analyzing means with phonemes arranged before and after the accented phonemes;
Phoneme accent clustering means for clustering accented phonemes clustered by the phoneme clustering means with phoneme accents arranged before and after the accented phonemes;
Phoneme string dividing means for dividing the text data converted into phonemes clustered by the phoneme accent clustering means into accented phoneme strings;
A speech data sequence generated by combining speech data corresponding to the accented phoneme sequence divided by the phoneme sequence partitioning unit is calculated by viterbi search, and the speech data that maximizes the connection score Voice data search means for searching for a sequence;
Voice data connecting means for connecting voice data strings searched by the voice data searching means;
A speech synthesis program characterized by functioning as

Images