JPS63172199A - Voice synthesizer - Google Patents

Abstract

(57) [Summary] This bulletin contains application data before electronic filing, so abstract data is not recorded.

Description

[Detailed description of the invention] (a) Industrial application field The present invention relates to a speech synthesis device based on rules, and by inputting an arbitrary character string to the device, it is possible to obtain any synthesized speech. It is something.

(b) Conventional technology When generating synthesized speech from sentences that are often used in daily life, such as sentences containing kanji and kana, word identification, reading assignments, accent position determination, breath position and depth, etc. requires linguistic processing such as determining the

Therefore, in order to perform these processes, reading words,
Use a word dictionary in which parts of speech, accent forms, etc. are registered as dictionary information. However, since there is a limit to the number of words that can be registered in the dictionary, correct word identification is not performed for all sentences, resulting in unregistered words. Conventionally, when an unregistered word occurs, processes such as applying an on-yomi reading to the kanji and reading it intermittently one character at a time, or generating a signal sound different from the audio output as disclosed in Japanese Patent Application Laid-Open No. 60-73725, have been carried out. Therefore, the sound quality was unnatural for unregistered words.

A speech synthesis device based on such conventional rules is shown in FIG.
According to the device shown in the figure, for example, when a sentence such as "He requested the discontinuation of publication" is input from the outside and is encoded and stored in the input sentence buffer memory (1), the input sentence buffer memory (1) The address counter (2) of the word dictionary memory (3) is incremented sequentially from the beginning, and the address counter (4) of the word dictionary memory (3) is incremented sequentially from the beginning. )
A matching circuit (5) performs matching between the information and the heading information (encoded word character string) in the word dictionary memory (3) as shown in FIG. Matching circuit (
According to 5), if several characters of the contents of the input sentence buffer memory (1) match the heading information of the word dictionary memory (3), Matuchisoge No. 13 (6) is generated. By the matching signal (6), the dictionary information such as pronunciation, part of speech, accent form, etc. of the word dictionary memory (3) as shown in Fig. 3 is written into the identified word memory (7), and the identified word memory (7) is also written. The address counter (8) is incremented, the address counter (2) of the input sentence buffer memory (1) is incremented by the number of characters of the matched word, and the address counter (4) of the word dictionary memory (3) is incremented. can be reset.

In the example sentence, the word "he" is first identified (matched), and then "ha" is identified.

However, when the address counter (2>) points to the character "obsolete" in the input sentence buffer (1), if the word "discontinued" does not exist in the word dictionary memory (3), Matching processing is performed for the Eastern word ``'' and then for the word ``publish''. In other words, the address counter (2) first reads the character ``waste゛'' and matching is performed, but if no matching word exists. First, when the unregistered word detection circuit (9) generates an unregistered detection signal (10), one character is registered as an unregistered word in the identification m-word memory (7). (2) Matching is performed. At this time, the word "kanpa" is identified and an identification signal (6) is generated. Figure 4 shows the identified word memory in the above example sentence.
It shows the content of the sentence, in which ``obsolete'' is an unspoken O word and ``published°'' is a suffix. Unregistered words in the identified word memory (7) are stored in the kanji dictionary memory (11), which maps the most frequently used pronunciations to the - character kanji, and in the matching circuit (
12), the matching is performed and the reading of the kanji is given. That is, "Yes" is given to "Abandonment".

Then, for "kanpa", "kan" is given as a suffix according to the word dictionary memory (3). Furthermore, (13) is the address counter of the kanji dictionary memory (11). Identification word memory (7) The part-of-speech information is matched by the accent combination rule table (14) and the matching circuit (15), and the accent phrases in which the words are combined, their accent positions, the degree of combination between words, etc. are stored in the accent phrase memory (16). Section 5 shows some of the rules for combining words to generate accent phrases, and matching with this rule is repeated until there are no more rules to apply. It is described in Proceedings of the Kansai Branch Federation Conference of Electrical Engineering Associations G15-11 'Ruled Synthesis Using Kanji-Kana-Mixed Sentences' in 2013. Also, the method of determining the position of the next breath, etc.
Figure 6, which is detailed in Publication No. 195596, shows example sentences.
It shows accent phrases, accent positions, etc. Following the information in the accent phrase memory (16), the speech rule synthesis section (17) connects the phoneme signals of the phoneme memory (18) and outputs synthesized speech from the speaker 19). Regarding the generation of pitch patterns in this synthesis section (17), see, for example, 2 of the Proceedings of the Acoustical Society of Japan.
-1-14' Examination of prosodic rules in rule synthesis of sentences''
has been reported.

Therefore, in the example sentence, “discontinued publication” is “high”.

“Kan゛,” “wo”, etc. - letters can be pronounced intermittently, and the synthesized speech of unregistered words becomes unnatural.

C) Problems to be Solved by the Invention The present invention attempts to alleviate the unnaturalness of synthesized speech in the unregistered word portions when synthetic speech is generated from sentences containing arbitrary kanji, kana, etc. It is.

(d) Means for Solving the Problems The speech synthesis device of the present invention includes an input sentence buffer memory in which character strings constituting a sentence are coded and stored, and a character type for determining each character type of the character string in the memory. A discrimination circuit, a word dictionary memory in which dictionary information such as part of speech, pronunciation, accent, etc. of words are registered, and words registered in this memory are matched with character strings in the sentence buffer memory, and character strings are identified by word punctuation. and an unregistered word processing unit that processes a continuous character string of the same type obtained from the character type discrimination circuit as an unregistered word when there is a character string that is not identified by the identification unit. and generates a synthesized speech according to the pronunciation of the grudge dictionary memory and the accent report for the word identified by the word identification unit, and
and a speech synthesis section that generates synthesized speech according to the pronunciation corresponding to each character of the character string of one word processed by the unregistered post-processing section, as well as typical actants for each word. It is something.

(e) Effect: According to the uninvented speech synthesis device, even unregistered words can be pronounced by adding a representative pronunciation of each letter and a representative accent position as a single word. This increases the probability of more natural and accurate pronunciation.

(F) Embodiment FIG. 1 shows a speech synthesis apparatus according to the rules of the present invention. In the device of the present invention shown in the same figure, the example sentence for the conventional device shown in FIG. Once stored, the address counter (2) of the input sentence buffer memory (1) is reset. on the other hand,! The address counter (4) of the word dictionary memory (3) is incremented sequentially from the beginning, and the matching circuit (4) matches the contents of the input sentence buffer memory (1) with the heading information of the word dictionary memory (3). 5). Similar to the conventional example, when a match is made, a matching number (6) is generated, the switch Swl is closed, and the 41 m dictionary memory (
Dictionary information such as pronunciation, part of speech, accent form, etc. as shown in FIG. 3 in 3) is written into the identified word memory (7), and the address counter (8) of the identified word memory 7) is incremented. Further, the address counter (2) of the input sentence buffer memory (1) is incremented by the number of characters of the matched word, and the address counter (4) is reset.

On the other hand, when the matching circuit (5) cannot find a match, when it cannot identify "out of print" in the example sentence and "out of print" cannot be identified, the unregistered word detection circuit (9) first identifies "out of print". The unspoken 0 word detected No. Z is sent to the unregistered word processing circuit (20). At this time, the character type discrimination circuit (21) determines the type of character output from the input sentence buffer (1) (for example, kanji, It distinguishes between characters (Hiromakana, Katakana, etc.) and outputs a signal indicating the type of character.

The difference between the device according to the present invention shown in FIG. 2 and that shown in FIG. When an unregistered word detection signal for each character is received from the word detection circuit (9), the switch Sw2 is opened until the character type discrimination signal from the latter circuit 21) changes.
Identification word memory by forcibly considering it as a group of words (7
).

For example, in the above case, since "discontinued" is unregistered, the character string "discontinued" including "publish", which is also a kanji character type, can be considered as a - word.Kirani, this unregistered word processing circuit (
20) assumes that all parts of speech are nouns for QL[, which consists of unregistered words, and that its accent form is 0 form1, and stores these part of speech and accent form information in the identified word memory ( 7) is added to the string of unregistered words.

In other words, the most likely unregistered words are nouns with a wide variety of types, and the most common words are words with the 0 form, which is a flat pronunciation with no accent position, so it is possible to add representative accents. There is.

FIG. 7 shows the identified word memory (7) for example sentences, and the reading of the unregistered word "discontinued" is determined by the operation of the kanji dictionary memory (11) and the matching circuit (12). The representative readings of "1 reading" are read from the memory (11) and written in the "1 reading" area. The contents of the identified word memory (7) are stored in a matching circuit (15) that matches the accent position of each phrase by connecting an appropriate number of words using the rules shown in Figure 5 of the accent combination rule table (14). Decide and accent phrase memory (1
6). Figure 8 shows accent phrase memory (16)
It shows the contents. Such a join rule table (14)
Regarding the method of forming accent phrases using
) page 2253 to page 2265 for details on Japanese accent combination rules for the purpose of speech synthesis from paper 1 text.

Thus, as shown in Figure 8, the accent phrase memory (16) contains example sentences such as “he”, “the discontinuation of publication”, and “requested”.
It is divided into three accent phrases, and especially for the second phrase that includes the unknown word ``discontinued publication'', information on the reading ``haigan wo゛, noun phrase, 0-form accent phrase (no accent position)'' is provided. is determined, which is likely to make the synthesized speech more natural.

That is, the speech rule synthesis section (17) connects the speech signal of the phoneme memory (18) according to the information of the accent phrase memory (16), and further adds an accent to the speech signal to generate a natural synthesized speech r Kareha, Haikanwo, etc. The sound will be emitted from the speaker (19).

In the above-described embodiment, a method has been described in which a character string of the same character type including an unregistered word is regarded as one word, but it is also possible to consider a character string of the same character type including an unregistered word as one word. In other words, to give an example of the four consecutive kanji characters "-tai tsu tsu tsu tsu tsu", even if "-" and " ゛ are registered words, "
Since ``tai'' and ``soku'' are unregistered, it is preferable to consider ``-tai ni tatsu'' as a group of words, rather than assuming the word ``taku ni tatsu''.

(G) Effects of the Invention As is clear from the above description, the speech synthesis device of the present invention has the following effects:
Since a representative pronunciation of each character and a representative accent position as a single word are assigned to continuous strings of the same character type in which unregistered words exist, the probability of speech synthesis with correct pronunciation and accent is increased, and the synthesized speech is improved. Naturalness can be improved.

[Brief explanation of the drawing]

FIG. 1 is a block diagram showing the configuration of an embodiment of the speech synthesis device of the present invention, FIG. 2 is a block diagram showing the configuration of a conventional device, and FIGS. 3 to 8 are memory diagrams. (1)... Input sentence buffer memory, (3)... Word dictionary memory, (7)... Identification word memory, (9)...
... unregistered word detection circuit, (14) ... accent combination rule table, (16) ... accent phrase memory,
(17)...Speech rule synthesis unit, (20)...Unregistered word processing circuit, (21)...Character type discrimination circuit.

Claims (1)

[Claims] (1) An input sentence buffer memory in which character strings forming a sentence are encoded and stored; a character type discrimination circuit that determines each character type of the character string in the memory; and dictionary information such as part of speech, pronunciation, and accent of words. a registered word dictionary memory; a word identification unit that performs matching between registered words in the memory and character strings in the text buffer memory to identify and distinguish character strings on a word-by-word basis; and character strings that are not identified by the identification unit. an unregistered word processing unit that processes a continuous character string of the same type obtained from the character type discrimination circuit as a single unregistered word when there is a character type discriminator; It also generates synthesized speech according to the accent information, and also generates synthetic speech according to the pronunciation corresponding to each character of the character string of one word processed by the unregistered word processing unit, as well as representative accents for each word. A speech synthesis device comprising a speech synthesis section that generates.