JP2006227425A - Speech reproducing device and utterance support device - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a speech reproducing device and an utterance support device for better communication with an audience by expressing a word or phrase difficult for a speaker or speech synthesis to pronounce in other words. <P>SOLUTION: The utterance support device 1 includes an input means 21 for acquiring document text data in the device, a word string extracting means 22 for extracting a word string, a substituent candidate word string retrieving means 23 for retrieving a substitutable candidate word string by matching the word string against a synonym database 24, a pronunciation difficulty database 26 in which difficulties in pronouncing respective word strings are quantitatively recorded in advance, a word string substituting means 25 for acquiring difficulties in pronouncing a word string to be substituted and a word string as a substitution candidate from the pronunciation difficulty database 26 respectively and selecting the word string having the minimum difficulty in pronouncing for substitution, and an output means 12 for outputting a text for dictation. <P>COPYRIGHT: (C)2006,JPO&NCIPI

Description

  The present invention relates to language processing technology, and in particular, a method of converting an original sentence so that words that are easy to hear by speech synthesis processing are used, or words that are easy for a speaker to pronounce when speaking in a foreign language. It is related to.

  In recent years, there are increasing opportunities to hear speech that has been artificially synthesized in many scenes of daily life. The sound quality has also improved considerably, and when outputting specific words, it is possible to achieve easy-to-listen voice output by finely defining intonation and the like.

  However, if speech synthesis is performed by inputting arbitrary text such as reading out Web content, it is still difficult to hear and the content may not be transmitted accurately. On the other hand, research to make it easy to hear by adding appropriate inflection to text is also progressing, but there is a problem that is unsuitable for high-speed playback due to heavy processing load related to speech synthesis processing.

Another case where the listener is difficult to hear is that when speaking in a foreign language, accurate pronunciation is difficult due to the difference in pronunciation method from the native language, making it difficult to communicate. In particular, since the manuscripts are prepared in advance for presentations at international conferences, etc., there are few grammatical errors. There are many things that are not transmitted.
As is well known in foreign language learning, for example, it is difficult for Japanese to pronounce English with L and R properly distinguished. This is considered to be due to the fact that there is no distinction between pronunciations corresponding to L and R in Japanese, and both are pronounced as “Laryl Lero”.

  Although pronunciation can be greatly corrected by training, it cannot be properly distinguished without significant training for some pronunciations such as L and R. This burden is great for non-experts such as interpreters. From the point of view of clearly communicating intentions, there are cases in which it is possible to avoid this by not using a word that is not always easy to communicate and training to make accurate pronunciation, but simply not using that word.

In the non-patent document 1 by the present inventors, candidates for deformation are given to the input text by the deformation section, the validity of the deformation is checked by the evaluation section, and the deformation is determined to be the most appropriate one. And a system that outputs it.
It has been introduced that there are various scales such as similarity, length and frequency between paraphrased phrases as scales used in the evaluation unit.

Masaki Murata, Hitoshi Isahara "Uniform model of paraphrasing, Utilization of deformation based on scale" Proceedings of the 7th Annual Conference of the Language Processing Society of Japan 2001

For example, if the length is used as a scale of the evaluation unit and a high evaluation is given to a shorter paraphrase, the sentence can be compressed.
In addition, if the frequency is used as a scale and the words are rephrased with higher frequency, it is possible to rephrase the difficult sentences as simple words.

  The inventors of the present invention have suggested a scale that does not use words that are difficult to pronounce in the above-mentioned document, but no specific research has been conducted on what scale it is. Therefore, there was no technical consideration on whether words that would be difficult to pronounce at the time of the document should be removed from the original text, or in other words, how they would be rephrased. I stayed in things.

The above technique is also disclosed in Patent Document 1 below. This disclosure provides a system that can easily convert a sentence or sentence that requires multiple types of paraphrasing into a target sentence or sentence.
Specifically, when the transformation processing unit inputs a conversion target sentence, a number of transformation candidates are generated using the transformation rules in the transformation rule storage unit. The evaluation processing unit evaluates the generated conversion candidates using a plurality of evaluation scales for evaluating whether or not the result of transforming the character string is a target appropriate conversion, and Select a string. The highly evaluated character string is output as a conversion result sentence.
The evaluation scale can be selected by the evaluation scale selection section, and the importance of the selected evaluation scale can be set by the evaluation importance setting section.

Japanese Patent Laid-Open No. 2003-76687

As a disclosure by a person other than the present applicant, the following Patent Document 2 can be cited. The disclosed technique is a technique for providing a natural language processing method capable of converting an expression that is difficult to hear regardless of a word or a compound word into an expression suitable for reading by speech.
More specifically, the pronunciation pattern extraction unit of the text conversion unit searches for a corresponding pronunciation pattern in the pronunciation rule table, and extracts a portion that is estimated to be difficult to hear when read out. Next, the text conversion processing unit uses the text conversion rule table to convert the extracted portion into text premised on reading. He searches for expressions that are difficult to hear from a new point of view, which is a combination of sounds when reading a document aloud, so that even compound word expressions can be replaced with appropriate expressions.

JP 2000-172289 A

However, Non-Patent Document 1 and Patent Document 1 disclose general paraphrasing techniques, and it cannot be said that a technique for easily replacing words that are difficult to pronounce is provided. That is, since what kind of scale is used and what kind of evaluation is performed is not disclosed at all, even if it can be simply replaced with a word that does not include L and R, the effectiveness and accuracy of the replacement can be determined. Cannot be evaluated.
In such a conventional method, there is no difference from performing mechanical term replacement work, and there is a problem that speech support by suitable paraphrasing cannot be realized.

  In addition, Patent Document 2 discloses a technique related to speech reading, but discloses a technique that makes it easy to hear mainly by inserting a particle or a pose at the time of replacement, and it is difficult to accurately reproduce speech itself. It is not applicable to any case. Of course, even if it is applied to a foreign language, a great effect cannot be obtained.

Like the prior art described above, there is no technology for rephrasing words that are difficult to pronounce when speaking in a foreign language, as in the prior art described above. A system that can be presented to the user could not be realized.
An object of the present invention is to provide an utterance support method and apparatus for expressing a phrase difficult to pronounce for a speaker by a suitable paraphrase and for promoting communication with the audience.

In order to solve the above problems, the present invention provides the following speech support apparatus.
That is, the invention according to claim 1 provides an utterance support device that outputs dictation text that is easy to utter by replacing difficult-to-pronunciate words in a manuscript text.
The apparatus includes input means for acquiring document text data in the apparatus, and word string extraction means for extracting words or word strings (hereinafter referred to as word strings) from the document text data.
Further, a synonym database having synonym phrases for the word string is stored in a storage medium, and the word string extracted by the word string extracting unit is compared with the synonym database to search for a replacement candidate word string that can be replaced. It has a replacement candidate word string search means.
In addition, a pronunciation difficulty level database in which the difficulty level of pronunciation of each word string is quantitatively recorded in advance is stored in a storage medium, and the word string before replacement and the replacement candidate word string are respectively pronounced from the pronunciation difficulty level database. It includes a word string replacement unit that acquires a difficulty level, selects and replaces a word string having the smallest pronunciation difficulty level, and an output unit that outputs dictation text in which the word string is replaced by the above unit.

  According to the second aspect of the present invention, the pronunciation difficulty level database defines difficulty levels according to the number of predetermined characters or phonetic symbols in the word string, and the word string replacement means extracts The number of the predetermined characters or phonetic symbols in the word sequence is counted, and the pronunciation difficulty level of the word sequence is calculated based on a calculation formula of the pronunciation difficulty level set in the pronunciation difficulty database .

  According to a third aspect of the present invention, the utterance support device includes a corpus database in the same language as the original text, and the word string replacement means includes a replacement candidate word string and k-grams before and after the replacement candidate word string. A frequency counting unit that counts the frequency of occurrence of sequences in the corpus database, and when the frequency is equal to or greater than a predetermined value, the pronunciation difficulty of the replacement candidate word string is minimized. In addition, it is possible to provide a configuration including a pronunciation difficulty comparison / replacement unit that replaces the replacement candidate word string with the highest frequency.

According to the fourth aspect of the present invention, the following speech support method can be provided. The method may be provided by a program executed on a computer.
That is, an utterance support method for outputting dictation text that is easy to utter by replacing difficult-to-pronunciate words in the manuscript text, wherein the input means acquires the manuscript text data in the apparatus, and the word string extraction means And a word string extracting step of extracting a word or a word string (hereinafter referred to as a word string) from the document text data.
Then, the replacement candidate word string search means using the synonym database provided with the synonym phrase for the word string collates the word string extracted in the word string extraction step with the synonym database and can be replaced. A replacement candidate word string search step for searching for a string, a word string replacement means using a pronunciation difficulty level database in which the difficulty level of pronunciation of each word string is quantitatively recorded in advance, and the word string before replacement and the replacement candidate word string The word string replacement step for obtaining each of the pronunciation difficulty levels from the pronunciation difficulty level database, selecting and replacing the word string having the minimum pronunciation difficulty level, and the output means, the word strings are replaced by the above steps There is provided an utterance support method including at least an output step of outputting dictation text.

  In the invention described in claim 5, the pronunciation difficulty level database defines difficulty levels according to the number of predetermined characters or phonetic symbols in the word string, and is extracted in the word string replacement step. Counting the number of the predetermined characters or phonetic symbols in the word string, and calculating the pronunciation difficulty of the word string based on the calculation formula of the pronunciation difficulty defined in the pronunciation difficulty database It is.

  According to the sixth aspect of the present invention, in the word string replacement step of the speech support method, a corpus database in the same language as the manuscript text is used, and the word string replacement means includes the replacement candidate word string and k-gram (k-gram (k Is an arbitrary or the same number of word strings before and after) and counts the frequency of occurrence in the corpus database, and when the frequency is equal to or higher than a predetermined value, the pronunciation difficulty of the replacement candidate word string is minimum In addition, there is provided a technique for replacing with a replacement candidate word string having the highest frequency.

The invention described in claim 7 proposes to add the above technique as a function on a word processor program.
That is, it provides an utterance support program that is used together with a word processor program and outputs dictation text that is easy to utter by replacing difficult-to-pronunciation words in a manuscript text.
Then, a document text data reading step for obtaining the edited document text data in the word processor processing means, and a word string extraction in which the word string extraction means extracts a word or a word string (hereinafter referred to as a word string) from the document text data. A step of extracting a word string from a document being created;

Further, the replacement candidate word string search means using the synonym database provided with the synonym phrase for the word string collates the word string extracted in the word string extraction step with the synonym database, and can be replaced. A replacement candidate word string search step for searching a string, a word string selection means using a pronunciation difficulty database that quantitatively records the difficulty of pronunciation of each word string in advance, the word string before replacement and the replacement candidate word string And a word string selection step of acquiring each pronunciation difficulty level from the pronunciation difficulty level database and selecting a word string having the minimum pronunciation difficulty level.
Then, the replacement word string presenting means presents the replacement word string selected in the word string selection step together with the original text data edited in the word processor processing means, and prompts the user to input whether or not to replace. The column presentation step and the word string replacement means provide an utterance support program used together with a word processor program including a word string replacement step for replacing a word string in accordance with a user input.

  In the invention according to claim 8, the pronunciation difficulty database is determined in accordance with the number of predetermined characters or phonetic symbols in the word string, and is extracted in the replacement word string presentation step The number of the predetermined characters or phonetic symbols in the word string is counted, and the pronunciation difficulty level of the word string is calculated based on the calculation formula of the pronunciation difficulty level defined in the pronunciation difficulty level database.

  According to the ninth aspect of the present invention, in the replacement word string presentation step of the speech support program, a corpus database in the same language as the manuscript text is used, and the replacement word string presentation means includes a replacement candidate word string and k-gram ( k is an arbitrary number of the same or different word sequences before and after), and the frequency of occurrence in the corpus database is counted, and when the frequency exceeds a predetermined value, the pronunciation difficulty level of the replacement candidate word sequence is A replacement candidate word string that is the smallest and the most frequent is presented.

The following effects are produced by the above invention. In other words, many substitution candidates are obtained from the synonym database, and the importance of substitution is numerically evaluated by referring to the pronunciation difficulty database that quantitatively determines pronunciation difficulty, and the most appropriate substitution candidate Can be selected.
In particular, it is possible to prevent unnatural substitution by checking whether or not an arrangement with preceding and following word strings appears in the corpus.

  In addition, when used in combination with a word processor program as in the inventions described in claims 7 to 9, a paraphrase candidate for a word string that is difficult to pronounce at any time while creating a dictation manuscript is presented. Since it can be provided, it contributes to efficient speech support.

Hereinafter, embodiments of the present invention will be described with reference to the drawings. In addition, in implementation of this invention, it is not limited to the following, A deformation | transformation, an application, etc. can be performed arbitrarily.
FIG. 1 is a block diagram of an utterance support apparatus (1) according to the present invention. The apparatus (1) is preferably constituted by a known personal computer or the like. These include a CPU (10) that controls arithmetic processing and text processing, a hard disk (11) that stores data being processed and a database used for processing, a display that displays data, and a memory that cooperates with the CPU (10). (13) etc. are provided.

  First, the document text (20) is input to the apparatus (1). The manuscript text is a manuscript for reading in a presentation or the like, and is created based on a document or a paper. Although these manuscript texts are grammatically correct and the usage of phrases is not wrong, they were created without considering the ease of pronunciation.

Then, the document text acquisition unit (21) inputs the text data into the CPU (10). This processing is usually a general processing such as data reading from the hard disk (11).
The input document text is extracted by the next word string extraction unit (22).
In this embodiment, an English presentation manuscript is taken as an example of manuscript text, and will be described below.

English, French, Korean, and the like are generally described in a way that leaves a space for each word, which is called a split-word. In many cases, the actual pronunciation is a pronunciation that is fused between adjacent words, but in principle, the pronunciation is interrupted for each word separated by spaces.
Therefore, when focusing on pronunciation as in the present invention, it is simple and reasonable to extract words separated by spaces.
The word string extraction unit (22) extracts words from the text data while referring to the blank code. At that time, the hard disk may be provided with dictionary data or the like, or a word string such as an idiom having one meaning among a plurality of words may be extracted using information included in a synonym database or a corpus to be described later.

However, the present invention is not limited to utterance support such as English, and may target Japanese. In that case, a word string may be extracted from Japanese text data using a known morphological analysis process.
Various known techniques can be used for the morpheme analysis processing. For example, morpheme analysis can be performed using JUMAN, which is a morpheme analysis program, and decomposed into morpheme strings. Note that grammatically strict morphological analysis is not particularly necessary in the practice of the present invention. Since the main purpose is to divide into units of pronunciation, a method of simply referring to dictionary data and extracting each word having a large number of characters may be used.

For each word string extracted by the word string extraction unit (22), it is necessary to perform paraphrasing, and if necessary, what words are to be paraphrased are determined by the following processing.
That is, the replacement candidate word string search unit (23) refers to the synonym database (24) stored in the hard disk (11) and searches whether there is a synonym (substitution candidate word string) in the extracted word string. To do.

As the synonym database (24), WordNet 2.0 (Non-patent Document 2), which is an English synonym database, was used in this example.
http://wordnet.princeton.edu/, Princeton University, Wordnet 2.0. 2003

  In the case of English, verb inflection forms and plural forms of nouns need to be extracted as synonym expressions, so these information are also added to the synonym database.

As a result of the search, the word string in which the replacement candidate word string exists is output to the next word string replacement unit (25) together with the candidate, and the other word strings are output from the replacement candidate word string search unit (23) as they are.
In the word string replacement unit (25), it is determined by calculation with reference to the pronunciation difficulty database (26) whether the replacement candidate word string should be paraphrased.

The pronunciation difficulty database (26) includes a data table as shown in FIG. That is, this database is a quantitative setting of the difficulty level of pronunciation, and gives a difficulty level to the characteristics of the word string.
The difficulty level may be set individually for each word and the difficulty level 13 of the word “right” as in the column (30). However, since extracting all difficult-to-pronounced words requires a huge amount of work, it is desirable that the word string replacement unit (25) calculates from the characteristics of the words as in columns (31) to (33). .

That is, the word string replacement unit (25) counts predetermined characters in the extracted word string, the number of phonetic symbols, and the like, and based on the calculation formula stored in the pronunciation difficulty database (26), Calculate the pronunciation difficulty.
Although the calculation formula can be arbitrarily set, in this embodiment, a simple method of multiplying the number of characters / phonetic symbols by a weighting value is used.

More specifically, the first level (31) focuses on how many L or R characters are included in the word string, and the degree of difficulty for the word is defined as the number × 1. For example, if it is left, the difficulty level 1 is calculated.
Next, in the second stage (32), attention is focused on the number of L or R included in the first accent sound in the word string. In general, the pronunciation of the first accent is important, and the person who hears it as a native language often judges the word based on this sound and the overall prosody. Therefore, the degree of difficulty when L or R is included in the first accent is defined as the number × 5. If the method of adding to the above difficulty is taken, the difficulty of left is 1 + 5 = 6.

The definition of the third level (33) focuses on whether or not another significant word string is formed when L and R in the word string are exchanged. For example, in the word “right”, even if R and L are interchanged, it becomes light.
Therefore, this number × 7 is defined as the difficulty level. According to the above, the right difficulty level is defined as 1 + 5 + 7 = 13 because it matches each condition one by one.

Since the structure of the pronunciation difficulty database (26) is defined by the relationship between the native language and the foreign language, it is desirable to prepare according to the combination, such as a database for Japanese utterances for Chinese. . As described above, when two pieces of information “native language” and “spoken language” are given to the device (1), a database that matches the two pieces of information is selected and used from a plurality of pronunciation difficulty databases. An apparatus configuration corresponding to multiple languages may be provided.
In addition, the speaker may be configured to manually add and rephrase a word string that is particularly difficult to pronounce.

In the word string replacement unit (25), the word string extracted from the original sentence as described above is compared with the pronunciation difficulty level of the replacement candidate word string, and the word string having the smallest difficulty level is selected. If one of the replacement candidate word strings is minimum, the extracted word string is replaced and output as dictation text (27). The output may be the display 12 or the result may be stored on the hard disk (11).
If a word string that does not become a substitution candidate in the synonym database (such as when L and R increase instead of substitution) is removed so as not to be included in advance, the substitution candidate word string is always an extracted word string Since the pronunciation difficulty level is smaller than that, the pronunciation difficulty levels may be compared only between the replacement candidate word strings.

Although the minimum configuration of the present invention is as described above, a technique for improving the accuracy of paraphrasing as the second embodiment will be described below.
FIG. 3 is an explanatory diagram for explaining an additional embodiment of the word string replacement unit (25). The configuration other than the word string replacement unit (25) is the same as described above.
In the present embodiment, it is possible to determine whether or not the sentence is a natural sentence even if the word string is replaced by paraphrase.

A flow as shown in FIG. 4 is used for this processing.
First, when a replacement candidate word string is extracted (51) by the replacement candidate word string search unit (23) with respect to the extracted word string (50), k-gram word strings before and after the original document text data are extracted. Is extracted (52). The front and rear k-gram extraction unit (40) acts on this.
Note that morphemes are extracted in principle, but in this embodiment, words separated by spaces may be used.
k is a constant and can be arbitrarily set independently before and after, but in this embodiment, it is 2-gram (k = 2) both before and after.

Further, the word sequence generation unit (41) generates (53) an array in which the preceding clause k-gram, the replacement candidate word string, and the subsequent clause k-gram are arranged side by side.
As a result, an array in which the front and rear k-grams are the same and only the replacement target word string is replaced is generated.

For the generated sequence, the appearance frequency fb2 of the same sequence is counted (54) from the corpus (43) by the appearance frequency counting unit (42).
As is well known, a corpus is a sentence with part-of-speech information, syntax information, and the like attached thereto. If it is only for the purpose of counting the appearance frequency, not only the corpus but also a text database having a certain amount of sentences may be used.

  In order to determine whether the sentence is a natural sentence, it is confirmed (55) whether the counted appearance frequency fb2 is 1 or more. If there is none, it is determined that the wording is unnatural and the process proceeds to the next word string.

In the case of 1 or more, the pronunciation difficulty level calculation / comparison unit (44) calculates the pronunciation difficulty level. For this purpose, the same method as described above is used, but in this embodiment, “number of R or L that is followed by a vowel” is used as information as one of the information provided in the pronunciation difficulty database (26). English vowels include a, i, u, e, o, and y.
According to this, the word string from which “R + vowel” and “L + vowel” are extracted and the frequency fb1 included in the replacement candidate word string are obtained (56), and the replacement candidate word string having the smallest value may be selected. In other words, this is nothing other than selecting a word string having the smallest pronunciation difficulty.

If there are a plurality of word strings having the minimum fb1, the one having the maximum appearance frequency fb2 in the corpus may be selected (57).
According to this selection method, it is possible to replace as much as possible a word string including an element having a high pronunciation difficulty such as “R + vowel”, and at the same time, it is possible to extract a sequence having a large appearance frequency in a correct foreign language sentence. Contributes to natural wording.

  When the word string to be replaced is determined as described above, the processing proceeds to the next extracted word string. After all word strings extracted from the manuscript text data are processed, the dictation text (27) is output.

In the present embodiment, the case where k = 2 is fixed has been described. However, by dynamically changing k, it is also possible to extract a combination with more meaningful front and rear clauses and compare appearance frequencies. That is, each phrase is extracted from the target word string until one content word (significant word that characterizes the sentence) appears toward the previous section or the subsequent section.
This is because it is difficult to determine that the paraphrase is correct even if a paraphrased expression appears in the corpus, etc., from the context of only the functional word. This is because it can be done correctly.

Specifically, a functional word dictionary is provided in a hard disk or the like, and word k-grams until m words (content words) that do not match the functional word dictionary appear (m is an arbitrary number) are extracted. At this time, m may be defined as 1, for example, and the upper limit of k may be defined as 7, for example, and may be dynamically changed between 1 and 7.
According to this configuration, in verifying whether the paraphrase is a natural sentence, the content word can be used effectively as an expression around the paraphrased expression, which can contribute to reliable paraphrase.

In the above description, specific characters such as L and R and their positions are used as conditions. However, since the present invention is intended for pronunciation, substitution may be performed based on pronunciation symbols.
As shown in Table 1, vowels and consonants are defined as phonetic symbols, and further accents are displayed. These symbols can be easily replaced from words to phonetic symbols by dictionary data, and the original text can be once converted into phonetic symbol notation by a phonetic symbol converter (not shown) of the CPU (10).

Then, with reference to the phonetic symbols, the number of [r] [l] or the number of vowels following [r] can be counted. When using phonetic symbols in this way, words that are difficult to pronounce or difficult to distinguish with only a character string can be clearly extracted as difficult to pronounce.
Even when phonetic symbols are used, the pronunciation difficulty level database may be provided with information in which the difficulty levels for the phonetic symbols are set.

As another embodiment 3, the present invention can be provided in a configuration in which the technology of the present invention is mounted on a word processor program (called word processor software).
FIG. 5 is a block diagram thereof. The CPU (60) has a word processor processing unit (61) which is a processing unit for operating a word processor. Here, a well-known word processor software operates, and text can be edited, read and written.

Then, the following processing unit according to the present invention performs an utterance support operation on the manuscript text (62) being edited on the word processing software. That is, the reading unit (63) reads the original text, and the word string extracting unit (64) extracts the word string. The operation is the same as above.
For each word string, the replacement candidate word string search unit (65) searches the replacement candidate word string using the synonym database (66).

  At this time, the entire text of the manuscript text does not necessarily have to be completed. For example, each process may be performed when one sentence is input (a period is entered).

Then, the word string selection unit (67) selects an optimum word string for replacement while referring to the pronunciation difficulty database (68). In this embodiment, after selection, the replacement word string presentation unit (69) presents a word string that recommends replacement to the user.
For the presentation, it is easy to use the display (70). For example, a word string that recommends replacement is displayed on the word string on the word processor software. Simultaneously with the presentation, a display prompting the user whether or not to replace is performed.

When the user consents with the keyboard (71) or the like, the word string replacement unit (72) replaces the word string with a replacement word string on the word processing software.
According to the above method, a user can present words that are difficult to pronounce while typing text for dictation on word processing software, and can perform paraphrasing only by an acceptance operation.
When implementing the system of the present invention, it is preferable to use a synonym database corpus having a large amount of information. However, in the system of this embodiment, the amount of information is presented to the user, and the user judges each time. If there is, it is sufficiently usable.

FIG. 6 shows the result of the above apparatus.
It can be seen that the words “approach” can be rewritten to “way”, and “length” can be rewritten to “size” to make it easier to pronounce. On the other hand, there are cases in which subtle differences in nuances cannot be rephrased in the current system. This can be achieved by improving the accuracy by reviewing the contents of the synonym database.
Further, as in the third embodiment, in other embodiments, candidates may be presented to the user and it may be determined whether or not the user is to paraphrase.

The present invention can also provide an audio reproducing apparatus using the above technique.
2. Description of the Related Art Conventionally, a voice synthesis processing technique is well known, and a technique for generating a voice waveform from text data by voice synthesis means and reproducing voice from a speaker or the like is known. At this time, it may be difficult to completely reproduce a part of the pronunciation and it may be difficult to hear. The above-mentioned Patent Document 2 aims to solve this problem, and attempts to solve the problem by supplementing particles and poses or by paraphrasing them mechanically.

However, in this method, only uniform paraphrasing is possible, and it is not possible to flexibly deal with ambiguous rules peculiar to languages. In particular, since all words that are assumed to be difficult to hear must be registered in advance, new words cannot be handled even if they are difficult to hear.
On the other hand, the technique of the present invention can deal with new words and phrases while determining whether or not pronunciation is difficult (difficult to hear) based on the rules as described above.

FIG. 7 is a block diagram of an audio playback device (80) according to the present invention.
Document text data (81) is input to the CPU (82) by text data input means (83). The text data input unit (83) is a collective term for various means for inputting text to the CPU. For example, a direct input means such as a keyboard, a reading means such as a magnetic disk, an information acquisition means from a network, etc. Including.

Then, the word string extraction unit (84) extracts the word string in the same manner as described above, and the replacement candidate word string search unit (85) searches the synonym database (86) for the replacement candidate word string.
Further, the speech reproduction difficulty level is calculated by the word string replacement unit (88) with reference to the speech reproduction difficulty level database (87) between the extracted word string and the replacement candidate word string.

The degree of difficulty in reproducing speech is the same as the degree of difficulty in pronunciation, but the contents included in the database are information related to sounds that are difficult to synthesize accurately when the next speech synthesizer (89) performs speech synthesis. It is.
That is, in Japanese, a specific sound in 50 sounds or a prompt sound is defined. For example, it is assumed that the sound synthesis unit (89) is not good at the prompt sound and the sound reproduction difficulty level is defined as 5 for the prompt sound “tsu”. It is assumed that the sound reproduction difficulty level is defined as 1 for all muddy sounds as a slightly unsatisfactory case.
At this time, “immediately” is extracted from the synonym database (86) for the phrase “soon”. Since the difficulty level of “immediately” is calculated as 1 including one muddy sound, “soon” with a difficulty level of 5 is replaced with “immediately”.

The voice synthesizer (89) generates voice waveform information from the text data by a known voice synthesis technique. As a method for outputting an arbitrary character string, a rule synthesis / text-to-speech synthesis method is known.
As an example, the case where the manuscript text data is Japanese text will be described. In the word string extraction unit (84), a word string is extracted by collation with a known morphological analysis technique or a dictionary and further stored in the hard disk (91). The word reading and accent (prosodic information) are defined based on the phonetic character correspondence database.

As a result, the input Japanese text is converted into a phonetic character string. The word string replacement unit (88) can perform the same calculation as in the above-described embodiment, such as weighting in the case of the first accent, with reference to the accent defined here.
Then, the phonetic character string is converted into a voice waveform in the voice synthesizer (89). The waveform of each phoneme is read from the waveform dictionary database in the hard disk (91), and a speech waveform is generated.

  The waveform of each phoneme varies significantly depending on what phoneme comes on both sides of the phoneme (phoneme context). Therefore, in general, different waveforms are prepared for the same phoneme but different phoneme contexts. How finely deal with phoneme context is greatly related to the clarity and smoothness of synthesized speech. The intelligibility is a sound listening rate (recognition rate) by a human and is a concept opposite to the speech reproduction difficulty level database in this embodiment.

In the speech synthesis technology, research corresponding to such phoneme contexts by complicated processing has been advanced. However, in order to simplify the processing in the speech synthesis unit (89), the present invention replaces word strings in advance. To do.
Accordingly, the speech synthesizing unit (89) is already provided with text that increases the clarity, and the generated speech waveform is input to the speech output unit (90).

  The sound output unit (90) is composed of a known sound card or the like, and generates and outputs sound (analog data) from the sound waveform. A speaker (92) connected to the CPU (80) is used for output.

Of course, in the present apparatus, when reproducing a foreign language such as English, it is possible to avoid and output a sound that is difficult for the audience to distinguish as in the above embodiment. For example, by reducing the number of words including L and R, it is possible to reduce the use of words that are difficult for Japanese to hear.
More generally, the phonetic symbols can be compared to rephrase words that contain many phonetic symbols with similar pronunciation. For example, phonetic symbols having similar pronunciations are prepared in advance in the speech reproduction difficulty level database, and the difficulty level is set in advance, so that a word string including many confusing pronunciations is rephrased. As a result, it is possible to provide an audio reproduction device that is easy to listen to a listener who is not fluent in foreign languages.

According to the above voice reproducing apparatus, the processing in the voice synthesizing unit (89) can be reduced, and at the same time, it is not necessary to prepare various waveforms for the phoneme context. In particular, when a plurality of languages are reproduced by the speech synthesizer, having all the speech contexts and waveforms in each language in advance has a problem that the amount of information becomes enormous.
However, in the present invention, if information on a sound that is poorly pronounced in the speech synthesizer (89) is given (preferably a phonetic symbol), word strings are replaced from the synonym database of each language even in a plurality of languages. By doing so, it is possible to realize voice reproduction with high clarity.

It is a block diagram of the speech support apparatus which concerns on the 1st Embodiment of this invention. It is an example of a pronunciation difficulty database. It is a block diagram of the word string replacement part which concerns on the 2nd Embodiment of this invention. It is a flowchart. It is a block diagram concerning the 3rd embodiment of the present invention. It is a figure which shows the support result by this invention. It is a block diagram of the audio | voice reproduction apparatus which concerns on the 3rd Embodiment of this invention.

Explanation of symbols

1 Utterance support device 10 CPU
DESCRIPTION OF SYMBOLS 11 Hard disk 12 Display 13 Memory 20 Manuscript text 21 Manuscript text acquisition part 22 Word string extraction part 23 Replacement candidate word string search part 24 Synonym database 25 Word string substitution part 26 Pronunciation difficulty database 27 Oral text

Claims (16)

A voice playback device that generates voice by voice synthesis processing when a manuscript text is input,
Input means for acquiring document text data in the apparatus;
Word string extraction means for extracting words or word strings (hereinafter referred to as word strings) from the manuscript text data;
A synonym database with synonym phrases for word sequences;
A replacement candidate word string search means for collating the word string extracted by the word string extraction means with the synonym database and searching for a replaceable replacement candidate word string;
A speech reproduction difficulty database that quantitatively records the difficulty of speech reproduction in speech synthesis processing;
Word string replacement means for acquiring the respective speech reproduction difficulty levels from the speech reproduction difficulty database for the word string before replacement and the replacement candidate word strings, and selecting and replacing the word string having the smallest voice reproduction difficulty level When,
Speech synthesis means for performing speech synthesis processing based on the dictation text with the word string replaced by the above means;
An audio reproduction apparatus comprising: audio output means for outputting audio. The speech reproduction difficulty level database defines a difficulty level according to the number of predetermined characters or phonetic symbols in a word string,
The word string replacement means counts the number of the predetermined characters or phonetic symbols in the extracted word string, and the speech of the word string is calculated based on the speech reproduction difficulty calculation formula defined in the speech reproduction difficulty database. The audio reproduction device according to claim 1, wherein the reproduction difficulty level is calculated. In the speech reproduction difficulty level database, regarding two or more characters or phonetic symbols that are difficult to distinguish, any number of predetermined characters or phonetic symbols included in a word string, or any predetermined character or phonetic symbol The number constituting the sound of the first accent in the word string, the number constituting a significant word string when any given character or phonetic symbol in the word string is replaced with the other character or phonetic symbol, Comprising at least one piece of information and defining a predetermined weighting value for these numbers,
The audio reproduction device according to claim 1, wherein the word string replacement unit performs a calculation of multiplying each constituent number by a weighting value and calculating a sum of the multiplications. The audio reproduction device includes a corpus database in the same language as the original text,
The word string replacement means is
A frequency counting unit that counts the frequency at which the sequence of the replacement candidate word strings and the word strings of k-grams before and after the k-grams (k is the same or different before and after) appear in the corpus database;
A speech reproduction difficulty comparison / replacement unit that replaces the replacement candidate word string with a replacement candidate word string that has the lowest frequency and the highest frequency when the frequency is equal to or higher than a predetermined value. The audio reproduction device according to any one of claims 1 to 3. An utterance support device that outputs easy-to-speak dictation text by replacing difficult-to-pronunciation words in manuscript text,
Input means for acquiring document text data in the apparatus;
Word string extraction means for extracting words or word strings (hereinafter referred to as word strings) from the manuscript text data;
A synonym database with synonym phrases for word sequences;
A replacement candidate word string search means for collating the word string extracted by the word string extraction means with the synonym database and searching for a replaceable replacement candidate word string;
A pronunciation difficulty database that quantitatively records the difficulty of pronunciation of each word string in advance;
Word string replacement means for obtaining the pronunciation difficulty level from the pronunciation difficulty database for the word string before replacement and the replacement candidate word string, and selecting and replacing the word string having the minimum pronunciation difficulty;
An utterance support apparatus comprising: output means for outputting dictation text in which a word string is replaced by the means described above. The pronunciation difficulty database defines the difficulty according to the number of predetermined characters or phonetic symbols in the word string,
The word string replacement means counts the number of the predetermined characters or phonetic symbols in the extracted word string, and based on the pronunciation difficulty calculation formula defined in the pronunciation difficulty database, the pronunciation difficulty of the word string The utterance support device according to claim 5. In the pronunciation difficulty level database, regarding two or more difficult-to-distinguish characters or phonetic symbols, the number of any given character or phonetic symbol included in the word string, or any given character or phonetic symbol is At least a number constituting the sound of the first accent in the word string, a number constituting a significant word string when any given character or phonetic symbol in the word string is replaced with the other character or phonetic symbol, It is configured to include any information and to define a predetermined weighting value for the number,
The utterance support device according to claim 5 or 6, wherein the word string replacement unit performs a calculation of multiplying each constituent number by a weighting value and calculating a sum of the multiplications. The utterance support device comprises a corpus database in the same language as the manuscript text;
The word string replacement means is
A frequency counting unit that counts the frequency at which the sequence of the replacement candidate word strings and the word strings of k-grams before and after the k-grams (k is the same or different before and after) appear in the corpus database;
A pronunciation difficulty comparison / replacement unit that replaces the substitution candidate word string with a substitution candidate word string that has the smallest and most frequent pronunciation difficulty when the frequency is equal to or higher than a predetermined value. The utterance support device according to claim 5. An utterance support method for outputting dictation text that is easy to speak by replacing difficult-to-pronunciation words in the manuscript text,
An input step in which the input means acquires document text data in the apparatus;
A word string extracting step in which a word string extracting unit extracts a word or a word string (hereinafter referred to as a word string) from the document text data;
The replacement candidate word string search means using the synonym database having synonym phrases for the word string collates the word string extracted in the word string extraction step with the synonym database, and determines a replacement candidate word string that can be replaced. A replacement candidate word string search step to search,
Using the pronunciation difficulty database that quantitatively records the difficulty of pronunciation of each word string in advance, the word string replacement means for each word pronunciation from the pronunciation difficulty database for the word string before the replacement and the replacement candidate word string A word string replacement step of acquiring a difficulty level and selecting and replacing a word string having the minimum pronunciation difficulty level;
The speech support method, wherein the output means includes at least an output step of outputting the dictation text in which the word string is replaced by each of the above steps. The pronunciation difficulty database defines the difficulty according to the number of predetermined characters or phonetic symbols in the word string,
In the word string replacement step, the number of the predetermined characters or phonetic symbols in the extracted word string is counted, and the pronunciation difficulty level of the word string is calculated based on the calculation formula of the pronunciation difficulty level defined in the pronunciation difficulty database The utterance support method according to claim 9. In the pronunciation difficulty level database, regarding two or more difficult-to-distinguish characters or phonetic symbols, the number of any given character or phonetic symbol included in the word string, or any given character or phonetic symbol is At least a number constituting the sound of the first accent in the word string, a number constituting a significant word string when any given character or phonetic symbol in the word string is replaced with the other character or phonetic symbol, It is configured to include any information and to define a predetermined weighting value for the number,
The utterance support method according to claim 9 or 10, wherein, in the word string replacement step, calculation is performed by multiplying each constituent number by a weighting value and calculating a total sum thereof. In the word string replacement step of the utterance support method,
Using a corpus database in the same language as the original text, the word string replacement means
Counting the frequency of occurrence of the replacement candidate word sequence and the sequence of k-grams before and after the word sequence in the corpus database (where k is the same number or different before and after)
The utterance support method according to any one of claims 9 to 11, wherein when the frequency is equal to or higher than a predetermined value, the replacement candidate word string is replaced with a replacement candidate word string having the lowest pronunciation difficulty and the highest frequency. An utterance support program that is used together with a word processor program and outputs dictation text that is easy to utter by replacing difficult-to-pronunciation words in a manuscript text,
An original text data reading step for obtaining original text data in an edited state in a word processor processing means;
A word string extracting step in which a word string extracting unit extracts a word or a word string (hereinafter referred to as a word string) from the document text data;
The replacement candidate word string search means using the synonym database having synonym phrases for the word string collates the word string extracted in the word string extraction step with the synonym database, and determines a replacement candidate word string that can be replaced. A replacement candidate word string search step to search,
The word string selection means uses the pronunciation difficulty database that quantitatively records the difficulty of pronunciation of each word string in advance, and the word string selection means for each word string from the pronunciation difficulty database for the word string before replacement and the replacement candidate word string. A word string selection step of acquiring a difficulty level and selecting a word string having a minimum pronunciation difficulty level;
The replacement word string presenting means presents the replacement word string selected in the word string selection step together with the original text data edited in the word processor processing means, and prompts the user to input whether or not to replace,
An utterance support program used together with a word processor program, wherein the word string replacement means includes at least a word string replacement step of replacing a word string in response to a user input. The pronunciation difficulty database defines the difficulty according to the number of predetermined characters or phonetic symbols in the word string,
In the replacement word string presenting step, the number of the predetermined characters or phonetic symbols in the extracted word string is counted, and the pronunciation difficulty of the word string is calculated based on the calculation formula of the pronunciation difficulty defined in the pronunciation difficulty database An utterance support program used together with the word processor program according to claim 13. In the pronunciation difficulty level database, regarding two or more difficult-to-distinguish characters or phonetic symbols, the number of any given character or phonetic symbol included in the word string, or any given character or phonetic symbol is At least a number constituting the sound of the first accent in the word string, a number constituting a significant word string when any given character or phonetic symbol in the word string is replaced with the other character or phonetic symbol, It is configured to include any information and to define a predetermined weighting value for the number,
The utterance support program used together with the word processor program according to claim 13 or 14, wherein, in the replacement word string presenting step, a calculation is performed to multiply each number by a weighting value and to obtain a sum thereof. In the replacement word string presentation step of the utterance support program,
Using a corpus database in the same language as the manuscript text, the replacement word string presenting means,
Counting the frequency of occurrence of the replacement candidate word sequence and the sequence of k-grams before and after the word sequence in the corpus database (where k is the same number or different before and after)
16. The word processor program according to claim 13, wherein when the frequency is equal to or higher than a predetermined value, the substitution candidate word string having the smallest pronunciation difficulty and the highest frequency is presented. Utterance support program.

Images