JPS62165267A - Voice word processor device - Google Patents

Abstract

PURPOSE:To simplify a statistical table, and to improve a processing speed by putting together into one generating the statistical table which checks a connection possibility between KANA (Japanese syllabary) character strings in Japanese from the appearance frequency of KANA characters only, and a bit of connection possibility information between the KANA character strings. CONSTITUTION:The connection possibility between each KANA character string is checked at a connection evaluation part 31 based on a candidate obtained from a speech recognition part 2 referring a statistical table 32, and the KANA character string which is not connected to others is eliminated. Furthermore, a morphemic analy sis for the KANA character string left at the connection evaluation part 31 is performed referring a word dictionary 34, and the KANA character string is selected. At a KANA-KANJI (Chinese character) conversion part 35, a KANA-KANJI conversion is performed for the KANA character string selected at a KANA character selection part 33 referring a KANJI dictionary 36, and a KANJI and KANA mixed text is generated. Finally, a route evaluation part 37, when an effective result is not obtained until a process at the KANA-KANJI conversion part 35, returns to another arranging route, then performing a conversion again. In this way, it is possible to perform the process at a high speed, and a memory capacity to store the table can be reduced.

Description

DETAILED DESCRIPTION OF THE INVENTION Technical Field The present invention relates to an audio word processing device that creates sentences by voice input, but it is also applicable to word processor devices that create sentences by not only voice input but also optical character reading devices and ordinary keyboard input. is also applicable.

Japanese Patent Application Laid-open No. 60-39238 discloses that, as a voice input sentence creation method, for kana character string candidates obtained from a speech recognition unit, the possibility of preceding connection and the possibility of backward connection of a certain candidate character are calculated respectively. Statistics table A.

A method is described in which a sentence is created by evaluating with reference to B, performing morphological analysis on connectable candidate kana character strings, and performing kana-kanji conversion. In order to improve the accuracy and speed of statistical processing, it is possible to reduce the number of kana characters in Japanese by combining the statistical tables into one, and to use the kana character string connection frequency, which takes into account the frequency of kana characters in Japanese, as statistical information. This reduces the amount of memory required for tables and enables high-speed processing.

In addition, when checking the possibility of connection between kana character strings in the kana-kanji conversion section, a statistical table that stores the appearance frequency information of individual kana characters and the possibility of connection between two kana characters (connection frequency) is stored. There is a method of evaluating candidate kana character strings while referring to another statistical table, but these two statistical tables can be combined into one as mentioned above, and in terms of system construction, This is preferable in order to reduce the amount of memory and improve processing speed.

The present invention was made in view of the above-mentioned circumstances,
In particular, in voice word processing devices that create sentences by voice input, statistical tables are used to check the connectability of kana character strings consisting of combinations of top candidates obtained from the recognition results of the speech recognition unit. Column appearance frequency and kana character string chain information are stored together as one, and after reducing kana character string candidates by referring to the statistical table for forward and/or backward connections, select the ones with a high possibility of connection. This was done with the purpose of providing a means for simplifying statistical tables and improving processing speed by performing morphological analysis preferentially starting from kana sequences while performing kana-kanji conversion processing.

Configuration In order to achieve the above object, the present invention includes a speech recognition unit that recognizes unknown input speech, and a storage that stores the top candidate kana character strings of the recognition results output from the speech recognition unit and their similarities. A statistical table that checks connections between top kana candidate strings, a selection and retention means that selects and retains candidate strings with a high probability of connecting kana strings, and converts the selected and retained kana strings into sentences containing kanji and kana. In an audio word processing device having a kana-kanji conversion means.

The feature is that a statistical table for checking the possibility of connection between Japanese kana character strings is created from the appearance frequency of individual kana characters and connection probability information between kana character strings and compiled into one. . Hereinafter, the present invention will be explained based on examples.

FIG. 1 is an electrical block diagram for explaining one embodiment of the audio word processor device according to the present invention.
1 is a microphone for voice input, 2 is a voice recognition section, 3 is a kana-kanji conversion section, 4 is a document editing section, 5 is a control section, and 6 is a printer for document output.

7 is a recognition result display section.

The voice input through the microphone 1 is converted into feature parameters in the recognition result section 2, and then matched with a standard pattern stored in advance. Usually, multiple kana character strings are retained along with their similarity. Then, a kana character string is formed as the arrangement of the kana characters, and while checking the connection status of kana characters, morphological analysis, grammar check, etc., the kana-kanji converter 3 converts it into a sentence containing kanji and kana, which is then sent to the document editor 4. and edited into the appropriate format. The control section 5 displays the document edited by the document editing section 4 on the display section 7, and can also make a hard copy to the printer 6 when necessary.

FIG. 2 is a block diagram for explaining in more detail the configurations of the speech recognition section 2 and the kana-kanji conversion section 3 in FIG. 24 is a recognition processing unit, 24 is a standard pattern storage unit, 25 is a recognition result candidate storage unit, 31 is a kana character string connection evaluation unit, 32 is a statistical table, 33 is a kana character string selection unit, and 34 is a word dictionary storage unit.

35 is a kana-kanji conversion unit, 36 is a kanji dictionary storage unit, and 37 is a kana character string route evaluation unit.

The audio input from the microphone 1 is processed by the preprocessing unit 21.
, after passing through a low-pass filter.

After being AD converted, the characteristic analysis unit 22 converts it into voice-specific characteristic parameters at regular intervals by bandpass filtering, LPG analysis, etc. #M$ pattern storage section 24
Type patterns such as monosyllables and words are stored in advance as similar feature parameter series, and the recognition unit 23 compares them with input speech patterns. At this time, the recognition result, a plurality of candidate kana character strings, and their similarities are stored in the candidate storage unit 25. Next, based on the candidates obtained by the speech recognition unit 2, the connection evaluation unit 31 checks the possibility of connection between each kana character string while referring to a statistical table 32, which will be described later, and identifies kana character strings that do not connect. exclude. Furthermore,
While referring to the word dictionary 34, the connection evaluation section 31 morphologically analyzes the remaining kana character strings and selects the kana character strings. The kana-kanji conversion unit 35 performs kana-kanji conversion on the kana character string selected by the kana character string selection unit 33 while referring to the kanji dictionary 36 to create a sentence containing kanji and kana. Finally, if a valid result is not obtained by the kana-kanji conversion section 35, the route evaluation section 37 returns to another arrangement route and performs the conversion again.

FIG. 3 is a diagram for explaining the method of creating the statistical table 32 shown in FIG.
32 is a kana sequence appearance chain frequency table (statistical table) created by the multiplication process of 8 and 9.

The kana appearance frequency table 8 represents the frequency of kana characters appearing in Japanese sentences, and the kana string chain probability table 9 represents the probability of two kana characters being linked.

For example, in the kana string chain probability table 9 in FIG.
The probabilities that the kana characters "Aate", "Ai", and "Ark" are chained are 0.02, 0.01°, and 0.01, respectively.

The statistical table 32 is created by multiplying the values of 8 and 9. For example, the frequency at which [Ai] is chained is (frequency of appearance of 'A' 100) x (r chain probability of Ai 0.01) )X(frequency of appearance of "i" 50)=50.

Figure 4 shows another implementation example in which a statistical table is created based on the chain probability between independent words and adjunct words or words and monosyllables, and the speech input, recognition, and conversion processing is performed between words and monosyllables. This kind of statistical table can be cited when doing it for each section.

Although the statistical table shown in Figure 3 represents the probability of linkage between kana character strings, the present invention is not limited to this; It is applicable to statistical tables representing connection chain probabilities between all delimited units.

Effects As is clear from the above explanation, the statistical table that takes into account the occurrence chain of kana sequences according to the present invention combines the appearance frequency table of kana characters in Japanese and the chain probability table between kana characters into one. As mentioned above,
Since the connection evaluation between kana character strings can be performed by referring to only one table, it is effective in speeding up processing and reducing the amount of memory for storing tables.

[Brief explanation of drawings]

FIG. 1 is an electrical block diagram for explaining one embodiment of the voice word processor device according to the present invention, and FIG. 2 is a detailed electrical block diagram of the voice recognition section 2 and kana-kanji conversion section 3 shown in FIG. The circuit diagram and FIG. 3 are diagrams for explaining an example of a method for creating the statistical table 32 shown in FIG. 2, and FIG. 4 shows another embodiment in which the statistical table is created based on chain probability. It is a diagram. 1... Voice input microphone, 2... Voice recognition section, 3... Kana-Kanji conversion section, 4... Document editing section, 5
...Control unit, 6...Printer, 7...Recognition result display section, 21...Preprocessing section, 22...Feature analysis section,
23... Recognition processing section, 24... Mark pattern storage section, 25... Recognition result candidate holding section, 31... Kana character string connection evaluation section, 32... Statistical table, 33... Kana character string selection section, 34... Word dictionary storage section, 35...
・Kana-kanji conversion section. 36...Kanji dictionary storage unit, 37...Kana character string route evaluation unit. Patent Applicant: Ricoh Co., Ltd. Figure I/Figure 2 Figure 41A Figure 3

Claims (3)

[Claims] (1) A speech recognition unit that recognizes unknown input speech; a storage unit that stores the top multiple candidate kana character strings and their similarity of the recognition results output from the speech recognition unit; It has a statistical table for checking connections, a selection holding means for selecting and holding candidate character strings with a high possibility of connection of kana character strings, and a kana-kanji conversion means for converting the selected and held kana character strings into sentences containing kanji and kana. A feature of the audio word processing device is that a statistical table for checking the possibility of connection between Japanese kana character strings is created from the appearance frequency of individual kana characters and connection probability information between kana character strings and combined into one. An audio word processing device that uses (2) Create an appearance chain frequency table to determine the possibility of connection between an independent word and an attached word, between an attached word and an attached word, and between an independent word and an independent word from the frequency of each occurrence and the connection probability information between them. The audio word processor device according to claim 1, wherein the information is created and compiled into one table. (3) Create an appearance chain frequency table to determine the possibility of connection between independent words or attached words and punctuation marks in a sentence from the frequency of occurrence of independent words or attached words and punctuation marks, and connection probability information between them. The audio word processing device according to claim (1), characterized in that the information is summarized into two tables.