JP2002221989A - Text input method and apparatus - Google Patents

Abstract

(57) [Problem] To provide a text input method which can be reduced in size. SOLUTION: Input means for inputting voice, candidate display means for creating and displaying word string candidates from the input voice,
A candidate selecting means for the user to select the candidate, wherein the user sequentially selects candidates in units of one to several words in order from the beginning of the utterance. The search process that has been performed is changed to a sequential search process using a user-initiated selection / confirmation operation. As a result, a search space in which many combinations of word string candidates are considered can be significantly reduced, and the apparatus can be downsized in terms of storage capacity and processing amount.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a text input method utilizing voice recognition, and more particularly to a text input method and apparatus for a small device such as a mobile phone.

[0002]

2. Description of the Related Art Conventionally, as a text input method using voice recognition, a speaker utters a word or a phrase unit,
There is a method of recognizing a speech for each utterance, and a method of recognizing a whole sentence at once by uttering in units of sentences or sentences.
In the former case, as described in Japanese Patent Application Laid-Open No. 2-298997, a predetermined number of candidates are displayed on a menu after the speaker has uttered, and the speaker can be selected from the menu. However, in this method, the speaker has to separate utterances in units of phrases, and must select the correct word each time, so that the input operation is very troublesome and time-consuming.

On the other hand, as the latter case, for example, “Japanese large vocabulary continuous speech recognition using words as recognition units” (Information Processing Society of Japan, Vol.40, No.4, pp1395-1403, Apl. 1999)
Are known.

FIG. 12 shows an operation flow of the conventional text input method, and the operation will be described.

[0005] First, the user inputs a voice (S12).
01). Next, the device automatically searches for a recognition result.
In searching for a recognition result, the device obtains an acoustic score of the entire utterance while connecting acoustic units such as phonemes. At the same time, a linguistic score is obtained for a series of linguistic units such as words. Then, the apparatus ranks the recognition results in descending order of the score obtained by integrating them. Usually, one utterance is a sentence composed of several words or tens of words. In order to output a highly accurate recognition result, the apparatus leaves many word string candidates in consideration of combinations of word candidates during the search (S1202).

Next, the apparatus displays the highest word sequence in the order of the recognition result for all the inputted voices (S1203). Next, the user corrects a part of the displayed recognition result that is different from his / her intention (S12).
04). Then, when all the corrections by the user have been completed, the device ends the input operation for one utterance (S1205).

[0007]

However, in the above-mentioned prior art, the recognition result candidate is corrected after performing the whole sentence recognition process. Therefore, in the case of a long utterance, the load of the recognition process is large, and the storage capacity is large. Many are needed. Therefore, there is a problem that the device cannot be downsized.

An object of the present invention is to realize a text input method which can be reduced in size and can input one or more sentences of continuous voice.

[0009]

A text inputting method and apparatus according to the present invention for solving the above-mentioned problems provide a method in which a user inputs utterances input in units of sentences or sentences, and sequentially inputs candidates in units of words or phrases from the beginning of a sentence. A search process for selecting and fixing is performed.

[0010] This eliminates the need to maintain a search space in which many combinations of word string candidates are taken into account, so that the storage capacity can be significantly reduced and the amount of speech recognition processing can be reduced. This makes it possible to reduce the size of the device. Furthermore,
Since the user can input a continuous voice in units of one or more sentences, the trouble of inputting in units of words is eliminated.

[0011]

DESCRIPTION OF THE PREFERRED EMBODIMENTS In a text input method according to a first aspect of the present invention, a step of continuously inputting speech and a step of selecting a word string candidate in units of one to several words from the head of the input speech are described. A candidate creating step for creating, a display step for displaying the candidate, and a selecting step for the user to select the displayed candidate. Based on the selected candidate, the next Creation steps;
The display step and the selection step are sequentially repeated.

[0012] Thus, the user performs an operation of deciding a candidate from the beginning of the input voice, and the unit is set to several words, so that the system needs to prepare only local candidates. Text input is possible with a small amount of processing and a small storage capacity, and the size of the apparatus can be reduced.

According to a second aspect of the present invention, in the text input method according to the first aspect, in the candidate creation step, a phrase unit candidate is identified by an expansion process of repeating word connection in accordance with a word unit chain probability. To decide.

[0014] Thus, while using information that is advantageous for miniaturization, such as the chain probability in word units, it can be presented to the user in an easy-to-understand manner by converting it into phrase units.

According to a third aspect of the present invention, in the text input method according to the second aspect, the candidate creating step further includes a candidate updating process based on an acoustic score.

[0016] By assigning the ranks of the candidates using the linguistic score and the acoustic score, the candidates are presented in the order of frequency of use and candidates having the closest pronunciation. Fewer candidates need to be read.

According to a fourth aspect of the present invention, in the text input method according to the third aspect, when the number of phrase candidates subjected to the expansion processing reaches a predetermined number from the top of the language score, the expansion is performed. This ends the processing.

Thus, by limiting the number of candidates to be presented to the user, when the speech recognition processing is unsuccessful, the user is prompted again for speech input without continuously displaying incorrect candidates endlessly. Can be processed.

A text input device according to a fifth aspect of the present invention comprises: an input unit for inputting a voice; a voice preprocessing unit for extracting a feature amount of the voice from the input unit; A word candidate creating unit for creating a next word candidate using the language model; and one or several words from at least one of the language model and the acoustic model from the extracted feature amount and the word candidate. A word string creation unit for creating a string candidate, a display unit for displaying the word string candidate, and an operation unit for the user to select the displayed word string candidate,
A candidate creation instructing unit that instructs the word candidate creating unit to create the next word candidate from the word string selected by the operation unit.

Thus, the user performs the operation of deciding a candidate from the beginning of the input voice, and the unit is set to several words, so that the system needs to prepare only local candidates. Text input is possible with a small amount of processing and a small storage capacity, and the size of the apparatus can be reduced.

According to a sixth aspect of the present invention, in the text input device according to the fifth aspect, the word string creating unit performs a phrase unit candidate by performing an extension process of repeating a word connection in accordance with a word unit chain probability. Is to create.

Thus, while using information that is advantageous for miniaturization, such as the chain probability in word units, it is possible to present to the user in an easy-to-understand manner by converting into phrase units.

According to a seventh aspect of the present invention, in the text input device according to the sixth aspect, the word string creating section further has an updating process based on an acoustic score.

In this way, by assigning the order of the candidates using the language score and the acoustic score, the candidates are presented in order from the candidate having the highest use frequency and the closest pronunciation, so that the user can obtain the desired candidate before the candidate is obtained. Fewer candidates need to be read.

According to an eighth aspect of the present invention, in the text input device according to the seventh aspect, the word string creating section determines that the number of phrase candidates subjected to the decompression processing is a predetermined number from the top of the language score. Is reached, the decompression process is terminated.

Thus, by limiting the number of candidates to be presented to the user, when the speech recognition processing is unsuccessful, the user is prompted again for speech input without continuously displaying incorrect candidates endlessly. Can be processed.

[0027] A mobile phone according to a ninth aspect of the present invention includes the text input device according to any one of the fifth to eighth aspects.

[0028] This makes it possible to reduce the size of the apparatus, especially in the case of a mobile phone, since only the local candidates are required for the system to prepare.

A program according to a tenth aspect of the present invention comprises an input step of continuously inputting voices, and a candidate generating step of generating word string candidates in units of one to several words from the beginning of the input voice. And a display step of displaying the candidate, and a selecting step of the user selecting the displayed candidate. The candidate creating step and the display of the next sound based on the selected candidate are performed. The step and the selecting step are sequentially repeated.

By doing this, the user performs an operation of deciding a candidate from the beginning of the input voice, and the unit is set to several words, so that the system needs to prepare only local candidates. Text input is possible with a small amount of processing and a small storage capacity, and the size of the apparatus can be reduced.

According to an eleventh aspect of the present invention, in the program according to the tenth aspect, in the candidate creating step, a phrase unit candidate is determined by an extension process of repeating word connection according to a word unit chain probability. Things.

Thus, while using information, which is advantageous for miniaturization, such as a chain probability in units of word strings, it is possible to present to the user in an easy-to-understand manner by converting it into units of phrases.

According to a twelfth aspect of the present invention, in the program according to the eleventh aspect, the candidate creating step further includes a candidate updating process based on an acoustic score.

In this way, by assigning the order of the candidates using the language score and the acoustic score, the candidates are presented in the order of frequency of use and candidates having the closest pronunciation, so that the user can obtain the desired candidate before the candidate is obtained. Fewer candidates need to be read.

According to a thirteenth aspect of the present invention, in the program according to the twelfth aspect, the expansion processing is performed when the number of phrase candidates subjected to the expansion processing reaches a predetermined number from the top of the language score. That's the end.

Thus, by limiting the number of candidates to be presented to the user, when the speech recognition processing is unsuccessful, the user is prompted again for voice input without continuing to display incorrect candidates endlessly. Can be processed.

A computer-readable storage medium according to a fourteenth aspect of the present invention comprises: an input step of continuously inputting voice; and a word string candidate in units of one to several words from the head of the input voice. A candidate creating step, a displaying step of displaying the candidate, and a selecting step of selecting the displayed candidate by a user, and generating the candidate for the next sound based on the selected candidate. The program records a program for executing a step, its display step, and its selection step sequentially.

Thus, the user performs the operation of deciding the candidate from the beginning of the input voice, and the unit is set to several words, so that the system needs to prepare only local candidates. Text input is possible with a small amount of processing and a small storage capacity, and the size of the apparatus can be reduced.

According to a fifteenth aspect of the present invention, in a computer-readable storage medium storing the program according to the fourteenth aspect, in the candidate creating step of the program, the linking of words according to the chain probability of each word is performed. The phrase-by-phrase unit is determined by repeated extension processing.

Thus, while using information that is advantageous for miniaturization, such as a chain probability in units of word strings, it is possible to present to the user in an easy-to-understand manner by converting it into units of phrases.

According to a sixteenth aspect of the present invention, in the computer-readable storage medium storing the program according to the fifteenth aspect, the step of creating a candidate for the program further includes a process of updating the candidate based on an acoustic score. .

In this way, by assigning the order of the candidates using the language score and the acoustic score, the candidates are presented in order from the candidate having the highest use frequency and the closest pronunciation, so that the user can obtain the desired candidate before the candidate is obtained. Fewer candidates need to be read.

According to a seventeenth aspect of the present invention, in the computer-readable storage medium storing the program according to the sixteenth aspect, the number of phrase candidates subjected to the program decompression processing is increased from the top of the language score. When the number reaches a predetermined number, the decompression process is terminated.

Thus, by restricting the number of candidates presented to the user, when the speech recognition processing is unsuccessful, the user is prompted again for voice input without continuously displaying incorrect candidates endlessly. Can be processed.

Hereinafter, an embodiment of the present invention will be described with reference to FIG.
This will be described with reference to FIGS.

(Embodiment 1) FIG. 1 is a block diagram of a text input device according to an embodiment of the present invention. In FIG. 1, an input voice fetched from an input unit 101 is input to a voice preprocessing unit 102, subjected to A / D conversion processing, and then performs feature amount extraction processing. The word candidate creation unit 103 refers to the language model 104 and creates a predetermined number of word candidates from the language model 104 following the phrase determined immediately before. Here, the language model 104 models the relationship between words in a word sequence. In the case of the first utterance, the candidate creation instructing unit 109 that has received the instruction from the operation unit 108 notifies the word candidate creation unit 103 of the instruction of the beginning of the sentence. Receiving this instruction, the word candidate creating unit 103 creates a candidate word having a high probability of being uttered at the beginning of the sentence with reference to the language model 104. The word candidates thus created are transmitted to the word string creating unit 106.

On the other hand, the creating unit 106
, The feature amount of the voice uttered in units of sentences is temporarily stored in the memory 110. The creation unit 106 performs an extension process on the word candidates from the word candidate creation unit 103 and an acoustic score update process with reference to the acoustic model 105 and the word dictionary 111, and creates a predetermined number of word strings as phrase candidates. The acoustic model 105 is a model of acoustic features. In the word dictionary 111, words to be recognized are registered as examples of phonetic symbols represented by acoustic models.
In addition, regarding the expansion processing and the acoustic score update processing,
Details will be described later.

The display unit 107 displays the created word string candidates. The user selects a correct phrase from the displayed candidates using the operation unit 108. Candidate creation instructing unit 1
09 receives the selected phrase from the word string creating unit 106 in accordance with the instruction from the operation unit 108 and outputs it as a confirmed phrase. On the other hand, the candidate creation instructing unit 109 also notifies the word candidate creation unit 103 of the determined phrase.

The word candidate creating unit 103 receives the determined phrase, and creates the next succeeding word candidate with reference to the language model 104 as described above. The above process is repeated until the input sentence is completed. After the end, the feature amount data of one sentence stored in the memory 110 is deleted.

FIG. 2 is a man-machine interface diagram using the portable telephone of the present embodiment. The voice button 201 is used to notify the start of voice recognition to the apparatus. The candidate button 202 is used to request display or change of a phrase candidate. The display screen 203 displays the determined text, phrase candidates, and the like. The confirm button 204 is for confirming a phrase candidate.

FIG. 3 is a flowchart showing the outline of the operation of the text input device of the present invention. The operation of the present invention will be described below with reference to FIGS.

First, the user presses the voice button 201, utters one sentence, and inputs a voice. Text input device
A / D conversion processing is performed on the input voice. Then, a feature extraction process such as an LPC cepstrum coefficient is performed on the converted audio signal in units of frames, for example, every 10 msec (S301).

Next, the user presses the candidate button 202 to make a phrase candidate display request (S302). The text input device creates a phrase candidate list using the feature amount of the input speech, the acoustic model, and the language model, and displays
03, one or more top candidates are displayed (S303).

This phrase candidate list is a list in which word strings are arranged in descending order of the integrated score which is the weighted sum of the acoustic score and the language score. Here, the acoustic score for the word string can be obtained as follows. The acoustic score as (i, j) for the input frame i and the dictionary frame j can be calculated by Expression (1).

[0055]

(Equation 1)

Here, "t" is transposed, "-1" is an inverse matrix, x (i)
Is the input vector corresponding to the input frame i, ((j),
((j) is the covariance matrix and the mean value vector of the feature vector corresponding to the dictionary frame j. The acoustic model described above is specifically a set of the covariance matrix and the mean value vector of these dictionary frames. The input vector is an extracted feature vector of the input speech, such as an LPC cepstrum coefficient vector, and the dictionary frame is a feature vector obtained by extracting a phrase registered in the phrase dictionary considered to correspond to the input frame from the acoustic model. Note that the feature data is not limited to the LPC cepstrum coefficient vector, but may be an MFCC (Mel Frequency Cepstral Coefficients).
Etc. can be used.

The acoustic score of a word can be obtained by finding the correspondence between an input frame and a dictionary frame by a matching method such as DP matching and adding the acoustic score on the optimal path connecting the correspondence. Furthermore, the acoustic score of a word string can also be obtained by adding the acoustic score of each word while considering the temporal consistency of adjacent words.

The language score for a word string can be obtained as follows.

The above-described language model specifically includes the word w
Chain probability P that (i) appears after n preceding words pre (i, n)
(w (i) | pre (i, n)). The language score for the word string is obtained by referring to the language model, determining the chain probability or its logarithmic value for each word in consideration of the preceding word, and adding them.

In this way, an acoustic score is obtained from the feature amount of the input speech and the acoustic model, and a language score is obtained from the word string hypothesis and the language model. A word string having a high score obtained by integrating them is registered as a phrase candidate in a list.

Next, the user confirms the displayed phrase candidate, and if not a desired candidate, presses the candidate button 202 to display the next candidate. When the desired candidate is displayed, the user presses the confirm button 204 to confirm the phrase (S304).

A determination operation is performed in units of phrases. If the phrase has not been determined until the end of the utterance, the process returns to step S302, and ends when the final phrase has been determined (S305).

As described above, in the present invention, after a candidate is determined by a user's operation for determining a phrase candidate,
There is no need to save other candidates and no need to perform recognition processing in order to create the next phrase candidate. As a result, the storage capacity required by the device can be reduced, and the device can be downsized.

Here, a linguistic unit will be considered.
Short units such as morphemes are suitable for miniaturization of the apparatus because the coverage can be increased with a small number of types. However, as a unit selected by the user, a longer unit such as a phrase is preferable because it is easy to understand. The present invention uses a morpheme as a linguistic minimum unit. In this embodiment, an example is presented in which more preferable morphemes in the interaction with a human are assembled in a phrase unit appropriately connected. This assembling process is called a morpheme decompression process.

The phrase candidate creation processing performed by the word string creation unit 106 will be described in detail below with reference to FIGS.

FIG. 4 is a flowchart showing the procedure of the phrase candidate creation process of the present invention. In this embodiment, first, a phrase candidate list is created by expanding morpheme unit candidates (S401 to S406). Next, a final phrase candidate list is created by adding the acoustic score to the result (S40).
7 to S412). Figure 5 shows the confirmed phrase "it" 50
It is an example of processing data when a phrase candidate list connected after 0 is created by decompression processing. FIG. 6 shows an example of processing data when a phrase candidate list is created by updating the acoustic score after the expansion processing.

In FIG. 5, first, a phrase candidate list 510 to be connected after the determined phrase “it” 500 is created. That is, the chain probability between “it” and all morphemes is obtained by a language model that has been learned in advance.
A list of the obtained morphemes sorted from those with the highest chain probability is referred to as a phrase candidate list 510. Each of the phrase candidates has a decompression end flag (FIGS. 5, 8, 10, and 11).
In this case, “completed” is displayed, and “0” indicating that there is a possibility of decompression is given as an initial value (S4).
01). This decompression end flag is set to “1” when the possibility of decompression has disappeared. In this state, the phrase candidates are too short to be easily understood. Therefore, the word string creation unit 1
06 searches for a relatively high linkage probability between the phrase candidate and the morpheme following it, and connects the phrase candidate to the phrase candidate.
Create longer phrase candidates.

Therefore, first, the word string creating unit 106 determines a phrase candidate to be expanded. The phrase candidates are referred to from the top of the list, and the first phrase candidate whose decompression end flag is "0" is selected (S402). Selected candidate is list 511
become that way.

Next, the word string creating unit 106 obtains a chain probability between the phrase candidate to be expanded and each morpheme that can be connected to the candidate. Here, a morpheme whose chain probability is less than a predetermined threshold,
Alternatively, morphemes whose chaining probability is smaller than the chaining probability for punctuation marks and punctuation marks are combined as “other morphemes”, and the sum of the chaining probabilities is obtained. The chain probability is the chain probability of “other morpheme” (S40
3). The calculated chain probability is as shown in list 512,
Probabilities other than "ta" and "te", which have relatively large chain probabilities from "shi", are grouped as "(other)". In the figure, (other) is used as a symbol corresponding to the concept of “other morpheme”. However, the lists 510, 520, and 530 displaying the extension completion flag (“completed”)
Are omitted from the symbols (others). (See Figure 6 below.
The same applies to FIGS. 8, 10 and 11. Next, a decompression candidate is created. The product of multiplying the chain probability of "it" → "shi" by the chain probability of "shi" → "ta",
Let it be the chain probability of "it" → "it". The phrase candidate "shi" has been expanded to "shi". Similarly, the word string creating unit 106 creates an extension candidate “do”. What is summarized as "other morphemes" has many branches to subsequent morphemes. In other words, it is considered to be suitable as a phrase boundary. Therefore, the word string creation unit 106
Regarding "other morphemes", it is considered that the decompression has been completed. Therefore, leave it as "Shi" and "She" →
Multiply “shi” by the probability of “shi” → “(other)” and use it as the chain probability. The decompression end flag is set to “1” (S
404). As a result, a list 513 of expanded candidates is created. Thus, one decompression process is completed.

Next, the word string creating unit 106 updates the phrase candidate list. That is, the word string creating unit 106 removes the pre-decompression candidate 511 from the phrase candidate list 510. Next, the word string creating unit 106 adds the post-expansion candidate 513. Then, they are rearranged according to the order of the highest chain probability (S405). As a result, the updated phrase candidate list 5
20 was obtained.

Next, the word string creating unit 106 makes an end determination. In the present embodiment, the process is terminated when the decompression process is performed 100 times, which is the preset number of times (S
406). If the decompression process is less than 100 times, it is determined that the process is not to be terminated, and the process returns to S402. By continuing the decompression process in this manner, candidates of a unit suitable for a phrase, such as “do”, “received”, and “determined”, as in the phrase candidate list 530, were obtained.

The end determination can be made when the number of phrase candidates for which the decompression end flag is set to “1” reaches a predetermined number from the top of the chain probability. In addition, it is also possible to terminate when there is no longer any phrase candidate whose decompression end flag having a chain probability greater than that of “other morphemes” is “0”.

Next, a description will be given of a method of creating a phrase candidate list in which the ranking is performed in consideration of the acoustic score.

In FIG. 6, the confirmed phrase “it (the end time 314)” 600 starts at the time when the voice button 201 is pressed in step S301 and ends when the “it” is uttered (the end time). ) Indicates 314 ms.

First, a linguistic score is obtained by performing logarithmic processing on the chain probability based on the phrase candidate list 530 created by 100 times of expansion processing. In the present embodiment, the language score is obtained from the chain probability by equation (2).

L = 20 log ₁₀ l (2) where L: language score, 1: linkage probability A suitably high value (here, 1.00) was set as the initial value of the acoustic score. The sum of the language score and the acoustic score was defined as the integrated score. Then, the word string creating unit 106
Sorts the phrase candidate list in descending order of the integrated score,
Listing 610 was sought. Also, as the end time of the utterance obtained by the acoustic matching, the end time 314 of the confirmed phrase is set as an initial value in each candidate (S40).
7).

Next, the word string creating unit 106 determines a candidate for updating the value of the acoustic score. In other words, the phrase candidates are referred to from the top of the list, and the first unupdated candidate for which the acoustic score has not been updated has been selected (S408). The determination as to whether or not the acoustic score has been updated is made based on whether or not the terminal time of the confirmed phrase and the terminal time of the phrase candidate are the same. In Listing 610,
"I did" was selected.

Next, starting at around time 314 ms,
An acoustic score for “Yes” is calculated (S409).
As a result of the acoustic matching, a relatively high acoustic score of 0.89 was obtained by the equation (1) in a speech section having a start time of 314 ms and an end time of 643 ms (List 61).
2).

A typical method of the acoustic matching includes AD processing of a voice signal, conversion to a characteristic parameter, calculation of a local distance to an acoustic model, and cumulative calculation of a local distance by DP matching. These processes can be divided into a process performed collectively in the voice input in step S301 and a process performed sequentially in the calculation of the acoustic score in step S409. Processing performed in a lump is advantageous in terms of the amount of processing because redundant calculations are prevented, and processing performed in a sequential manner is advantageous in terms of storage capacity because there is no need to save intermediate results. How to distribute the data should be determined according to the actual hardware configuration. In the present embodiment, the calculation of the local distance to the acoustic model and the cumulative calculation of the local distance by DP matching are performed in step S409.

Next, the word string creating unit 106 updates the value of the phrase candidate. That is, the acoustic score is updated to 0.89,
The integrated score was updated by finding the sum of the language score and the acoustic score. The end time of the phrase candidate was updated with reference to the matching section (S410). As a result, the new candidate is listed 613.

Next, the phrase candidate list is updated. That is,
The word string creating unit 106 deletes the pre-update acoustic score candidate 611 from the phrase candidate list 610. Next, the word string creating unit 106 stores the updated candidate 613 in the phrase candidate list 610.
Add to Then, they are rearranged in descending order of the integrated score (S411). As a result, a phrase candidate list 620 was obtained. The above process is called an acoustic score update process.

Next, the word string creating unit 106 makes an end determination. In the present embodiment, the process ends when the acoustic score updating process is performed 100 times, which is the preset number of times (S412). If the number is less than 100, it is determined that the processing is not to be ended, and the process returns to step S408. By continuing the acoustic score updating process in this manner, a phrase candidate list that is frequently used and has a high score for acoustic matching with utterance is created. This list is in the form of phrase candidates arranged in descending order of score.

Note that the end determination can be made when the number of phrase candidates whose end time is different from the determined time reaches a predetermined number from the top of the integrated score.

The text input device displays the phrase candidates at the top of the phrase candidate list obtained as described above. As a result, the text input device can perform the text input with a small processing amount and a small storage capacity by performing the voice recognition processing specified for the target phrase at that time. In addition, since one or more top candidates can be displayed in descending order of the integrated score, the number of candidates presented until the user obtains a desired candidate can be reduced. Further, by displaying the candidates in units of phrases, the selection and presentation are easy for the user to understand.

(Embodiment 2) This embodiment is different from Embodiment 1 in that the decompression process and the acoustic score update process performed by the word string creation unit 106 simultaneously update the phrase candidate list.
And different. The block configuration diagram, man-machine interface, etc. of the other text input devices are all the same.

FIG. 7 is a flowchart showing a processing procedure of a phrase candidate creation process of the text input device according to the second embodiment of the present invention.

FIG. 8 shows a confirmed phrase “it” 500
3 shows a flow of processing data when a phrase candidate list to be connected next to is created by alternately repeating decompression processing and acoustic score processing.

The details will be described below with reference to FIGS. 7 and 8.

First, the determined phrase “that” 500
The step S701 of creating a phrase candidate list 801 to be connected to the next is the same as step S401 of the first embodiment. Next, an integrated score is obtained by adding the linguistic score and the acoustic score logarithmically processed to the linkage probability of the candidate list 801 to create a candidate list 802 with an acoustic score (S7).
02). Next, from the top of the candidate list with the acoustic score, a search is performed for a candidate that has not been subjected to decompression processing, and the first candidate is set as a decompression processing candidate (S703). In the list 802, the candidate is “shi”. For this candidate, "ta", "te" and "(other)" having a relatively large chain probability are obtained from "shi" in the same manner as in S407 using the language model (S704). These phrase candidates are added to the candidate list 802 in the same manner as in the first embodiment, and rearranged in descending order of the integrated score to obtain a new phrase candidate 803 (S705).

Next, a search is made for a candidate whose terminal time is the same as the terminal time of the determined phrase candidate from the top of the candidate list.
An acoustic score of the first candidate is obtained (S706). In the list 803, “receive” corresponds to this. When an acoustic score was obtained for this candidate in the same manner as in S409, an acoustic score of 0.02 was obtained in a speech section having a start time of 314 ms and an end time of 640 ms. This is reflected on the phrase candidate list 803 (S707), and the phrase is rearranged in the descending order of the integrated score to obtain a new phrase candidate 804 (S70).
8). The processing from step S703 to step 708 is repeated a preset number of times, and a phrase candidate list 806 is obtained. In this embodiment, the process was repeated 100 times. The result of the present example was the same as the result of Example 1.

In this embodiment, the end determination is made when the expansion processing and the acoustic score update processing are repeated a predetermined number of times. However, the number of the phrase candidates for which the expansion end flag is set to “1” is determined. Can be terminated when a predetermined number is reached from the top.

Further, the end determination can be made when the number of phrase candidates whose end time is different from the determined time reaches a predetermined number from the highest rank of the integrated score.

Alternatively, the end determination can be made by the method using the decompression end flag or the method using the end time, whichever is earlier.

(Embodiment 3) In this embodiment, the expansion processing and the acoustic score update processing performed by the word string creating section are performed in the first embodiment.
Is different from the first embodiment in that the steps are performed in the reverse order. The block configuration diagram, man-machine interface, etc. of the other text input devices are all the same.

FIG. 9 is a flowchart showing a processing procedure of a phrase candidate creation process of the text input device according to the third embodiment of the present invention.

FIG. 10 shows a confirmed phrase “that” 50.
This shows the flow of processing data when a phrase candidate list to be connected after 0 is created by performing decompression processing after completing the acoustic score processing.

Hereinafter, a detailed description will be given with reference to FIGS. 9 and 10.

First, the determined phrase “that” 500
A step S901 for creating a phrase candidate list 1001 to be connected to the next is the same as step S401 in the first embodiment. Next, an integrated score was obtained by adding the linguistic score and the acoustic score that were logarithmically processed to the linkage probability of the candidate list 1001, and a temporary acoustic score candidate list 1002 was created (S902). Next, a candidate having an end time different from the end time 314 of the confirmed phrase candidate is searched from the top of the candidate list 1002, and the first candidate is determined as an acoustic score calculation candidate (S903). The acoustic score for this candidate is calculated in the same manner as in S409 (S904). Listing 10
In 02, when “shi” was selected and the acoustic score was calculated, a relatively high acoustic score of 0.90 was obtained in the speech section of the start time 314 ms and the end time 510 ms (S904). The result is reflected in the phrase candidate list 1002 (S905), and the phrase is rearranged in descending order of the integrated score to obtain a new phrase candidate 1003 (S906). Steps S903 to S906 described above were repeated a preset number of times (S907). In this embodiment, the candidate list 1004 is obtained 100 times.

Next, a decompression process was performed on the candidate list 1004 using a language model. First, the first candidate whose decompression end flag is not set to “1” is selected from the top of the candidate list 1004 (S908).

Next, referring to the chain probability of the language model (S
909), "ta", "te" and "(other)" having a relatively large chain probability are obtained from "shi" in the same manner as in step S403 (S910).

These phrase candidates are added to the candidate list 1004 in the same manner as in the first embodiment, and are rearranged in descending order of the integrated score to obtain new phrase candidates 1005 (S911).

Steps S908 to S911 are repeated 100 times, which is a preset number of times (S912), to obtain a phrase candidate 1006. The result is
Since only the value of the first morpheme is used as the acoustic score, the results are different from those of the first and second embodiments, but the same phrase is displayed in the higher rank.

In this embodiment, the end of the acoustic score update process is determined by repeating the update process a predetermined number of times. However, the number of the phrase candidates whose end time is different from the determined time is determined by the integrated score. It is also possible to end when a predetermined quantity is reached from the highest order.

Also, the end determination of the decompression process can be terminated when the number of phrase candidates for which the decompression end flag is set to “1” reaches a predetermined number from the top.

FIG. 11 shows the result of the expansion process after the completion of the acoustic score update process as in FIG. However, this processing data example shows that, in this decompression processing, after generating the decompressed candidates in step 910, the acoustic score for the connected morphemes is calculated and added to the acoustic score before decompression. Different from 10.

The end times of “do” and “do” in the candidate list 1105 of FIG. 11 are “643” and “64”, respectively.
0 ". It is more preferable to update the acoustic score in conjunction with the decompression processing, since an accurate acoustic score of a phrase candidate is obtained.

In the first to third embodiments, a description has been given of an example in which a phrase candidate list is created, the candidate is determined by inputting a confirm button, and then the next phrase candidate is created. However, in order to shorten the time from when the candidate is determined by the user to when the next phrase candidate is displayed, the process of creating the next phrase candidate is performed using the display candidate when the candidate is displayed. It is also possible. Alternatively, if there is no desired candidate list, the voice button may be pressed again, and only the phrase to be recognized is uttered to cause the apparatus to recreate the candidate.

[0108]

As described above, according to the present invention, it is possible to perform a search process in which the user sequentially selects candidates from the beginning of a sentence and determines the utterance inputted in units of sentences or sentences in units of words or phrases. Accordingly, an advantageous effect is obtained in which text input can be realized while achieving both a reduction in the size of the apparatus and a reduction in the complexity of voice input.

[Brief description of the drawings]

FIG. 1 is a block diagram of a text input device according to a first embodiment of the present invention;

FIG. 2 is a man-machine interface diagram of the text input device according to the first embodiment of the present invention.

FIG. 3 is a flowchart showing an operation of the text input device according to the first embodiment of the present invention;

FIG. 4 is a flowchart of a processing procedure of a phrase candidate creating process of the text input device according to the first embodiment of the present invention;

FIG. 5 is a diagram showing an example of processing data in a decompression process of the text input device according to the first embodiment of the present invention;

FIG. 6 is a diagram showing an example of processing data in an acoustic score update process of the text input method according to the first embodiment of the present invention.

FIG. 7 is a flowchart of a processing procedure of a phrase candidate creating process of the text input device according to the second embodiment of the present invention;

FIG. 8 is a diagram showing an example of processing data in a phrase candidate creating process of the text input device according to the second embodiment of the present invention;

FIG. 9 is a flowchart of a procedure of a phrase candidate creating process of the text input device according to the third embodiment of the present invention;

FIG. 10 is a diagram showing an example of processing data in a phrase candidate creating process of the text input device according to the third embodiment of the present invention.

FIG. 11 is a diagram showing an example of processing data in a more preferable phrase candidate creation process of the text input device according to the third embodiment of the present invention;

FIG. 12 is a flowchart of a conventional text input method.

[Explanation of symbols]

 Reference Signs List 101 input unit 102 speech preprocessing unit 103 word candidate unit 104 language model 105 acoustic model 106 word string creation unit 107 display unit 108 operation unit 109 candidate creation instruction unit 110 memory 201 voice button 202 candidate button 203 display screen 204 confirmation button

 ──────────────────────────────────────────────────続 き Continued on the front page (72) Inventor Natsuki Saito 1006 Kazuma Kadoma, Kazuma, Osaka Matsushita Electric Industrial Co., Ltd. F-term (reference) 5D015 HH23 KK02 LL08

Claims (17)

[Claims] 1. A step of continuously inputting voices, a step of generating a word string candidate in units of one to several words from the beginning of the input voice, a display step of displaying the candidates, A selection step for the user to select the displayed candidate, and a text that repeats the candidate creation step, the display step, and the selection step sequentially for the next sound based on the selected candidate. input method. 2. The text input method according to claim 1, wherein, in the candidate creation step, a phrase unit candidate is determined by an expansion process of repeating word connection according to a word unit chain probability. 3. The text input method according to claim 2, wherein said candidate creating step further includes a candidate updating process based on an acoustic score. 4. The number of phrase candidates subjected to the decompression processing is as follows:
The text input method according to claim 3, wherein the decompression process is terminated when a predetermined number is reached from the top of the language score. 5. An input unit for inputting a voice, a voice preprocessing unit for extracting a feature amount of the voice from the input unit, and a next word candidate using a language model from the determined word string. A word candidate creating unit for creating, a word string creating unit for creating a word string candidate of one to several words from the extracted feature quantity and the word candidate using at least one of a language model and an acoustic model, A display unit for displaying a word string candidate, an operation unit for selecting the displayed word string candidate by the user, and instructing the word candidate creation unit to create a next word candidate from the word string selected by the operation unit A text input device comprising: 6. The text input device according to claim 5, wherein the word string creating section creates a phrase unit candidate by performing an extension process of repeating a word connection according to a word unit chain probability. 7. The text input device according to claim 6, wherein the word string creating unit further has an updating process based on an acoustic score. 8. The text according to claim 7, wherein the word string creating unit ends the decompression processing when the number of phrase candidates subjected to the decompression processing reaches a predetermined number from the top of the language score. Input device. 9. A mobile phone having the text input device according to claim 5. 10. An inputting step of continuously inputting voices, a candidate generating step of generating word string candidates in units of one to several words from the head of the input voices, and a displaying step of displaying the candidates. Selecting a displayed candidate by a user, and sequentially repeating the candidate creating step, the displaying step, and the selecting step for the next sound based on the selected candidate. A program that lets you do things. 11. The program according to claim 10, wherein in the candidate creating step, a phrase unit candidate is determined by an expansion process of repeating word connection according to a word unit chain probability. 12. The program according to claim 11, wherein the candidate creating step further includes a candidate updating process based on an acoustic score. 13. The non-transitory computer-readable storage medium according to claim 12, wherein the expansion processing is terminated when the number of phrase candidates subjected to the expansion processing reaches a predetermined number from the top of the language score. 14. An inputting step of continuously inputting voices, a candidate generating step of generating word string candidates in units of one to several words from the beginning of the input voices, and a displaying step of displaying the candidates. Selecting a displayed candidate by a user, and sequentially repeating the candidate creating step, the displaying step, and the selecting step for the next sound based on the selected candidate. A computer-readable storage medium that stores a program for executing the above. 15. The computer-readable storage medium storing a program according to claim 14, wherein, in the candidate creating step, a phrase unit candidate is determined by an expansion process of repeating word connection according to a word unit chain probability. . 16. A computer-readable storage medium storing a program according to claim 15, wherein said candidate creating step further includes a candidate updating process based on an acoustic score. 17. The computer-readable recording medium storing the program according to claim 16, wherein the expansion processing is terminated when the number of phrase candidates subjected to the expansion processing reaches a predetermined number from the top of the language score. Storage medium.