JPH0475098A - Voice recognition device - Google Patents

Abstract

PURPOSE:To automatically remove a meaningless recognition result which is added to the heat of a word by erasing all recognition results corresponding to the syllable of the head of the work and re-editing a recognition result, and then instructing a process based on the recognition result after the re-editing to a language processing part and a display part. CONSTITUTION:There is the meaning less recognition result due to a noise at the position corresponding to the work head of a paragraph candidate displayed in a menu display area 11 and when an operator judges that the meaningless recognition result should be removed, the removal of the recognition result of the word head is instructed by depressing a key on a keyboard 8. The syllable candidate corresponding to the syllable of the word head in a generated syllable lattice is erased under the control of a control part 6. Then a paragraph candidate array is generated again according to a re-edited syllable lattice and displayed in the menu display area 11 by a display part 7. Consequently, the meaningless recognition result added to the word head of each candidate array can automatically be removed.

Description

[Detailed description of the invention] [Industrial application field]

The present invention relates to a speech recognition device that recognizes input speech and creates a candidate sequence that is considered to be grammatically correct.

[Conventional technology]

When the input voice is recognized by the voice recognition device, noise such as the operator's breath or the operation sound of the voice recognition device is introduced before and after the human voice, and meaningless recognition results due to this noise are generated before and after the input voice. There are cases in which the correct recognition result may or may not be obtained due to the addition of the recognition result. Conventionally, when noise appears before or after a human voice as described above, the meaningless recognition result due to the noise is removed by operating the "reverse" key.

[Problem to be solved by the invention]

As mentioned above, when removing meaningless recognition results by operating the "backward" key, if meaningless recognition results are added after the input speech, the meaningless recognition results are after the speech section. Therefore, the meaningless recognition results can be easily erased by operating the "backward" key. However, when a meaningless recognition result is added before the input speech, the meaningless recognition result is at the beginning of the word, so key operations to move the cursor to the meaningless recognition result are cumbersome. There is. Furthermore, in order to avoid troublesome key operations, there is also a method of canceling the recognition result of the input voice and re-uttering the same voice. However, in this case, it is not always possible to obtain impressive recognition results as a result of repeating the voice, and there is a problem in that the user must rely on cumbersome key operations. Therefore, the purpose of this invention is to automatically remove meaningless recognition results added to the beginning of words without having to perform cumbersome key operations or canceling/re-voicing recognition results. An object of the present invention is to provide a speech recognition device that can eliminate this problem.

[Means to solve the problem]

In order to achieve the above object, the present invention recognizes human speech in units of phonemes or syllables by a recognition unit based on feature parameters extracted from an input speech signal by a speech analysis unit, and based on the recognition results. A speech recognition device that selects and outputs a correct candidate from a generated candidate string, comprising: a recognition result storage unit that stores recognition results obtained by phoneme unit or syllable unit recognition in the recognition unit; a language processing unit that performs language processing using the recognition results stored in the storage unit to generate the candidate string on a word-by-word or phrase-by-phrase basis; and a display that displays a menu of the candidate strings generated by the language processing unit. Department and
When a meaningless recognition result is added to the beginning of a word in a candidate string displayed in the menu by the display section, an instruction means is operated by an operator to instruct removal of the meaningless recognition result; When instructed by the instruction means to remove meaningless recognition results, all recognition results corresponding to the initial syllable of a word are deleted from among the recognition results stored in the recognition result storage section and the recognition results are deleted. The present invention is characterized by comprising recognition result reorganization means for instructing the language processing unit and display unit to perform processing based on the reorganized recognition results after the reorganization.

[Effect]

A voice analysis unit extracts feature parameters from the input voice signal. Then, based on the extracted feature parameters, the recognition unit recognizes the input speech in units of phonemes or units of syllables. Then, the obtained recognition result is stored in the recognition result storage section. Then, using the recognition results stored in the recognition result storage section, language processing is performed by the language processing section to generate a candidate string on a word-by-word or phrase-by-phrase basis. The generated candidate string is then displayed as a menu on the display unit. The candidate columns displayed in the menu on the display section are referred to by the operator. If it is determined that a meaningless recognition result is added to the beginning of a word in each candidate string, the operator operates the instruction means to instruct removal of the meaningless recognition result. Then, the recognition result reorganization means deletes all the recognition results corresponding to the initial syllable of the word from among the recognition results stored in the recognition result storage section. Further, the language processing unit and display unit are instructed to perform processing based on the recognition result after the reorganization. Then, the language processing unit generates a candidate string again based on the recognition results after removing and reorganizing the meaningless recognition results added to the beginning of the word, and eliminates the meaningless recognition result added to the beginning of the word. An incorrect recognition result is displayed on the display. Therefore, just by operating the above-mentioned instruction means by the operator, meaningless recognition results added to the beginning of each word of each candidate string are automatically removed, and recognition results in which meaningless recognition results are not added to the word order are obtained. is displayed in a menu on the display.

【Example】

Hereinafter, the present invention will be explained in detail with reference to illustrated embodiments. FIG. 1 is a block diagram showing one embodiment of the present invention. The audio signal input from the microphone l is sent to the audio analysis unit 2.
A/D conversion is performed by , and feature parameters are extracted for each frame. The time series of feature parameters extracted in this way is cut out into syllables to obtain a feature pattern, and the obtained feature pattern is sent to the syllable recognition unit 3. In the syllable recognition unit 3, the input speech is recognized in units of syllables based on the similarity between the input feature pattern and the syllable standard pattern created in advance, and the recognition result (i.e., syllable lattice) obtained as a result of the recognition is is stored in the recognition result storage section 4. In the language processing section 5, phrase candidates are created from the combinations of syllable lattices found by the syllable recognition section 3, and a grammatically correct phrase candidate string is created using a word dictionary or the like. The phrase candidate string thus obtained is displayed on the display section 7. The keyboard 8 is used to enter various instructions manually, such as an instruction to move a cursor when selecting a correct clause candidate from the list of clause candidates displayed on the display section 7. The control unit 6 also controls the voice analysis unit 2. Syllable recognition unit 3. Language processing section 5, display section 7 and keyboard 8
to perform voice recognition processing, which will be described in detail later. In the present invention, noise such as the operator's breath or the operation sound of the keyboard 8 is input to the microphone l when voice recognition processing is executed under the control of the control unit 6, and When a meaningless recognition result is added to the beginning of a word, this meaningless recognition result is automatically deleted. Hereinafter, the operation of the speech recognition device having the above configuration will be specifically explained, taking as an example the case where the phrase /shakaiwa (society wa)/ is input by voice. The operator speaks into the microphone.
When the human voice starts to be uttered, the voice analysis unit 2 performs an acoustic analysis of the human voice /shakaiwa/ and sequentially outputs the characteristic parameters. Then, the syllable recognition unit 3 performs syllable recognition using the degree of similarity between the feature pattern in which the output feature parameters are cut out into syllables and the syllable standard pattern, and obtains the recognition result as shown in FIG. At this time, the recognition results are output as a syllable lattice and stored in the recognition result storage section 4, since speech recognition includes ambiguity due to erroneous detection of syllable sections, erroneous recognition of syllables, etc. Here, suppose that noise occurs and a meaningless recognition result is added to the beginning of the word just before the operator utters /shakaiwa/ into the microphone l. Therefore, in Figure 2, the syllable string /sha/, /ka/
The recognition results (syllable candidates) for noise are added before the syllable lattice for , /i/, /wa/. Assume that, at the stage where the voice/challenge/manpower has been completed, an instruction is given from the keyboard 8 for the language processing unit 5 to perform language processing. Then, in the language processing section 5, first, the syllable recognition section 3
With reference to the syllable lattice shown in FIG. 2 generated by
A clause candidate /kyusakaiwa/ with the first candidate rank consisting of only kyū/, /sa/, 7'ka/, /i/, and /wa/ is generated and output. Thereafter, the syllable candidates constituting the phrase candidate /kyusakaiwa/ are sequentially replaced with the syllable candidates read from the syllable lattice according to a predetermined rule to create phrase candidates one after another, and the phrase candidates are output in order of cumulative similarity. As a result, a phrase candidate string as shown in FIG. 3 is generated. next,
The language processing unit 5 compares the phrase candidate sequence generated as described above with the contents of a word dictionary (not shown), and rejects grammatically incorrect phrase candidates that are not in the word dictionary. Then, under the control of the control section 6, the display section 7 uses a well-known technique, for example, a window display technique, to display a phrase candidate string consisting of a plurality of phrase candidates generated by the language processing section 5, as shown in FIG. Menus are displayed in the menu display area 11 of the display screen in descending order of candidate ranking. At this time, the display area 12 at the top of the menu display area If displays the cursor among the phrase candidate strings for the already recognized phrase "Japan" and the input speech /Shiyakaiwa/ for which the recognition process is currently being performed. The sentence ``Japanese Shusakaiha'' consisting of the clause candidate ``Shusakaiha'' specified by No. 13 is displayed at the same time. In this case, as mentioned above, the meaningless recognition result (/shu/ or /kiyu/) due to noise is added to the beginning of the word, so the correct phrase candidate ``Kujiyakaiwa'' cannot be obtained. There isn't. However, as can be seen from FIG. 4, the meaningless recognition result "shu" caused by the above noise is removed from the phrase candidate "shushakaiha" displayed in the second row from the top of the menu display area 11. Then, the correct phrase candidate ``
It turns out that Shiyakai can get J. Therefore, in this embodiment, the removal of meaningless recognition results caused by the noise added to the beginning of a word is instructed by one action of pressing one key on the keyboard 8, for example. Then, under the control of the control unit 6, one noise (initial syllable) is selected from among the syllable candidates constituting the syllable lattice related to the input voice /Jakaine/', which is stored in the storage unit.
The syllable lattice is reorganized by erasing the syllable candidates corresponding to . Thereafter, the reorganized syllable lattice from which the syllable candidates corresponding to the noise have been deleted is expanded again to obtain a string of phrase candidates as shown in FIG. Then, the language processing section 5 performs language processing again, and the results of the language processing are displayed as a menu again in the menu display area 11 of the display screen, as shown in FIG. That is, the keyboard 8 constitutes the instruction means, while the control section 6 constitutes the recognition result reorganization means. In the above case, since meaningless recognition results due to noise have already been removed, the correct clause candidate "Shakaiha" is displayed in the second row from the top of the menu display area 11. Therefore, the cursor 13 is moved to the position of the bunsetsu candidate "Shiyakai wa" using the keys from the keyboard 8, and the bunsetsu candidate "Shiyakai wa" is selected. In this way, the bunsetsu candidate "Shiyakai wa" is selected. in a state where
For example, when you operate the kana-kanji conversion key, the phrase candidate "shakaiwa" is selected from the phrase candidate string displayed in the menu display area 11, and the input voice /shakaiwa/
The recognized phrase ``Shiyakaiha'' is converted into kana and kanji characters and displayed in the display area 12 as ``Japanese society wa''. In this manner, in this embodiment, the syllable recognition results (ie, syllable lattice) from the language processing unit 5 are used to generate a string of phrase candidates arranged in the order of candidate ranks. and,
The generated sentence candidate string is displayed as a menu in the menu display area 11 by the display unit 7. Here, there is a meaningless recognition result that seems to be caused by noise at the beginning of the phrase candidate string displayed in the menu display area IT, and the operator has determined that this meaningless recognition result should be removed. If you do,
Removal of the recognition result of the beginning of a word is instructed by pressing a key on the keyboard 8 or the like. In this way, when an instruction is given to remove the recognition result of the beginning of a word, syllable candidates corresponding to the beginning syllable of the word in the syllable lattice generated as described above are deleted under the control of the control unit 6. Then, based on the syllable lattice rearranged in this way, a line of phrase candidates is created again and the menu is displayed in the menu display area 1.1 by the display unit 7. Therefore, the sounds of the operator's breath, the operation of the keyboard 8, etc. When a meaningless recognition result is added to the beginning of a word due to noise, it can be automatically removed with a simple instruction such as pressing a key. It is possible to automatically obtain a string of phrase candidates that are free from the effects of noise without having to perform cumbersome key operations for movement or canceling/re-voicing the recognition results. The recognition unit 3 recognizes each syllable, and the language processing unit 5 obtains a phrase candidate string based on the recognition result.However, the present invention is not limited to this. The word candidate string may be obtained by recognizing each phoneme.

【Effect of the invention】

As is clear from the above, the speech recognition device of the present invention is
The language processing unit generates a candidate string for each word or phrase. Then, when the operator refers to the candidate string displayed on the menu on the display and operates the instruction means to instruct that the meaningless recognition result added to the beginning of the word be removed,
After the recognition result reorganization means erases all recognition results corresponding to the initial syllables of words stored in the recognition result storage section and reorganizes the recognition results, the language processing section and display section are configured to display the reorganized results on the language processing section and display section. Since the process for the recognition result is instructed, the meaningless recognition result added to the beginning of each candidate string can be automatically removed by one action of operating the instruction means. Therefore, according to the present invention, meaningless recognition results added to the beginning of words due to noise etc. during voice input are automatically removed without bothersome key operations or canceling/re-speaking the recognition results. You can do it.

[Brief explanation of the drawing]

FIG. 1 is a block diagram of one embodiment of the speech recognition device of the present invention, FIG. 2 is a diagram showing an example of a syllable lattice generated by the syllable recognition unit in FIG. 1, and FIG.
FIG. 4 is a diagram showing an example of a phrase candidate string generated by the language processing unit in FIG. , FIG. 5 is a diagram showing an example of a phrase candidate string generated based on the reorganized syllable lattice, and FIG. 6 is an example of a phrase candidate string displayed in a menu based on the phrase candidate string shown in FIG. FIG. ■...Microbon, 2...Speech analysis department, 3.
・Syllable recognition unit, 4...Recognition result storage unit, 訲...
Language processing unit, 6. Control unit, 7. Display unit,
8... Keyboard, 11 Menu display area, 12... Display area, 13... Cursor. Figure 1 Figure 4 Figure 5 Input noise /sha/ /ka, //i/ /wa/ Similarity 1
Ranking: 2nd place
Shusiya is in 3rd place
Figure 3.

Claims (1)

[Claims] (1) The recognition unit recognizes the input speech in units of phonemes or syllables based on the feature parameters extracted from the input speech signal by the speech analysis unit, and the correct candidate is selected from the candidate string generated based on the recognition results. A speech recognition device that selects and outputs a recognition result storage unit that stores recognition results obtained by phoneme unit or syllable unit recognition in the recognition unit; a language processing section that performs language processing using the computer to generate the candidate string on a word-by-word or phrase-by-phrase basis; a display section that displays the candidate string generated by the language processing section as a menu; If a meaningless recognition result is added to the beginning of a word in a candidate string, the instruction means is operated by an operator to instruct the removal of the meaningless recognition result; When deletion is instructed by the instruction means, all recognition results corresponding to the initial syllable of the word are deleted from among the recognition results stored in the recognition result storage section, the recognition results are reorganized, and then the recognition results of the language A speech recognition device comprising recognition result reorganization means for instructing a processing section and a display section to perform processing based on the reorganized recognition results.