JPH0555880B2 - - Google Patents

Description

[Detailed description of the invention]

<Technical Field> The present invention relates to an improvement of a speech recognition device that performs a recognition determination by calculating the similarity between a speech feature pattern to be recognized and a plurality of pre-registered speech feature standard patterns, The present invention relates to a speech recognition device that provides means for determining whether a standard pattern of speech characteristics registered in advance is good or bad. <Prior art> Devices have been put into practical use that register a plurality of feature standard patterns and perform speech recognition by matching the standard patterns with input feature patterns. Once a feature standard pattern is registered in , it is not possible to quantitatively know the quality of the registered feature standard pattern, and it is necessary to judge the quality of the feature standard pattern (voice pattern) empirically from the recognition results. Ta. <Purpose> The present invention has been made in view of the above points, and includes means for quantitatively determining whether or not a feature standard pattern registered in advance is necessary based on an evaluation value according to a recognition result. It is an object of the present invention to provide a speech recognition device that can select and replace registered patterns depending on the quality of feature standard patterns. <Example> Hereinafter, the present invention will be explained in detail by giving an example. FIG. 1 is a block diagram showing the configuration of an apparatus according to an embodiment of the present invention, which recognizes speech uttered word by word in monosyllable units, performs dictionary checking on a plurality of word candidates, and calculates the recognition results. An example of a speech recognition device for output is shown. In FIG. 1, reference numeral 1 indicates a microphone for picking up speech input, and reference numeral 2 indicates a microphone that picks up speech input, and 2 indicates a system in which the speech that is uttered word by word and input via the microphone 1 is analyzed for each monosyllable to form an input pattern, which is stored in a standard pattern memory 3. A monosyllable recognition unit that matches a standard pattern and an input pattern and outputs a recognition result; 3 is a standard pattern memory that holds registered standard (feature) patterns; 4 is each standard stored in the standard pattern memory 3; A counter memory means for storing a predetermined count value corresponding to a pattern, 5 a word dictionary memory for storing words necessary for dictionary comparison, and 6 a standard for storing a plurality of words for testing standard patterns. A pattern test word memory 7 is a keyboard input device, for example, as shown in FIG. 2, there are a kana key 7a, a conversion/next candidate key 7b for ending word input and calling up the next candidate, and an instruction to confirm the recognition result. A confirmation key 7c for instructing the recognition result, a correction key 7d for instructing correction of the recognition result, and the like are provided. Further, 8 is a display device for displaying recognition results, etc., 9 is a floppy disk device used for saving standard patterns, etc.;
10 is a controller (CPU) that controls each of the devices 2 to 9 described above. The above standard pattern memory 3 has “A” to “N”.
Up to 5 monosyllabic characteristic patterns (A
~E) are stored. Further, an initial value "80", for example, is set and stored in the counter memory means 4 corresponding to each of the above-mentioned standard patterns, as shown in FIG. 3, respectively. The initial registration operation of information in the standard pattern memory 3 and counter memory means 4 is performed according to the initial registration flow shown in FIG. That is, when a predetermined key of the key input device 7 is operated to set the device to the initial registration operation mode, the CPU 10
A single syllable to be uttered, for example "a", is displayed on the display device 8 under the control of (steps n1, n2). When the operator confirms the monosyllables displayed on the display device 8 and utters the syllables, the uttered syllables are input via the microphone 1 (n3), and the syllables are input to the monosyllable recognition unit 2.
A feature pattern is created for the input speech (monosyllabic), and this analyzed input pattern (feature pattern) is stored by the CPU 10 at a predetermined location in the standard pattern memory 3 (for example, a location corresponding to _A ).
(n4), and an initial value "80" is set at a predetermined position in the counter memory means 4 corresponding to this registered standard pattern ( _A ) (n5). Such a series of operations is performed for all the standard patterns, and as a result, the counter value for each standard memory in the counter memory means 4 becomes the third
As shown in the figure, each is set to the initial value "80". Then a value (e.g. "80") as above
The operation in which the value set in the counter memory means 4 increases or decreases in accordance with the recognition operation, etc. will be explained. (i) Increase/decrease of counter value during recognition The processing flow during recognition is shown in Figure 5.
An example of recognizing the input voice "red" will be explained. If you now put the device in recognition operation mode and utter the voice to be recognized, for example "/a//ka//i/" (red), this voice will be input through microphone 1 (n11, n12) and will be monosyllable. The recognition unit 2 sequentially recognizes the input speech monosyllable by monosyllable, and as a result of recognizing "a", the following patterns are close to the standard pattern: " _AB ,"" _HaD ,"" _AC ," and " _PaA ." Recognized monosyllable candidates are obtained. Next, "ka" is recognized, and "i" is similarly recognized, resulting in the recognition results for each syllable as shown in Table 1 (n13).

[Table] Here, when an instruction is input using the conversion key 7b of the key input device 7 to indicate the end of the word voice input (n14), a syllable latitude as shown in Table 2 is created under the control of the CPU 10.

[Table] Next, a string of candidate words is created from this syllable latex in order of likelihood as shown in Table 3.

[Table] After that, each created candidate string is compared with the word information stored in the word dictionary memory 5, and the matched candidate strings as shown in Table 4 are sent to the CPU 10.
The dictionary matching results are stored in the buffer (n15),
The first candidate of the matching results is displayed on the display device 8.

[Table] The operator checks the display and, if the word is the one he was trying to input, instructs ``Confirm'' by operating the key 7c of the key input device 7, and if not, instructs ``Next Candidate'' by operating the key 7c of the key input device 7. The user operates the key 7b to instruct the user to display the next recognition result on the display device 8. In addition, if the word to be input is not included in the candidates, the user must correct the word using the "modify" key 7d or the kana key 7a, and then input the "confirm" key 7c (n17 to n22). In this example, "Red" is the first one as input.
Since it is a candidate, immediately press the "Confirm" key 7c.
(n17, n18). This operation moves to steps n23 and n24, and first the recognition result of the first syllable (n=1) is checked. In step n25, it is determined whether the first standard pattern is correct or not, and the recognition result for "A" is that the first standard pattern is " _AB " as shown in Table 1 above. It is determined that the answer is correct, and the process moves to step n26, where the contents of the counter memory means 4 at the position corresponding to the standard pattern " _AB " are incremented by "+2". Next, the process moves to step n28, where it is determined whether or not the second-place standard pattern is correct.As shown in Table 1, the second-place answer for "A" is " _HaD ", so
The process moves to step n30, and the stored content of the counter memory means 4 at the position corresponding to the standard pattern of ` <sub>`D</sub> '' is set to ``-1'', and then the process moves to step n31, where n
=2, returns to n24 through n32, moves to the operation of checking the recognition result of the second syllable, and the same operation is repeated thereafter. Now, the recognition result for the second syllable "ka" is determined to be correct because the first place is "ka <sub>C</sub> ", and the value for "ka <sub>C</sub> " in the counter memory means 4 is "+2".
Since the second place is " <sub>A</sub> ", it is determined that the answer is incorrect, and the value for " <sub>A</sub> " in the counter memory means 4 is set to "-1". Also, the third syllable “i”
The recognition result for `` _HIC '' is determined to be incorrect because the first place is ``HIC'', and the value of the counter memory means 4 for `` _HIC '' is set to ``-2'', and the second place is ``HIC''.
Since the answer is " _B ," it is determined that the answer is correct, and the value for " _B " in the counter memory means 4 is incremented by "+1." Through the above operations, the contents of the standard pattern counter memory means 4 change as shown in FIG. (ii) Increase/decrease of the counter value in the test mode The processing flow of the test mode is shown in FIG. The stored words (words to be uttered) are sequentially read out and displayed on the display device 8 (n42). If an instruction to say "Akai" is now made on the display device 8 (n42), the operator speaks the word displayed on the display device 8 and inputs the voice (n43). This input voice “/a//ka//i/”
is sequentially recognized monosyllable by monosyllable in the monosyllable recognition unit 2, and recognition results are obtained for each syllable as shown in Table 1 above, as in the above recognition mode. In the test mode, since the uttered word has been determined, the same operations (steps n46 to n57) as those after the operation of the "confirm" key (n23) in the recognition mode are performed thereafter. Further, in this series of operations, words are sequentially read out from the standard pattern test word memory 6 and displayed, and similar processing is executed. By selecting and storing words in the standard pattern test word memory 6 so that each syllable appears evenly, it is possible to perform a test on all standard patterns. (iii) Increase/decrease of counter value in recognition process between standard patterns The process flow of recognition between standard patterns is shown in Figure 8. When the recognition process between standard patterns starts in the speech recognition device (step n61). ), the CPU 10 selects one of the standard patterns stored in the standard pattern memory 3 and supplies it to the monosyllable recognition unit 2 as an input pattern (n62).
This input pattern is matched with each standard pattern stored in the memory 3 in the monosyllable recognition unit 2, and the recognition result is sent to the CPU 1.
Give to 0 (n63). The CPU 10 determines whether the given first recognition result matches the input pattern (n64), and if it does not match, it determines that some kind of hard error has occurred, and the error occurs. Display is performed on the display device 8, and subsequent operations are canceled (n65, n66). If the first recognition result matches the input pattern, the process moves to step n67 to check the second to sixth recognition results. Now, for example, as shown in Table 5, as a result of recognition processing using "A _A " as an input pattern, the first place is "A _A ", and the second and sixth places are "A".
_{``C</sub> '', ` <sub>`A</sub> '', `` _AE '', `` <sub>KaB</sub> '', `` _AD '',
Those in the same category as the input pattern are “ _AC ”,
It was determined that there were three characters, “ _AE ” and “ _AD ” (n68).
The evaluation value of the input pattern is calculated as ``6'', which is 2 times ``3'' (n69), and this value is added to the value in the counter memory location corresponding to the standard pattern of ` <sub>`A} '' (n71), and then The next standard pattern is selected after passing through step n72 and proceeding to n73.
Returning to n62, the same operation is performed.

【table】

Claims (1)

[Scope of Claims] 1. A recognition device that performs a recognition determination by calculating the similarity between a feature pattern of a voice to be recognized and a standard pattern of features of a plurality of types of voices registered in advance, comprising: A counter means provided for each pattern, an input means for inputting whether the candidate string obtained as a result of recognition determination is correct or not, and a counter means for inputting whether or not the candidate string obtained as a result of recognition determination is correct, and in response to this input means, it is determined which order of the candidate in the candidate string is correct or not. and means for increasing or decreasing the content of the corresponding counter means based on the evaluation value weighted by.