JPS5987500A - voice recognition device - Google Patents

Abstract

(57) [Summary] This bulletin contains application data before electronic filing, so abstract data is not recorded.

Description

[Detailed Description of the Invention] [Field of Application of the Invention] The present invention relates to a speech recognition device, and in particular to a standard speech pattern (
The present invention relates to a speech recognition device suitable for simplifying the registration of standard patterns (hereinafter abbreviated as standard patterns).

[Prior art]

With the recent remarkable spread of office computers and industrial robots, technologies to facilitate the human-machine system, especially technologies related to speech synthesis and recognition, are suddenly attracting attention. If computers and robots can respond using the voices we use in everyday life, they will become more familiar to us and will become more widespread.

Speech recognition devices are divided into speaker-specific and speaker-independent systems, depending on the degree of constraints imposed by the user.

The specific speaker method, which recognizes only the voice of a specific speaker, registers a standard pattern of the voice of a specific speaker in advance and uses the degree of similarity with that pattern as a criterion for judgment, or uses the registered standard pattern to By performing optimal analysis and setting conditions for each Wl, the problem of individual difference normalization, which is one of the most troublesome problems in speech recognition, is avoided. As a result, high recognition ability has been achieved with a relatively simple device. However, it is complicated because it is necessary to register a standard pattern for each user. On the other hand, in the speaker-independent system, it is not necessary to register all standard patterns for each user, but there is a limit to recognition ability. In order to increase recognition ability, the speaker-independent method requires a plurality of standard patterns to be prepared at all times, which increases the amount of recognition processing.

However, in phoneme and M-clause recognition, the amount of information per pattern is small, so the recognition ability is generally unstable, and in the speaker-specific method, the standard pattern that happens to be registered is biased. In order to reduce the recognition ability, a countermeasure has been taken by registering multiple times and creating an average I-turn or central pattern, but the registration is complicated and not practical. Unfortunately, the speaker-independent method also lacks recognition ability for input with little information.

In this way, the specific speaker method and the non-specific speaker method each have their advantages and disadvantages, but unlike text recognition devices, voice recognition devices can be used exclusively by a specific user for a certain period of time. In other words, inputs from multiple speakers do not coexist on one device at a given time (preferably).

Further, even if a speech recognition device has a considerably high recognition ability immediately after the start of use, the device cannot fully demonstrate its original recognition ability due to the user's inexperience. In such a situation, even if a standard pattern is registered as a speaker-specific system, it is difficult to obtain a good standard pattern, and re-registration is usually required. In early situations, even speaker-specific methods cannot fully utilize their cognitive capabilities.

[Purpose of the invention]

The purpose of the present invention is to avoid the complexity of standard/turn registration of the speaker-specific method and to substantially achieve the recognition ability of the speaker-specific method, in order to improve the conventional drawbacks as described above. The purpose of this invention is to provide a speech recognition device that can perform the following tasks.

[Summary of the invention]

In order to achieve the above object, the present invention registers all standard I turns for unspecified speakers in advance, and as the use progresses, the input voice during use, < turn (hereinafter referred to as input/( turn) The standard pattern is corrected based on the accuracy of the recognition result (omitted), and as the number of uses increases, the standard pattern is gradually brought closer to the average (omitted) pattern of the user's unique pattern. do.

[Embodiments of the invention]

An embodiment of the present invention will be described below with reference to the drawings.

FIG. 1 shows the structure of one embodiment of the present invention. Reference numeral 1 denotes a control unit, which is composed of a processor such as a microcomputer. 2 is a speech recognition unit, which analyzes all the features of the input pattern; a feature analysis unit 3 that analyzes all the features of the input pattern; and inputs based on the features analyzed by the feature analysis unit 3 (feature buffer memory 4 that stores all turns, features). (It has a matching unit 5 that performs all matching with the characteristics of the input pattern stored in the sofa memory 4 and the standard pattern. 6 is the second standard pattern memory and the standard pattern transferred from the first standard pattern memory 7.・Cut turns are modified by input cover patterns. 7 is the first standard pattern memory, which stores multiple types of standard patterns. 8 is the third standard pattern memory, which stores modified patterns. 20th standard pattern (Store all standard patterns in turn memory 6. 9 displays the standard pattern with the best matching degree on the console, and stores the standard pattern in the second standard pattern memory 6 and the third standard pattern.
(Instructs the transfer of memory contents between the two memories 8.

Next, the operation of this embodiment will be explained.

A plurality of standard pattern memories stored in the first standard pattern memory 7 are copied and transferred to the second standard pattern memory 6 under the control of the control section 1, and then transferred to the second standard pattern memory 6 by the matching section 5.
Then, the input cover patterns stored in the 1% mark buffer memory 4 are compared with each other. Note that the operation of the matching section 5 will be described in detail later. Each verification result is sent to the control section 1. The control unit 1 compares the results of each comparison between the incoming cover pattern and a plurality of standard patterns, and the one with the highest degree of darkness is displayed on the display unit of the console 9 as the recognition result. The speaker who has generated the introductory pattern looks at the displayed result, determines whether the recognition result is correct or not, and inputs the determination result from the input section of the console 9.

Now, let the feature vector of the input pattern stored in the feature buffer memory 4 be X, and the first standard pattern memory 7
Standard pattern from ¥1 Second standard pattern memory 6
Let the standard < turn in the middle be Z, and the corrected standard pattern be Z'. When a speaker first starts using it, Z=Y since it has not yet been modified. When it is determined that the feature vector X of the input pattern has been correctly recognized.

・・・・・・・・・・・・・・・ (Like υ, second
The contents of standard pattern memory 6 are modified by control unit 1.

If it is determined that the recognition result is incorrect, the control unit 1
is the entire contents of the second standard pattern memory 6,...
・・・・・・・・・ Correct (2).

Note that the values of Wl and W vary depending on whether the unit of recognition is monophasic or phonological, but they are generally set to a value of 1/1082.
It's okay. It is clear that there are other methods for modifying the standard pattern that take into account the physical properties of the rigid pattern. However, the concern of the present invention is not in the content of this modification itself, but in modifying the standard pattern by the input cover pattern, so it is clear that the content of this modification procedure itself does not affect the present invention. be.

When the above series of input/correction processing is completed, the user gives an instruction from the console 9 to save the contents of the second standard/turn memory 6 to the third standard/turn memory 8.・When performing correction processing, the uncorrected first
Instead of the standard) ζ turn memory 7, the entire contents of the modified third standard turn memory 8 are transferred to the second standard turn memory 6, and the input/correction process is restarted.

Here, the operation of the matching section 5 will be explained with reference to FIG.

Figure 2 shows the case of al fd single word recognition.The input cover turn
Stored in. Let be the feature (multidimensional vector) ix in the th period, then X=x1 X,...
...Xj...XM. Word feature dictionary 1 also contains standard patterns AI, 11, . . . Ashi, etc. expressed in the same format as X. Word matching unit 201; jX and AI.

Compare At, ...A and ..., X and AI
If they are most similar, use i money. In other words, input speech X is recognized as word i.

Figure 2 (aJ) matches words in units of words, whereas Figure 2 (1) l first performs matching in units of phonemes.Then, the results of phoneme recognition are used to match words. Although the processing is more complicated than the case in Fig. 2(a), it is suitable for common speech, unspecified speakers, continuous utterances, etc. Pattern al e at +・・・・・・aj・・
. . . and the input cover pattern X are matched by the phoneme matching unit 204. The phoneme recognition result is the phoneme symbol string J+ + Jt
...... (207) is output.

The academic dictionary 205 contains the phonetic symbol string of each word. For example, the word Ak is , (k), ! k)°°°°
It has a shape like °′. This and the phonetic symbol string JI obtained earlier
J2... (207) in word matching section 2
Compare with 06. Single nJ A 1 , A 2 +...
...If Ak is closest to X among Ak, k is output as a single recognition result.

The first standard pattern memory 7 and the third standard pattern memory 8 can be easily implemented if they are configured with removable external memories such as floppy disks. The console 9 can be composed of a microcomputer or a terminal used in a normal computer. Historically, the recognition unit 2tj: It is also possible to use the techniques of various speech recognition devices that are already widely used as they are. When the input is continuous speech, the standard pattern and the corresponding input pattern can be determined using the method of the invention previously filed by the present inventors, so the standard pattern can be corrected in this case as well. is possible.

Note that the first standard pattern memory 7 and the third standard pattern memory 8 may be the same device.

Of course it is possible. It is also desirable to prepare the initial values Z to be stored in the first standard pattern memory 7 separately for men and women, etc., in order to improve correction efficiency and recognition ability.

In addition, after a certain number of uses, the correction of the descending pattern is no longer sufficient and the recognition becomes stable, so from the viewpoint of processing efficiency, the control unit is configured to omit the correction process according to instructions from the console 9. It is desirable to configure 1.

[Effects of the Invention] As explained above, according to the present invention, it is possible to modify the standard pattern according to the input pattern, thereby avoiding the complexity of standard pattern registration in the speaker-specific method, and It is possible to maintain the level of cognitive ability of a person's method.

[Brief explanation of the drawing]

FIG. 1 is a diagram showing a general configuration 9 of an embodiment of the present invention, and FIG. 2 is a diagram for explaining the operation of the matching section shown in FIG. 1.

Claims (1)

[Claims] 1. A first memory in which the standard voice pattern 'f-, Rirakashi is registered, a second memory in which the contents of the first memory are stored, and a second memory in which the contents of the first memory are stored; means for recognizing input speech using a standard speech pattern, means for determining whether the input speech recognition result is correct or incorrect, and a standard speech pattern stored in the second memory based on the result of the correctness judgment and the characteristics of the input speech '(il- A speech recognition device characterized in that it is provided with means for correcting. 2. A third memory is provided for storing the contents of the second memory after being corrected by the correcting means; The speech recognition device according to claim 1, wherein the second memory stores the contents of the third memory for input speech from the first time onward. 3. Memory of MiJ memory 1 and the third memory are removable memories. 4. The first memory and the third memory are provided in the same storage device. The speech recognition device according to claim 2, characterized in that: 5. means for determining whether the input speech is correct or incorrect and means for correcting the standard speech pattern when the input speech is input a certain number of times or more; 5. The speech recognition device according to claim 1, 2, 3, or 4, further comprising means for stopping each operation.