JPS6232798B2 - - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION The present invention relates to an improvement of a voice input device using a pattern matching method.

Conventionally, technology has been put into practical use to recognize speech by registering previously uttered speech as a standard pattern and performing pattern matching between the newly uttered speech and the standard pattern (see,
Information Processing Society of Japan Research Group Material MMS23-2, 1976 1
May 20, ``Continuous speech recognition system using DP'' (hereinafter referred to as cited document (1)).

In such a device, it is necessary to register the uttered voice in advance as a standard pattern. At this time, conventionally, standard patterns were registered in a predetermined order according to a predetermined word dictionary. When using numbers or alphanumeric characters as words to be registered, these should be 1, 2, 3,
..., A, B, C, etc. It is normal and easy to utter the words in this order. However, in this case, even if the words are uttered separately, there is a risk that a certain intonation will be added, or that the pronunciation before and after two words that are next to each other will change continuously. For this reason, there is a possibility that a standard pattern different from that in the case where each word is uttered independently may be registered.

Further, in general, standard patterns are registered in the voice input device repeatedly, for example, once a day. At this time, if you get used to saying the words in the order of the word dictionary, the patterns with certain peculiarities that are different from when you say each word independently are registered as standard patterns for general words, just like numbers and alphanumeric characters. There is a risk of Furthermore, if the speaker becomes accustomed to the order of standard pattern registration and remembers the next word to be uttered, the utterance of the word may become inaccurate.

The purpose of the present invention is to prevent the occurrence of peculiar or inaccurate utterances due to the order of occurrence that occurs when standard patterns are registered in a voice input device using such a pattern matching method, and to obtain reliable standard patterns. It is in.

In order to achieve the object, the voice input device of the present invention includes a feature extraction section that extracts features from uttered speech and converts it into a pattern, a standard pattern registration section that registers the pattern as a standard pattern, and a standard pattern registration section that registers the pattern as a standard pattern. a similarity calculation unit that calculates the degree of similarity between the input pattern patterned in the feature extraction unit and the standard pattern from the voice obtained by the feature extraction unit; a result output unit that calculates a recognition result based on the degree of similarity; The present invention is characterized in that it includes a random number generation section that gives a random order when registering the patterns to the standard pattern registration section.

Examples will be described in detail below.

FIG. 1 is a block diagram of a voice input device using a pattern matching method. In the figure, 1 is a microphone, 2 is a feature extraction unit, 3 is a standard pattern registration unit,
4 is a similarity calculation section, 5 is a result output section, and 6 is a mode changeover switch. The audio signal WV input from the microphone 1 is patterned by the feature extractor 2 to become an audio pattern P. In a voice input device using a pattern matching method, its operation is divided into two: a standard pattern registration mode and a recognition mode. First, in the standard pattern registration mode, the mode changeover switch 6 is set to the registration mode TR. The audio pattern P input at this time is registered in the standard pattern registration section 3 as a standard pattern.
In the recognition mode, the mode changeover switch 6 is set to the recognition mode OP, and the similarity calculation unit 4 calculates the similarity S between the input speech pattern P and the standard pattern RP output from the standard pattern registration unit 3. and output. Based on this similarity S, a recognition result R is outputted by the result output unit 5. The above is the operation of a voice input device using the pattern matching method, and the device shown in document (1), for example, also operates in a similar manner. The feature of the present invention is the operation in the above standard pattern registration mode, and is not limited to the operation in the recognition mode, that is, the pattern matching method. Therefore, the operation of the standard pattern registration mode will be described in more detail.

FIG. 2 shows a partial configuration necessary when registering a standard pattern, and shows the portion surrounded by dotted lines in FIG. 1 extracted. In the figure, 11 is a random number generator, 12 is a word dictionary, 13 is a display,
14 is a standard pattern memory. Note that the microphone 1, feature extraction section 2, and standard pattern registration section 3 surrounded by dotted lines in FIG. 2 are the same as in FIG. 1, and the mode changeover switch 6 is omitted. First, when registering a standard pattern of N words, N word names are set in the word dictionary 12 in advance. This word name is displayed on the display 13 according to the address signal AD from the random number generator 11. The speaker is this display 1
Read the word name displayed in 3 and say it into microphone 1. The uttered voice is patterned by the feature extractor 2 and then stored in the standard pattern memory 14. At this time, it is stored in the area in the standard pattern memory 14 indicated by the same address signal AD as that given to the word dictionary 12. As a result, a standard pattern corresponding to the word name is registered. When the voice input is finished, the voice detection signal SD is output, and the random number generation unit 11 generates the next address signal AD.
is output. The random number generator 11 outputs a number from 1 to N as an address signal AD, and the address signal AD changes to the next number in response to the voice detection signal SD.
That is, the numbers 1 to N are randomly rearranged and output. An example of a method for performing this random sorting will be shown. FIG. 3 shows an example of the configuration of the random number generator 11, where 31 is an address counter, 32 is an N-bit address flag memory that can be accessed one by one, and 33 is a random number counter.
34 is a random number generator, and 35 is a random number maximum value register. First, the contents of the address flag memory 32 are all set to 1, and the random number maximum value register 35 is set to N. The random number generator 34 generates a random number less than the value set in the random number maximum value register 35. There are many methods of generating random numbers, such as a method using an M sequence, but there is also a method of counting the time intervals between vocalizations and using them as random numbers.

The address counter 31 is reset by the address output signal SD, and the random number generated from the random number generator 34 is set in the random number counter 33. Next, the address counter 31 receives the clock signal.
Counted up by CLK. Data R is read from the address flag memory 32 in response to the address signal from the address counter 31, and if this value is 1, the random number counter 33 counts down. This is repeated and the random number counter 33
When becomes 0, a zero signal ZF is output, whereby the address signal from the address counter 31 is output to the word dictionary section 12 and the standard pattern memory 14. At the same time, the clock signal CLK to the address counter 31 is masked and the address flag memory 32 is masked.
0 is written to the currently addressed portion of . At the same time, the random number maximum value register 35 is subtracted by 1 and becomes N-1. Then, the random number generator 34 again generates a random number with a maximum value of N-1. Next, when the address output signal SD is input, the above operation is repeated. This causes the address signal AD to be output every time the address output signal SD is input.
The numbers from 1 to N are randomly rearranged and output. The above is an example of the operation of the random number generator 11. Although various other methods can be considered, the present invention is not limited to this method of random rearrangement.

The key point of the present invention is to register standard patterns in a different random order each time, and is not limited to the embodiments. For example, display 1
The word name displayed at 3 may be outputted by voice, and the address output signal inputted to the random number generation section 11 may be inputted using another key.

According to the voice input device of the present invention, since standard patterns are registered in a different order each time, it is difficult to get accustomed to a fixed order, and the conditions are similar to those when uttering independently. Also, since it requires some attention when uttering, the utterance becomes more accurate.

[Brief explanation of the drawing]

FIG. 1 is a block diagram of a voice input device using a pattern matching method, FIG. 2 is a block diagram of a standard pattern registration section in an embodiment of the present invention, and FIG. 3 is a block diagram of a random number generation section. In the figure, 1 is a microphone, 2 is a feature extraction unit, and 3
is a standard pattern registration section, 4 is a similarity calculation section, 5 is a result output section, 6 is a mode changeover switch, 11 is a random number generation section, 12 is a word dictionary, 13 is a display, 14 is a standard pattern memory, and 31 is an address counter. , 32 is an address flag memory, 33 is a random number counter, 34 is a random number generator, and 35 is a random number maximum value register.

Claims (1)

[Claims] 1. A feature extraction unit that extracts features from uttered audio and converts it into a pattern, a standard pattern registration unit that registers the pattern as a standard pattern, and an input that is patterned by the feature extraction unit from newly input audio. a similarity calculation unit that calculates the similarity between a pattern and the standard pattern; a result output unit that calculates a recognition result based on the similarity; and a random order when registering the standard pattern in the standard pattern registration unit. What is claimed is: 1. A voice input device comprising: a random number generating section that generates a random number.