JP2002132293A - Speech recognizer - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a speech recognizer which exactly recognizes the speech inputted with simple constitution. SOLUTION: This speech recognizer comprises a word dictionary data base for speech recognition, a speech input section which computes the speech characteristic quantity from the inputted speech, a collation section which determines the similarity of the computed speech characteristic quantity and the respective words stored in the dictionary database, a specifying means which specifies the word not having another words to maximize the similarity afterward among the words maximizing the similarity determined in the collation section and an output means which outputs the word specified by the specifying means as the recognized word.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a speech recognition device for recognizing an input speech and outputting a recognition result.

[0002]

2. Description of the Related Art Conventionally, a voice input through a microphone or the like is recognized, and the recognized content is output to a display means such as a display, or is output as a destination information to a navigation system mounted on an automobile. Speech recognition devices are known.

[0003] As a speech recognition method, a specific speaker recognition method in which words uttered by a speaker are registered in advance and a recognition word dictionary is created, and a word uttered by the speaker is registered instead of a specific speaker recognition method.
There is an unspecified speaker recognition method for creating a recognition word dictionary from a text document or the like. In each of these methods, the input voice is recognized by the speaker uttering a predetermined word. In a conventional speech recognition device, the most similar word is selected from words registered in a word dictionary based on a speech input within a certain period after detecting a speech by a speaker.

[0004]

However, in the above-mentioned conventional speech recognition apparatus, the speaker utters unnecessary words irrelevant to the recognition, such as "er" or "ano", and then utters words registered in the dictionary. In this case, words registered in the dictionary cannot be correctly recognized.

[0005] A word spotting method is known as one method of speech recognition that solves this inconvenience. The word spotting method is to sequentially collate words in the standard dictionary in a certain time unit, output words whose similarity exceeds a predetermined threshold, and continue collation otherwise. is there. Examples of the speech recognition method using the word spotting method include, for example, a “word speech recognition method using a duration control type state transition model” (Transactions of the Institute of Electronics, Information and Communication Engineers, vol. J72-D-II, No. .11, pp.1769-p
p. 1777, November 1989). This speech recognition method obtains good recognition performance while reducing the amount of computation by adding duration information to phonemes included in a dictionary to be recognized.

In the word spotting method described above, even if the speaker utters an unnecessary word such as "Eh" or "Ah" unrelated to recognition and then utters a word registered in the dictionary, the necessary word can be accurately detected. Can be recognized. However, since the collation is always performed with words in the dictionary irrespective of the presence / absence of the utterance of the person, there is a problem that a so-called swelling phenomenon occurs in which the utterance sound of the third person other than the person is recognized.

An object of the present invention is to provide a speech recognition device capable of performing speech recognition more accurately.

[0008]

According to a first aspect of the present invention, there is provided a speech recognition apparatus comprising: a word dictionary database for speech recognition; a speech input unit for calculating a speech feature amount from inputted speech; A matching unit for calculating the amount and the similarity with each word stored in the dictionary database; and, among the words having the largest similarity determined by the matching unit, other words having the largest similarity later It is characterized by comprising a specifying means for specifying a word having no word, and an output means for outputting the word specified by the specifying means as a recognized word.

[0009] The second speech recognition device of the present invention comprises the first speech recognition device.
The word recognition means, wherein the similarity calculated by the calculation means is equal to or greater than a predetermined threshold and the other words whose similarity is equal to or greater than a predetermined threshold And a word having no maximum word.

A third speech recognition apparatus according to the present invention is any one of the above speech recognition apparatuses, further comprising a sound volume measuring means for measuring a sound volume of a speaker, and the output means being specified by the word specifying means. And output means for outputting, as a recognized word, words for which the measured value of the sound volume measuring means exceeds a predetermined threshold value.

[0011]

BRIEF DESCRIPTION OF THE DRAWINGS FIG. 1 is a block diagram showing a speech recognition apparatus according to an embodiment of the present invention. FIG.
1 is a configuration diagram of the voice recognition device 100. The voice recognition device 100 includes a central processing unit (hereinafter referred to as a CPU) 1
, A microphone 3 for collecting voice, an A / D converter 2 for converting an analog signal collected by the microphone 3 into a digital signal for recognition processing, a ROM 4 storing a voice recognition processing program, a voice recognition process And a RAM 5 on which the above-mentioned program is developed at the time of execution of the program, an output device 6 such as a display for outputting the result of the speech recognition processing, and a standard dictionary database 7 for predetermined words. Note that a configuration is also conceivable in which a navigation system mounted on an automobile or the like is connected instead of the output device 6, and the result of voice recognition is output as destination information of the navigation system.

FIG. 2 is a diagram showing processing blocks of a speech recognition process executed by the CPU 1. The sound picked up by the microphone 3 is converted into a digital signal in the A / D converter 2 and then converted into a digital signal.
1 is output. The feature extraction unit 50 extracts a speech feature amount necessary for speech recognition from the input speech. Specifically, a predetermined time unit called a frame (for example, 20 m
s) A 10th order Meru cepstrum is determined every time.

The standard dictionary database 7 stores, for each word,
Average melu-cepstral vector for each phoneme generated from a phoneme string (character string) of a word, duration of each phoneme,
Further, information such as an automaton indicating a state transition of each phoneme is stored. For example, as described above, when the speech recognition result is used as destination information of the navigation system, the standard dictionary database 7 includes words such as place names and building names.

The collating unit 52 includes the information of each word stored in the standard dictionary database 7 and the feature extracting unit 5.
0, a distance vector (hereinafter referred to as similarity) S is determined while judging the state transition of the phoneme of each word, and the calculated similarity S of each word is determined. The data is output to the result determination output unit 54 together with the identification number (for example, JIS code).

On the other hand, the volume detecting section 51 detects the frame (20
Every ms), the maximum volume V of the input voice is obtained, and the obtained volume value V is output to the result determination output unit 54.

FIG. 3 is a diagram showing the transition between the similarity S and the maximum volume value V when the speaker himself speaks, and the transition between the similarity S and the maximum volume value V when a third person speaks. . As shown in the figure, the maximum sound volume value V generally becomes large when the voice is consciously uttered. In the case of unconsciously uttered voice, ambient noise, or utterance by a third party other than the speaker, the maximum sound volume value V is relatively small.

In the result judgment output unit 54, the word having the maximum similarity (the word that has started to fall after the similarity has increased)
Is detected, and it is checked whether or not there is another word having the maximum similarity between the time when the predetermined response time T _th elapses after the similarity of the word becomes maximum. If no other word having the maximum similarity is detected during the response time _Tth , the volume maximum value V at this time is set to the predetermined threshold value _Vth.
The word having the maximum similarity is output as a recognized word on condition that the number of words exceeds the threshold.

Meanwhile, if prior to the expiration of the response time T _th word in which the maximum similarity to the other is detected, until again the response time from the time of detecting T _th of the word has passed , And whether there is any other word having the highest similarity.

Even when no other word having the maximum similarity is detected during the response time _Tth , if the maximum sound volume value V does not exceed the predetermined threshold value _Vth , Then, after determining that the word is a word uttered and spouted by a third party and initializing the measured values of the similarity and the volume, the above process is repeated.

FIG. 4 is a flowchart of the voice recognition process executed by the CPU 1. FIG. 5 is a diagram showing the transition of the degree of similarity between the two words “Shinyokohama” and “Shinyokohama” in the result determination output section 54. Hereinafter, the procedure of the voice recognition processing will be described with reference to FIG.

First, the values of the similarity and the maximum volume of the voice input for each frame (20 ms) are initialized (step S1). A voice input is accepted (step S2).
The feature amount is extracted (step S3). A similarity S is obtained for each word stored in the standard dictionary database 7 (step S4). If there is a word whose similarity S has reached the maximum _Smax , for example, "Shinyokohama" as shown in FIG. 5 (YES in step S5), the response time timer is initialized (step S6). Here, after confirming that the similarity S _{ma x} exceeds the threshold value S _th (YES at step S7), and starts a timer, or to already continue operation if you started (step S
8). If the similarity S _max does not exceed the threshold value S _th (NO in step S7), the process returns to step S2, and the next word having the maximum similarity is detected.

Although the similarity S _max exceeds the threshold value S _th (YES in step S7), the response time timer value t
_Does not exceed the threshold value _Tth (step S9).
NO), the process returns to step S2, and another word having a maximum similarity is detected. Here, before the response time timer value t exceeds the threshold value T _th , another word having a similarity S of the maximum S _max , for example, “Shinyokohamakita” as shown in FIG. 5 is detected. (Yes in step S5)
ES), the response time timer is initialized again (step S6). The value t of the response time timer is set to the threshold value T _th without detecting any other word having the maximum similarity S at the maximum S _max.
Has elapsed (YES in step S9), the condition is determined on the condition that the maximum volume value V exceeds a predetermined threshold value _Vth (YES in step S10).
In the example of (1), “Shinyokohama Kita” is determined (step S1
1) This is output (step S12).

[0023]

According to the first speech recognition apparatus of the present invention, when a word having the highest similarity is detected, another word having the highest similarity later is immediately recognized from the words that have been recognized. Check for any. By performing the processing, for example,
"Shinyokohama" and "Shinyokohama" can be accurately recognized.

According to a second speech recognition apparatus of the present invention, in the first speech recognition apparatus, a word having a similarity greater than a predetermined threshold is output. Thereby, more accurate recognition processing can be realized.

Also, in the third speech recognition device of the present invention,
In any of the above speech recognition devices, in addition to the degree of similarity, it is determined whether the word is a word spoken by the speaker based on the volume of the input voice or a word spoken by a third party who is in the vicinity. Do. Thereby, erroneous recognition due to the source can be prevented.

[Brief description of the drawings]

FIG. 1 is a configuration diagram of an image processing system according to a first embodiment.

FIG. 2 is a system configuration diagram.

FIG. 3 shows a similarity S and a maximum sound volume V for a certain word.
FIG.

FIG. 4 is a flowchart of a voice recognition process executed by a CPU.

FIG. 5 is a diagram illustrating a state of a voice recognition process performed on two words “Shinyokohama” and “Shinyokohama”.

[Explanation of symbols]

1 CPU, 2 A / D converter, 3 microphone, 4 RO
M, 5 RAM, 6 output devices, 7 word dictionary database, 50 feature extraction unit, 51 volume detection unit, 52 collation unit, 53 standard dictionary, 54 result judgment output unit.

Claims (3)

[Claims] 1. A word dictionary database for voice recognition, a voice input unit for calculating a voice feature amount from an input voice, a voice input unit for calculating the voice feature amount, and each word stored in the dictionary database. A matching unit that determines the similarity, a specifying unit that specifies a word that has no other word that has the highest similarity later, and a specifying unit that specifies a word that has the highest similarity calculated by the matching unit. And an output unit that outputs the recognized word as a recognized word. 2. The method according to claim 1, wherein the identifying unit determines that the similarity obtained by the matching unit is equal to or more than a predetermined threshold value and is the maximum, and then the similarity is equal to or more than the predetermined threshold value and is the maximum. The speech recognition device according to claim 1, wherein a word having no word is specified. 3. The voice recognition device according to claim 1, further comprising: a sound volume measuring unit that measures a sound volume of a speaker, wherein the output unit outputs a word of the word specified by the specifying unit. Output means for outputting, as a recognition word, a word whose measured value by the volume measurement means exceeds a predetermined threshold value.