JPS6347797A - Word voice preselection system - Google Patents

Abstract

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

Description

DETAILED DESCRIPTION OF THE INVENTION Technical Field The present invention relates to a method for preselecting word sounds in a speech recognition device.

Conventionally, as a preliminary selection method for word sounds, a comparison is made with the same weight in a standard pattern (word dictionary) based on a certain feature extracted from unknown input speech, and it is compared with the same weight in a standard pattern (word dictionary). The usual method is to select a pattern. However, with these methods, there may be variations in the number of recognition targets narrowed down by preliminary selection, and preliminary selection methods based on certain criteria may have problems with the nature of the input word (e.g.
In some cases (when there are many similar words in the standard pattern), inconveniences often occur, such as not being selected as a candidate (that is, being rejected).

Purpose The present invention was made in view of the above-mentioned circumstances.
In particular, when recognizing large word sounds with a large number of recognition targets, in a word preliminary selection method that limits the number of candidate words for the purpose of shortening the recognition processing time, the number of candidates is reduced to a certain value or less. Make m selections in advance using the preselection method of .

Eliminating variations in recognition processing time and increasing the processing speed ratio by '4'
It was done for the purpose of being happy.

1-1 In order to achieve the above-mentioned object, the present invention provides a microphone for inputting voice, a characteristic part prayer unit for determining a characteristic time sequence in voice, and a voice recognition device for common words. A preliminary selection unit that selects word candidates, a dictionary for preliminary selection that is checked when performing preliminary selection, a recognition processing unit that recognizes candidate words narrowed down by the preliminary selection unit, and words that are referred to during recognition processing. It consists of a standard pattern storage unit and a recognition result output terminal that outputs recognition results, and is characterized by performing preliminary selection using a plurality of different preliminary selection methods until the number of candidate words falls below a predetermined value. That is. The following will explain based on five embodiments of the present invention.

FIG. 1 is an electrical block diagram for explaining one embodiment of the present invention, in which 1 is a voice input microphone;
2 is a voice feature extraction unit, 3 is a word preliminary selection unit, 4 is a preliminary selection dictionary storage unit, 5 is a recognition processing unit, 6 is a standard pattern (dictionary) storage unit, 7 is a recognition result output unit, and 7 is a recognition result output unit. The voice inputted from 1 is converted by the feature extractor 2 into a time series of characteristic parameters specific to the voice. Preliminary selection section 3
Now, using this series of feature parameters, matching is performed with a preliminary selection dictionary 4 created in advance from the human language award word dictionary to narrow down candidate words. In the recognition processing section 5, the words narrowed down by the preliminary selection section 3 are matched with the word standard patterns in the dictionary storage section 6, and the pattern name closest to the input cover pattern is outputted to the recognition result output section 7 as a recognition result.

FIG. 2 is a flowchart showing the contents of the process in the preliminary selection section 3 shown in FIG.
2, 34, and 35 are M different preliminary selection method units, 33 is a determination processing unit that determines the end of processing, and 36
is the processing end terminal. As a preliminary selection process, first,
A preliminary selection method unit 32 selects candidate words based on certain criteria. The selection method involves extracting characteristic parameters from unknown input speech and selecting a standard pattern having the same or similar characteristics. At this time, the number of selected candidate patterns (number of words) is set to .

On the other hand, the number of candidate patterns to be finally narrowed down is set in advance to N. It goes without saying that the value of N depends on the number of words to be recognized and the performance of the preliminary selection method, but the value of N in particular is determined empirically according to the performance of the desired recognition system. There are many things.

In this manner, when the number of candidates K becomes equal to or less than N through the processing of the preliminary selection method section 32, the preliminary selection process is terminated. However, normally, when the number of recognition targets is large or when N is small compared to the number of recognition targets, it is rare that a single preliminary selection method satisfies the determination condition (K≦N) of the determination processing unit 33. It is. In such a case, it is necessary to further narrow down the candidate words using a second preliminary selection method section 34 different from the preliminary selection method section 32. Different preliminary selection methods for 32, 34, and 35 include, for example, comparing the duration of speech, the number of silent intervals, etc., classifying the phonemes of the initial and middle parts of words, the presence or absence of power dips, and other special selection methods from speech patterns. Various methods can be used, such as a method based on patterns extracted by various processes, and these methods can also be used in any order. Therefore, if the process flow shown in FIG. 2 is followed, the preliminary selection can be terminated when the number of candidates sent to the recognition processing unit 5 is always smaller than N, and the unknown input It is possible to suppress variations in the number of recognition target candidates due to differences in voice characteristics and to suppress unnecessary preliminary selection processing.

Effects As is clear from the above explanation, according to the present invention, in advance, the number of candidate patterns to be finally narrowed down is reached.
Since different preliminary selection methods are used in series and judgment processing is performed at the end of each process, it is possible to limit the increase in time required for recognition processing to within a certain amount of time, and to achieve efficient preliminary selection. Selection processing and recognition processing become possible.

[Brief explanation of the drawing]

FIG. 1 is an electrical block diagram for explaining one embodiment of the present invention, and FIG. 2 is a flowchart for explaining the processing contents of the preliminary selection section 3 shown in FIG. DESCRIPTION OF SYMBOLS 1... Microphone, 2... Feature extraction part, 3... Preliminary selection part, 4... Preliminary selection dictionary storage part, 5... Recognition processing part. 6... Standard pattern, 7... Recognition result output section, 32
゜34.35... Preliminary selection section, 33... Judgment section. Figure 1 Figure 2

Claims (1)

[Claims] A microphone for inputting speech, a feature section for finding characteristic time sequences in speech, a preliminary selection section for selecting word candidates prior to recognition of large vocabulary word speech, and a pre-selection section for making preliminary selections. It consists of a preliminary selection dictionary for comparison, a recognition processing section that recognizes candidate words narrowed down by the preliminary selection section, a word standard pattern storage section for reference during recognition processing, and a recognition result output terminal that outputs recognition results. A word-sound preliminary selection method, characterized in that preliminary selection is performed using a plurality of different preliminary selection methods until the number of candidate words becomes equal to or less than a predetermined constant value.