JPS60250400A - 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 the Invention The present invention relates to a speech recognition device that mainly recognizes registered words for a specific speaker.

Conventional configuration and its problems In a speech recognition device that recognizes registered words for a specific speaker, a time series of feature vectors of registered words is stored in memory as a standard button in advance, and the input speech button is used as a standard button. Performs batan matching with each standard batan,
A method is generally used in which the recognition result is the one with the highest degree of similarity.

However, even if the same speaker utters the same word, the spectral pattern changes over time, so if the above speech recognition device is used for a long time, recognition performance will deteriorate. This has the problem of causing the problem.

OBJECT OF THE INVENTION The present invention solves the above-mentioned conventional problems by automatically changing the input voice button to a standard button when the recognition processing result is sufficiently reliable. The purpose of the present invention is to provide a speech recognition device that can respond to fluctuations.

Structure of the Invention The present invention includes a voice analysis means for extracting voice feature vectors, an energy peak detection means for detecting main peaks on the voice energy envelope, and a time series of registered voice feature vectors that is stored as a standard beat. By matching the registered voice button with the input voice button,
It has a recognition means for deriving recognition candidate speech, and a speech length comparison means for comparing the speech length of the input speech and the recognition candidate speech,
When the recognition candidate speech is sufficiently accurate according to the results of the recognition means, energy peak detection means, and speech length comparison means,
This is a voice recognition device that uses the input voice button as a standard button and replaces a part of the standard button.

DESCRIPTION OF THE EMBODIMENTS FIG. 1 is a block diagram showing a speech recognition device according to an embodiment of the present invention. In FIG. 1, reference numeral 1 denotes a voice input section, into which voice from a speaker is input through a microphone or the like. 2 is a voice analysis means that extracts a feature vector from the input voice signal.

Reference numeral 3 denotes an energy peak detection means that detects major peaks on the energy envelope of the input voice.

Reference numeral 4 denotes a recognition means that performs bang matching between the standard bangs stored in the storage means 7 and the input voice bangs. 5
Reference numeral 6 denotes a time length comparing means for comparing the time lengths of the input speech and the recognition candidate speech obtained by the recognition means 4, and 6 is a bang exchanging means for exchanging the input speech bang and the standard bang.

FIG. 2 is a circuit diagram showing the configuration of this embodiment.

10 is a microcomputer for realizing each of the energy peak detection means 3, the recognition means 4, the voice length comparison means 5, the pattern exchange means 6, and the storage means 7 shown in FIG. A storage unit 12 that stores the time series of parameters and the peak number of each word obtained by the energy peak detection means 3, and an arithmetic control unit that performs matching of the input voice and the standard button, and determines whether or not to replace the standard button. 13, and input section 11,
The output unit 14 is configured isometrically.

8 is a microphone for inputting audio, and 9 is an analog-to-digital converter (hereinafter referred to as an A/D converter) for converting the audio signal input from the microphone 8 into analog-to-digital conversion and extracting characteristic parameters. 15 is a switch for starting the recognition registration process, and 16 is a recognition result display for displaying recognition candidate voices.

FIG. 3 is a flow chart of essential parts for explaining the operation of the microcomputer of this embodiment.

The operation of this embodiment with the above configuration will be described in detail for the case where three standard buttons for each registered word are prepared.

The voice recognition device according to this embodiment first enters a voice input waiting state by pressing the switch 15. When audio is input, in step 17, audio signal input processing is performed. This is done by the A/D converter 9, and the analog-to-digital converted audio signal is sent to the microcomputer 1o.
is inputted to , the arithmetic control unit 13 extracts a feature vector,
The time series of this feature vector is stored in the storage unit 12. Thereafter, in step 18, main peaks on the energy envelope of the input voice are detected and the number of peaks is stored in the storage unit 12. In the detection of the number of energy peaks in step 18, in order to detect peaks with high energy in the vowel part as much as possible, the sum of feature vectors on the low frequency side (for example, below IKHz) is regarded as the energy value, and the sum of the feature vectors on the energy envelope is Dip and peak parts are first detected.

When the difference in energy between the dip portion and the peak portion is equal to or greater than a threshold value, this is regarded as one peak, and the number of peaks on the entire energy envelope is counted. When this peak detection is completed, in step 19, the storage unit 1
2, the standard drums stored in 1.2 are matched with the input voice drums, and the 1st and 1st. Select the second candidate word.

In step 2o, it is checked whether the selected first candidate word and second candidate word are the same word, and if they are the same, the process proceeds to step 21, and the distance difference between the first candidate word, the second candidate word and the input speech is determined as a threshold value. If it is below, in step 22, recognition candidate words and the like are displayed on the recognition result display 16 as a recognition result. If the conditions in steps 20 and 21 are not satisfied, it is determined that the pattern matching result is not accurate and the process returns to step 17.

When the recognition result is obtained, in order to adopt the current input voice pattern as the standard button, it is determined whether the input voice is the same as the recognition result by the process described below.

First, in step 23, the number of energy peaks of the input voice and the first candidate are compared, and if they are the same, the number of peaks of the input voice and the second candidate are compared in step 24. If the number of peaks is the same, it is assumed that the energy pattern of the input speech is very similar to that of the recognition result, and the process proceeds to the next step.

In step 25, the input speech and the first candidate are compared in speech length. This calculates the voice length ratio of the input voice and the first candidate, and checks whether D is within the range of 1-a(D(1+a(0(a(1)).If this condition is met. Then, in step 26, the input speech and the second candidate's speech length are compared in the same way as in step 25.Here, a is
O is the threshold for voice length comparison. When all these conditions are met, the input voice is
It is determined that the word is the same as the recognized word, and the standard button is replaced in step 27. In this standard baton replacement, the first. The remaining standard drums other than the two second candidate standard drums are replaced with the drums of the input voice and stored in the storage unit 12 again.

In this way, the standard button is changed to a new button one by one. If the conditional statements in steps 23, 24, 25, and 26 are not satisfied, the process returns to step 17 and waits for voice input.

According to the configuration of the above embodiment, when the recognition result is determined to be accurate by using the results of bang matching, energy peak information, and voice length comparison results, the standard baton is changed to a new baton. By doing so, it is possible to respond to temporal fluctuations in the speaker's utterances.

Effects of the Invention The present invention provides an energy peak detection means for detecting a peak on the energy envelope of an input voice, and a voice length comparison means for comparing the time length of a human voice and a standard bang as a recognition result. As a result of slam matching, the first. When the second candidates are the same, the distance difference between the input voices is small, the number of energy peaks all match, and the length of each voice is within a certain range, the recognition result is judged to be reliable, and a part of the standard baton is By replacing the input speech pattern with the input speech pattern, it is possible to provide a speech recognition device that can update the standard stamp in response to changes in the speaker's utterance over time.

Furthermore, since the standard button is automatically updated, there is no need to re-register each time, and it is possible to provide a speech recognition device that can be used for a long time.

[Brief explanation of the drawing]

FIG. 1 is a block diagram of a speech recognition device according to an embodiment of the present invention, FIG. 2 is a circuit diagram showing the configuration of the same device, and FIG. 3 is a flowchart of essential parts for explaining the operation. 2... Voice analysis means, 3... Energy peak detection means, 4... Recognition means, 5... Time length comparison means, 6... Bang replacement means, 7...
...Storage means, 8...Microphone, 9...
...A/D converter, 10...Microcomputer. Name of agent: Patent attorney Toshio Nakao and 1 other person No. 1
Figure 2

Claims (1)

[Claims] a speech analysis means for extracting a % signature vector from input speech;
energy peak detection means for deriving an energy envelope from the feature vector YI obtained by the voice analysis means and detecting the number of main peaks on the envelope; and storing a plurality of time series of feature vectors of each registered voice as standard beats. a recognition means for deriving a recognition candidate voice by matching a standard bang of the registered voice stored by the storage means with the input voice, and comparing the voice lengths of the input voice and the recognition candidate voice. As a result of bang matching by the speech length comparison means and the recognition means, the top two candidates are the same word and sufficiently similar to the input speech, and the number of peaks determined by the energy peak detection means match both the input speech and the recognition candidate speech. , and further comprising a button replacement means for replacing one of the standard clicks other than the top two candidates of recognition candidate voices with an input voice button when the voice lengths of both voices are within a certain range as a result of the voice length comparison means. A voice recognition device featuring: