JPH0432900A - Sound recognizing device - Google Patents

Abstract

PURPOSE:To prevent the generation of a failure in the input of the same vocalization repeated several times by reducing the values of two thresholds when the 1st candidate of the preceding vocalization coincides with that of the current vocalization. CONSTITUTION:When the 1st candidate of the current vocalization is equal to that of the preceding vocalization in the sound recognizing device provided with a collating part 4 for collating a sound pattern with a registered sound pattern and calculating the 1st candidate and its recognition reliability, a recognized result deciding part 6 and a candidate storage part 5, the threshold of recognition decision is reduced to decide a recognized result. Consequently, an effective recognized result for a revocalized input can be obtained.

Description

DETAILED DESCRIPTION OF THE INVENTION Yokoyoshi M - The present invention relates to a speech recognition device, and more particularly, to a method for determining recognition results in a speech recognition device.

Voice recognition is widely used for voice control of machines and devices, voice data input, etc. In this case, erroneous voice recognition manifests itself as an error in control input or data input, and the effects thereof are extremely serious. In particular, when instructing the operation of a machine by voice input, it is necessary to prevent abnormal operation due to erroneous recognition. In such cases, conventionally, if the action instruction content of the recognition result of the first candidate is dangerous, the candidate with harmless instruction content is L&referenced based on the instruction content of the higher ranking candidate other than the first candidate. or disable voice input to prevent abnormal operation of one machine,
Alternatively, if the first candidate is output as a recognition result and the machine is instructed to operate, if there is a possibility that the machine will perform an abnormal operation, the user may be given instructions based on the recognition reliability of the first candidate. They asked for confirmation and invalidated the number one candidate.

Therefore, in the conventional technology as described above, even if the input speech is compared and the first candidate is the correct result, if the recognition reliability is smaller than the second threshold, the recognition result is invalid. The user will repeat the same utterance.

At this time, if the cause of the low recognition reliability is sudden noise or temporary voice modulation (coughing, voice overturning, etc.), the second utterance may produce an effective recognition result. is likely to be obtained. However, if the cause is constant ambient noise or changes in voice over time, effective recognition results will not be obtained no matter how many times the same utterance is repeated.

The present invention was made in view of the above-mentioned circumstances.
In particular, the purpose of this invention is to provide a speech recognition device that can obtain effective recognition results for inputs that are reworded while preventing abnormal machine operations.

In order to achieve the above object, the present invention includes a collation unit that collates an input voice pattern with registered voice patterns and calculates the first candidate and the recognition reliability of the candidate. , when the recognition reliability is larger than the first threshold, the first candidate is set as the recognition result, and when the recognition reliability is larger than the second threshold and smaller than the first threshold, the first candidate is set as the recognition result. a recognition result determination unit that displays a display prompting confirmation of the first-ranked candidate and does not select the first-ranked candidate as a recognition result if the recognition reliability is smaller than a second threshold; a candidate storage section for storing the first candidate memory of the current utterance;
If the candidate for the first position is equal to the candidate for the first position of the previous utterance, one or both of the first and second thresholds are lowered to determine the recognition result. Hereinafter, the present invention will be explained based on examples.

FIG. 1 is a block diagram for explaining one embodiment of the present invention, in which 1 is a voice input section, 2 is a voice pattern conversion section, and 3 is a block diagram for explaining an embodiment of the present invention.
is a standard pattern storage section, 4 is a matching section, 5 is a candidate storage section,
Reference numeral 6 denotes a recognition result determination section, in which a voice signal inputted from a voice input section 1 such as a microphone is converted into a voice pattern by a voice pattern conversion section 2 . Various conversion methods are known that are effective for speech recognition, and for example, one may be used that extracts the outputs of a bandpass filter group of 15 channels with a frame period of 10 m5. In the standard pattern storage unit 3, the speech of the recognition target word is converted into a voice pattern and stored as a standard pattern in advance. In the matching unit 4, the input voice pattern X and the standard pattern Y1,
Y3,...YM (M is.

(Number of registered standard patterns). For example, DP matching may be used as the matching method. for example,
If Yj has the smallest distance from X among all standard patterns, then the word name, command name, etc. corresponding to Yj will be the first candidate.

If this distance is Dl, the recognition reliability may be set to 1/D1. Or normalization coefficient Z set for each word
Using j, it is sufficient to set L=Zj/D□.

FIG. 2 is a flowchart for explaining the operation of the recognition result determination section 6. In the recognition result determination section 6, two
Using the threshold value T iJ t T z j of the type, L>T
In the case of process j, the first candidate is output as the recognition result (however, in the flowchart in Fig. 2, L: recognition reliability, j: first candidate, j-old: previous first candidate). candidate, Txj+ T2: J: Threshold value set for each word, α□β: Positive constant) 6T2j<L<
In the case of TyuJ, a display prompting the user to confirm whether or not to output the first candidate as a recognition result is displayed, and the first candidate is displayed only if the user is instructed to output the first candidate as a recognition result. 1
The position candidates are output as recognition results. As a method of giving the instruction, a method of inputting yes/no by voice or a method of assuming that an OK instruction has been given if the cancel button is not pressed within 3 seconds may be used.

In the case of L < T 2 J, the first candidate is rejected, the recognition result is invalidated, and a state of waiting for the next voice is entered. The candidate storage unit 5 stores the first candidate, and when the next voice is compared, the word name or command name corresponding to Yj in the first candidate is the same as the one stored in the candidate storage unit 5. In this case, Tlj and T2j are lowered and the recognition result determination unit 6
Processing is performed.

The change of T-work J + Tzz is as follows: T-work j = T1j - αL' (
1) T2:J = T2j-βL'
(2), where α and β are positive constants, and L' is the previous recognition reliability.

As is clear from the above explanation, according to the present invention, when the previous first-place candidate and the current first-place candidate are equal, the values of the two thresholds T 1jt T 2 j is made smaller. Therefore, if the first-ranked candidates from last time and this time are the same and there is a high possibility that this candidate is the correct recognition result, there is a high probability that this candidate will be output as is even if the current recognition reliability is low. This solves the problem of not being able to input even if you repeat the same utterance many times.

[Brief explanation of drawings]

FIG. 1 is a block diagram for explaining an embodiment of the speech recognition apparatus according to the present invention, and FIG. 2 is a flowchart for explaining the operation of the recognition result determination section shown in FIG. 1. 1... Audio input section, 2... Audio pattern conversion section, 3
Standard pattern storage section, 4... Collation section, Death... Candidate storage section, 6. Recognition result determination section. Figure 2 Figure 1 T2J=T1j-βT2 recognition result recognition result invalid

Claims (1)

[Claims] 1. A collation unit that collates the input voice pattern with the registered voice pattern and calculates the first candidate and the recognition reliability of the candidate, and if the recognition reliability is greater than the first threshold, , the first candidate is the recognition result, and if the recognition reliability is larger than the second threshold and smaller than the first threshold,
A recognition result determination unit that displays a display prompting confirmation of the first-ranked candidate and does not select the first-ranked candidate as a recognition result if the recognition reliability is smaller than a second threshold; In the speech recognition device, when the first candidate of the current utterance is equal to the first candidate of the previous utterance, one of the first and second thresholds is Alternatively, a speech recognition device is characterized in that the recognition result is determined by lowering both.