JPH0570837B2 - - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION (Field of Industrial Application) The present invention relates to a speech section detection device for detecting the range of input speech used in speech recognition technology.

(Prior Art) Various methods have been tried as conventional voice segment detection devices. Among these methods, a typical method is to set thresholds for characteristic parameters such as power and number of zero crossings in advance, and then set a threshold value such as the power of the input signal, the number of zero crossings, or the energy calculated from the power within a unit time. There is a method that detects a voice section based on whether or not it exceeds the range.

The principle of a conventional voice detection device will be explained below with reference to the drawings.

FIG. 1 is a block diagram showing the principle of a conventional voice detection device.

The voice input from the microphone 1 is converted into a time series of characteristic parameters including power in the voice parameter extraction section 2.

Next, detection is performed in the voice section detection section 4 according to a threshold value set in the threshold value storage section 3 in advance. This method is performed by comparing the characteristic parameters of the input speech with a fixed threshold or a threshold value adapted to the noise level at each time. The processing in the voice detection section 4 is described in, for example, Japanese Patent Application No. 58-156098.
A method described in the specification "Variable Threshold Type Speech Detector" is known.

This method calculates the average power of the noise signal in the input signal, sets a threshold value for power, power accumulation, etc. from that value, and then compares it with the input signal.
The starting end and ending end are determined from the results for each threshold value.

After detecting the voice section, if the input voice is a registered pattern, switch S1 to A and then store it in the registered pattern buffer memory 5. If the input voice is a recognized pattern, switch S1 to B and then store it in the recognition pattern buffer memory 6.
Each pattern is memorized.

(Problems with the prior art) However, in the speech segment detection device as described above, the threshold value for detecting speech segments does not adapt to individual differences in speech. False positives may occur. This means that a recognition error occurs during recognition. Especially when such false positives occur during registration,
No matter how good the recognition method is, a problem will arise in that recognition errors will increase significantly.

(Object of the Invention) An object of the present invention is to provide a speech section detection device that accurately detects speech sections of registered patterns and has a function of learning an optimal threshold value for each speaker.

(Structure of the Invention) The speech section detection device according to the present invention requires the following parts. That is, a voice parameter extraction unit extracts a parameter time series including power from input voice, a pattern buffer memory stores the voice parameters obtained by the voice parameter extraction unit, and a pattern buffer memory for each input registered pattern. A rule storage unit that records predetermined rules; an optimal threshold learning unit that learns an optimal threshold based on registered patterns in the pattern buffer memory and rules in the rule storage unit; and an optimal threshold learning unit that uses the optimal threshold obtained therein. a voice section detecting section that detects the registered pattern in the pattern buffer memory and the voice section of the recognition pattern obtained by the voice parameter extracting section; and storing the registered pattern voice detected by the voice section detecting section. These sections include a registered pattern buffer memory and a recognition pattern buffer memory that stores recognition patterns voice-detected by the voice section detecting section.

(Operation/Principle of the Present Invention) The principle of the present invention can be considered in the following three steps. First, in the first step, an optimal threshold value is learned from a plurality of threshold values obtained for a plurality of registered patterns.

That is, a rule is given to the conditions under which voice segment detection can be accurately obtained for each registered pattern, and a threshold value that satisfies the rule is determined for each registered pattern. The plurality of threshold values obtained in this way are obtained for the registered pattern, and by obtaining the optimal threshold value from among these, it is possible to obtain a threshold value suitable for the voice of the speaker. In the second step, audio detection of the registered pattern is performed using the optimal threshold obtained in the first step. Here, since the optimal threshold value is obtained from the registered pattern, reliability of voice detection can be expected. In the third step, after detecting the voice of the registered pattern, voice detection is performed on the recognized pattern in the same manner as in the second step.

(Example) Examples of the present invention will be described in detail below with reference to the drawings.

FIG. 3 is a block diagram showing an embodiment of the speech section detection device of the present invention, which includes a microphone 1, a speech parameter extraction section 2, a pattern buffer memory 7, a rule storage section 8, an optimum threshold value learning section 9, and a speech section detection device. It consists of a detection section 4, a registered pattern buffer memory 5, and a recognized pattern buffer memory 6.

Here, the input voice is limited to numbers, and the registered patterns are ``zero'', ``ichi'', ``nii''... ``kyu''.
An example will be explained in which a total of 10 patterns are used each time, and one pattern of "Niisan" is used as a recognition pattern.

In the first step, after switching the switch S2 to A, the data "ZERO" to "QUE" inputted from the microphone 1 are converted into a time series of characteristic parameters including power in the audio parameter extraction section 2.

The parameters thus determined are stored in the pattern buffer memory 7. In the rule storage unit 8, conditions are defined so that the voice section can be reliably detected for each data of "Zero" to "Kyuu".

For example, for "Roku", E _ROKU <E _io ...(1) NP=2 ...(2) (P'(t)=P(t)-P(t-1) (t=2,
...T), and if P'(x)×P'(x+1)<0 and P'(x)>0, it takes the maximum value at time x). E _ROKU is an energy threshold predetermined for the pattern "ROKU", and E _io is an integral of power for a speech section temporarily detected for the input speech. Also, NP is the number of x that takes the maximum value, P(t) is the power at time t, T
is the time length of the input audio.

In other words, the condition here is ``if the integral of power is
It can be rephrased as "is larger than E _ROKU and has two maximum values."

When the second half of “Roku”, “ku”, is devoiced,
Because the power level is so low, the "ku" cannot be detected using standard audio detection levels. Therefore, in order to prevent such cases, the condition "having two maximum values" is given as a rule.

The next process is to find an optimal threshold value in the optimal threshold value learning section 9 from the rules for all registered patterns and the registered patterns. The configuration of the optimal threshold value learning section 9 is as shown in FIG. 2. The process will be explained below using the registered pattern "Roku" as an example. The threshold value stored in the temporary threshold value storage unit 91 is set as the initial value, and the input registered data "ROKU" is set as the initial value.
Using a rule for , the threshold determining unit 92 determines a threshold that satisfies the rule. Figure 2 2
2 shows how the threshold value determination unit 92 detects the sound of "Roku". L _S is a provisional threshold, L _F is a threshold that satisfies the rule, and the threshold is gradually lowered until the number of poles NP = 2.

The configuration of the threshold determining section 92 is as shown in FIG. 2.

The microprocessor 921 operates according to the flowchart shown in FIG.
A threshold value for L _F is determined.

In block A in FIG. 2, voice detection is performed by a conventional method, for example, the method described in the above-mentioned Japanese Patent Application No. 156098/1985, and detects the start end (T _S ) and the end end (T S ). _F ). Also L _V
is the threshold value that is lowered at once. The threshold value determined in this manner is stored in the threshold value storage section 93. Similarly, thresholds that satisfy the respective rules for all registered patterns are stored in the threshold storage section 93.

Once the threshold values have been determined for all registered data, the optimal threshold value determination unit 94 then determines the optimal threshold value.

Here, for example, it is possible to take the minimum value of the threshold values in the threshold value storage section 93. The optimum threshold value determination section 94 is constituted by a minimum value detection circuit, and the optimum threshold value obtained here is stored in the speech section detection section 4.

As explained above, the first step has a function of adapting the threshold value to the input registered data, and as compared to the conventional method, a threshold value more suitable for the speaker can be obtained.

In the second step, after switching the switches S3 and S4 to A, the voice section detecting section 4 detects the voice of the registered pattern using the optimum threshold value determined in the first step. Registered pattern buffer memory 7
Since the parameters of each registered data are already stored in the , it is possible to use them.

The processing in the detection unit 4 is similar to that described above.
It is possible to use the voice detector described in -156098.

In the third step, after switching S2, S3, and S4 to B, the speech section of the recognition pattern is detected.
The characteristic parameters of "Nii-san" inputted through the microphone 1 are extracted by the voice parameter extraction unit 2 in the same manner as at the time of registration.

Next, using the optimum threshold value determined in the first step, detection is performed in the voice section detecting section 4 in the same manner as in the second step.

Although one embodiment of the present invention has been described above, it is obvious that the data to be handled need not be limited to numbers and can be applied to any voice data. Furthermore, although an example of a rule has been given in the explanation, the description content and method of writing the rule do not change the essence of the present invention in any way in the sense that the rule is a condition for determining a threshold value. Therefore, it is included in the present invention.

Furthermore, it is not necessary to determine the optimal threshold value using all the threshold values of all the registered data, and it is also possible to determine the optimal threshold value using only the threshold value for one or more of the registered data.

Further, as a method for determining the optimal threshold value from a plurality of threshold values in the optimal threshold value determination unit 94, there are a method of taking an average value in addition to the minimum value, a method of taking a weighted average value, and the like.

(Effects of the Invention) As described above, the voice section detection device according to the present invention has the following advantages, even though the number of voices is not significantly different from that of the conventional method.

First of all, it can accommodate the individuality of vocalizations. In other words, in order to learn the optimal threshold from the registered pattern,
Compared to a fixed threshold type voice detection device, a threshold value that is more suited to the voice of the speaker can be used for detection.

Second, since the voice section detection of the registered pattern is performed again using the optimal threshold value determined from the threshold value for the registered pattern, it can be said that false detection of the registered pattern is reduced compared to the conventional method. This shows that recognition errors caused by poor quality registered patterns can be significantly reduced.

[Brief explanation of the drawing]

FIG. 1 is a block diagram showing a conventional speech interval detection device, FIG. 2 1 is a block diagram showing an optimal threshold value learning section 9, FIG. FIG. 3 is a circuit diagram of the threshold determining section 92, FIG. 2 is a flowchart of the threshold determining section 921, and FIG. 3 is a block diagram showing a voice section detection apparatus according to the present invention. In the figure, 1...Microphone, 2...Audio parameter extraction section, 3...Threshold value storage section, 4...Audio detection section, 5...Registered pattern buffer memory, 6...Recognition pattern buffer memory, 7... . . . Pattern buffer memory, 8 . , 921...Microprocessor, 922...
I/O port, 923...Memory, S1, S2,
S3, S4...Switch.

Claims (1)

[Claims] 1. A voice parameter extraction unit that extracts a parameter time series including power from input voice, a pattern buffer memory that stores the voice parameters obtained by the voice parameter extraction unit, and a predetermined judgment for each registered pattern. using a judgment condition storage unit that stores conditions, a threshold determination unit that determines a threshold that matches the registered pattern in the pattern buffer memory and the output of the judgment condition storage unit, and the threshold obtained in the threshold determination unit, a registered pattern in the pattern buffer memory and a voice section detecting section that detects the voice section of the recognition pattern obtained by the voice parameter extracting section; and a registered pattern that stores the registered pattern voice detected by the voice section detecting section. A speech segment detection device comprising: a buffer memory; and a recognition pattern buffer memory that also stores a recognition pattern detected by the speech segment detection section.