JPS6118199B2 - - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION The present invention relates to improvements in speech recognition devices.

Speech recognition devices, which are devices that automatically recognize speech uttered by humans, are considered to be very effective as a new means for inputting data and commands from humans to computers and various machines. For example, if you use a device that recognizes voice numbers and connects them to a computer, you can input numerical data such as slips, but voice signals can be easily transmitted via inexpensive telephone lines, so data can be sent from remote locations. This enables a new and useful form of use in which the user inputs the following information. In addition, by using a device that recognizes the various command words necessary to operate various industrial machines, it is possible to control the machine simply by uttering the words, and the hands, feet, and even eyes can be used for other purposes, allowing multiple commands to be used at the same time. This means that work can be carried out and higher work efficiency than ever before can be achieved.

However, the speech recognition devices that have been developed to date cannot operate normally and may cause erroneous recognition when ambient noise is present or when the speech is unclear. Therefore, if the purpose of use is for absolutely no errors, it is necessary to have a configuration that displays the judgment result and requests confirmation from the user once the user has finished speaking and the judgment result has been finalized. . In this case, the user repeats utterance and judgment result recognition one after another, and continues inputting the previous data in sequence unless there is an error. If an error is discovered during the confirmation stage, the user speaks the data again and inputs it. The following work procedure will be executed. In order to execute this work procedure at high speed, it is necessary for the recognition device to operate at high speed and quickly display the determination result so that the user can confirm it. The problem here is the fact that even if the utterance has actually ended, the recognition result cannot be confirmed for several hundred milliseconds. In other words, most recognition devices to date detect the end of input speech by testing the amplitude level of the speech, but when the amplitude level is instantaneously 0 (actually a sufficiently small value), ), it is not possible to determine that this point is the end of the input audio, and the end can only be determined after the amplitude level 0 continues for, for example, 250 milliseconds. If the moment when the amplitude level becomes 0 is immediately determined to be the end, the following inconvenience will occur. Now, if you say the number "6" (/roku/), there will normally be a pause between /ro/ and /ku/,
The amplitude level becomes 0 (hereinafter, this pause section will be referred to as a pause section within a word). Therefore, if we assume that the point when the power level becomes 0 is the end of a word, /ro/ will be treated as one word, and will probably be determined as "5" (/go/). In order to avoid such inconveniences, it is necessary to accurately determine whether the section in which the amplitude level reaches o, that is, the pause section, is a pause section within a word, or whether it is truly a pause section at the end of a word. Pauses within words are usually only about 200 milliseconds or less in length. Taking advantage of this fact, the pause interval of amplitude level o can be set to a predetermined length, e.g.
If it is 250 milliseconds or less, it is determined that it is a pause within a word, and if it exceeds 250 milliseconds, it is determined that it is a pause at the end of a word. Therefore, for example, 250 milliseconds after the end of the utterance, the end of the word is not determined, so the determination result cannot be displayed.

For the reasons mentioned above, users of conventional speech recognition devices cannot know the recognition result until they waste several hundred milliseconds after the utterance ends.
Also, the next data could not be uttered until after the recognition results were displayed and confirmed. For this reason, it has been impossible to input data at high speed with conventional speech recognition devices.

The present invention improves the above-mentioned drawbacks of conventional speech recognition devices, namely that the recognition result cannot be known until several hundred milliseconds have passed after the end of utterance, and therefore the data input speed is low. The object of the present invention is to realize a speech recognition device in which the judgment progress is displayed immediately after the utterance is finished, and the data input speed is high.

The speech recognition device according to the present invention has a function of generating a pause section detection signal when a pause section of input speech is detected, and generating an end detection signal when the duration of the pause section exceeds a predetermined threshold. a recognition section having a function of performing a recognition operation and promptly outputting a recognition result signal by assuming that the input voice has ended at the time when the pause section detection signal is generated; The recognition result display section includes a recognition result display section for displaying the recognition result, and an output section that allows the recognition result signal to pass through and be output at the time when the end detection signal is generated.

With such a configuration, the recognition result is displayed immediately after the utterance is finished, and confirmation can be performed quickly. Furthermore, since the recognition result is output after the pause period continues for a sufficiently long time and it is determined that it is the end, there is no possibility that an erroneous recognition result will be output during the pause section in a word.

FIG. 1 is a block diagram showing one embodiment of the present invention, and FIG. 2 is a time chart showing an example of its operation. In the figure, reference numeral 1 denotes a microphone that converts an input sound wave into an electrical signal and sends it to the analysis section 10 via a signal line i. The analysis section 10 has a function of analyzing the input audio waveform sent via the signal line i, extracting the recognition parameter a and the amplitude level l, and sending them to the recognition section 13, the start end detection section 11, and the end end detection section 12, respectively. has. The analysis section is, in short, a circuit that extracts parameters necessary for recognition from the input speech waveform, and can be configured with, for example, a bandpass filter analyzer, an autocorrelation analyzer, or a linear prediction analyzer. can. FIG. 3 shows an example of an analysis section mainly composed of a bandpass filter analyzer. The broken line block 110 is a well-known bandpass filter analyzer that analyzes the input signal i into, for example, 10 frequency bands and outputs the output amplitude of each channel as an analog signal of α ₁ to α ₁₀ . do. These signals are sent to analog multiplexer 12
0, and is further converted into a digital signal a by the A/D converter 130 and sent to the recognition section 13. On the other hand, the input signal i is directly applied to the rectifier circuit 140 and rectified, then smoothed by the low-pass filter 150, further converted into a digital signal by the A/D converter 160, and output as an amplitude level signal l. . The multiplexer 120 and the A/D converter 160 operate in synchronization with the pulses of the separately provided frame period signal Cl. The pulse interval of this frame period signal is set to 10 milliseconds, for example. Therefore, the output signal a from the analysis unit 10 is a time series of input vectors ai = (α _1i , α _2i , ..., α _xi , ..., α _10i ) (1), and the output signal l is l _i (2 ) is the time series of the amplitude signal. Here, α _xi is the analysis output of the xth channel at the time sample point, ie, the signal α _x (digitized).

The starting edge detection unit 11 is given as the signal l.
Test l _i in equation (2) to detect the start of the input voice,
At that point, a starting edge detection signal pulse S is generated. Various methods can be considered for starting edge detection, but one example is a predetermined threshold θ _b
The simplest method is to compare the above-mentioned l _i and define the point l=b at which l _i >θ _b (3) as the starting point. In any case, when the starting edge of the input audio is detected, a starting edge detection signal pulse S is generated. This pulse S is sent to the recognition section 13.

The termination detection unit 12 performs a detection operation for a pause section and an termination. The rest period is defined as a period in which the amplitude level is smaller than a predetermined threshold θ _e . Furthermore, the end point is defined as the time point i=e immediately before the pause section when the pause section continues for longer than a predetermined time length L. Fourth
The figure shows the definition of the pause section and the end. Fourth
The upper part of the figure shows how the amplitude level l _i changes over time when the word "3" (/san/) is uttered, for example. The time point after which l _i is smaller than a predetermined threshold θ _e , that is, the time point indicated by reference numeral 40, is the rest period,
If a pause exceeds a predetermined time length L, this pause is considered to mean the end of a word. In this case, the time point immediately before reference numeral 40, that is, the time point indicated by reference numeral 41, becomes the end point i=e of the input speech.

FIG. 5 is a diagram showing an example of the configuration of the termination detection section. The amplitude level signal l _i from the analysis section 51 is compared with a preset threshold value θ _e by the comparison circuit 51 . If l _i >θ _e , the output of the comparator circuit becomes y=0,=1. The former signal y is output as is as the rest period detection signal e2. Since the latter signal is given to the counting circuit 53 as a reset signal, as long as l _i >θ _e , the counting circuit 5
The output m of 3 is 0. The signal m output as the contents of this counter is compared with a preset threshold L by a comparison circuit 54. As an example, this threshold L is set to L=25, which is 250 milliseconds when counted at the frame period of 10 milliseconds.
The output e1 of the threshold circuit 54 is e1=1 only when m≧l
When m<l, e1=0.
Therefore, as long as l _i >θ _e, the termination detection signal e1
is 0. Enters the pause section at the end of the input audio.
At the point when _i < θ _L , the output of the comparison circuit 51 is y
=1, =0. Since the former signal is output as the rest period detection signal e2, e2=1.
Moreover, this signal y and the frame periodic signal pulse
The logical product Z with Cl is calculated by the logic circuit 52 and provided to the counting circuit 53. At this time, since the signal is 1, the pulses of the signal Z are counted by the counting circuit 53. Therefore, the content m of the counting circuit 53 continues to increase by being added by 1 in synchronization with the frame periodic signal pulse Cl while l _i <θ _e , that is, during the pause period. Thus, when the content of the numeral circuit 53 becomes equal to or greater than the constant L, the output of the comparison circuit 54, that is, the end detection signal e1 becomes 1.

While the end detection section is performing the operation of detecting the end of the input audio as described above, the recognition section 13 performs the following operations. The time point when the start edge detection signal S (pulse) from the start edge detection section 11 is given, i.e., i=
After b, the input vector a _i (formula (1)) input from the analysis unit 10 is received and a recognition operation is executed. After the start edge is detected, the input vectors are inputted one after another in synchronization with the pulses of the frame periodic signal Cl. Until the rest period detection signal e2 is applied, preprocessing for recognition is mainly performed, and no final determination is made. At the time when the pause section detection signal e2 is applied, the following input vector sequence is input and held in the recognition unit 13.

A=a _b , a _b+1 ,..., a _i ,..., a _e (4) Below, this input vector sequence is used as input pattern A
It is called. The recognition unit 13 regards this input pattern A as one input voice (that is, assumes that the input voice ends when the pause section detection signal e2 becomes 1) and makes a determination.

In reality, the operation during this period will take various forms depending on the recognition principle adopted by the recognition unit 13, but one of the features claimed in the present invention is an interval with a low amplitude level (a pause interval). ) is detected,
That is, when the pause section detection signal e2 is received, the determination is made immediately on the assumption that the input audio has been completed at that point. Various recognition principles can be considered for the recognition unit 13, and it is not intended to be specific to one type, but as an example, there is a One possible example is the DP matching method described in the paper titled "Comparison of Various DP Matching Methods." That is, the standard pattern C ⁿ = C ⁿ ₁ , C ⁿ ₂ , ..., C ⁿ _Jo (5) is built in for all necessary words n (n = 1 to N), and the above-mentioned pause section is When the detection signal becomes 1, DP matching is performed between the input pattern A and each standard pattern C ⁿ to calculate the similarity S (A, C ⁿ ) (6). Next, the input pattern A is determined to be the word r by comparing the similarity calculated with each standard pattern C ⁿ and determining the maximum word name n=r. The specific configuration of this recognition unit is not claimed as a right of the present invention, so it will be omitted.

The recognition result r determined by the recognition section 13 is sent to the output section 16 and the recognition result display control section 14 via the signal line r shown in FIG. The output section 16 is the termination detection section 1
It has a function of outputting the recognition result signal r as an output signal x only when the termination detection signal e1 from 2 is 1. Therefore, while the termination detection signal e1 is 0, no output x is generated. The recognition result display control section 14 sends the recognition result signal r to the recognition result display section 17 via the signal line d for display while the pause section detection signal e2 given by the end detection section 12 is 1. The recognition result display section 17 is constituted by, for example, a character display of a light emitting diode, and promptly informs the speaker of the recognition result r by displaying the characters specified by the signal line d.

The pause section continues and eventually the end detection signal e1 becomes 1.
Then, the output unit 16 outputs the recognition result signal r as the output signal x. This output signal x is sent to another device (hereinafter referred to as a target device) connected to and controlled by the above speech recognition device.

The input ready state display section 15 and the display lamp 18 are
This is to inform the speaker of when it is permissible to utter the input voice. A current is passed through the signal line t for which the end detection signal e1 is 1, and the indicator lamp 18 is turned on. When e1 is 0, no current is passed through the signal line t, and the indicator lamp 18 is turned off. therefore,
The indicator lamp 18 is turned off after the start of the audio is detected until the end is detected.
When the end is detected, the light is lit again to indicate that the next input voice may be uttered.

According to the above operations, the final recognition result is determined at the time point indicated by reference numeral 21 in FIG. 2, but is actually displayed at the time point indicated by reference numeral 20. The time from the time point 22 at which the pause section is detected to the time point 20 at which the recognition result is displayed can be made sufficiently short by speeding up the operation of the recognition unit 13, so that the speaker actually finishes speaking. You can know and check the recognition results almost at the same time. Moreover, since the recognition result is output to the target device at time point 21 after the pause has continued for a sufficiently long period of time, there is no possibility of misidentifying a pause within a word from a pause at the end of a word, as mentioned above. .

FIG. 6 is a time chart to clarify the operation when there is a pause section in a word. When you say the number “6” (/roku/), it becomes /o/
Since there is a pause section indicated by the reference numeral 61 between and /k/, the pause section detection signal e2 becomes 1, and the recognition result r is generated accordingly and displayed as d. At this point, the input pattern of equation (4) that is input to the recognition unit 13 is up to the /ro/ part, so it may be displayed incorrectly as "5" (/go/). However, since this pause period is shorter than the L set in the end detection section (for example, a time length of 250 milliseconds), the end detection signal e1 never becomes 1, so it is output as the output signal x. There's no fear. When the utterance ends up to the /ku/ part, a pause section at the end of the word indicated by reference numeral 62 is detected, so the recognition result r is determined by the same operation as above.
is determined and displayed as d. This time, the entire word “6” has been entered, so it is correctly written as “6”.
is recognized and displayed. In this way, the end detection signal e1 becomes 1 and the output signal x is sent to the target device only when the pause period 62 continues for the length L or more and reaches the time point indicated by the reference numeral 63. According to this operation, an incorrect recognition result is displayed once during the utterance, but since the speaker knows that the utterance is in the middle of utterance, no particular problem will occur if the speaker ignores the displayed result.

By displaying and outputting the recognition results in this way, the speaker can quickly know and confirm the recognition results, and there is also the possibility that a pause in the middle of a word may be mistakenly recognized as a pause at the end of a word, resulting in erroneous recognition. will not be carried out. This effect allows the speech recognition device to be used quickly and smoothly.

The principles of the present invention have been explained above based on examples, but these descriptions do not limit the scope of the present invention, and various modifications can be made to the method of detecting the end in the end detecting section. . For example, past
A time window k-A ≦ i ≦ k with a time width of about 250 milliseconds is set, and the sum of the amplitude levels within this window, that is, l _kA +l _k-A+1 +... l _k is a predetermined threshold. A possible method is to generate the termination detection signal at the time point k when the value becomes larger than θ′ _e . According to this method, the effect of instantaneous noise can be weakened to some extent by taking the summation, so that more stable operation can be expected than in the method described in the previous embodiment. In addition, by taking advantage of the property that the number of zero crossings of the input audio wave is low in the pause section, comprehensive judgment is made by combining the amplitude level and the number of zero crossings.
A method of detecting the pause section and the end is also effective.

Further, in the previous embodiment, the end detection signal e1, the rest period detection signal e2, and the like were used as intermittent DC signals, but the same operation can be achieved by using them as pulse signals. Furthermore, it is clear that the principle of the present invention can be implemented even if the amplitude level signal l _i and the like are processed as analog signals.

[Brief explanation of the drawing]

FIG. 1 is a block diagram showing an embodiment of the present invention, in which 1 is a microphone, 10 is an analysis section, 11 is a starting edge detection section, 12 is an end detection section, 13 is a recognition section, 1
4 is a recognition result display control section, 15 is an input enable state display section, 16 is an output section, 17 is a recognition result display section, 1
8 is a display lamp. FIG. 2 is a time chart for explaining the operation of the embodiment shown in FIG. 1, and FIG. 3 is a block diagram showing an example of the configuration of the analysis section. In FIG. 3, block 110 indicated by a broken line is a bandpass filter analyzer, block 120 is an analog multiplexer, 130 is an A/D converter, 140 is a rectifier, 150 is a low-pass filter, and 160 is an A/D converter.
It is a converter. FIG. 4 is a diagram for explaining the operation of the termination detection section 12, and FIG. 5 is a block diagram showing an example of the configuration of the termination detection section 12. In FIG. 5, 51 is a comparison circuit, 52 is an AND circuit, 53 is a counting circuit,
54 is a comparison circuit. FIG. 6 is a time chart showing part of the operation of the apparatus of the embodiment shown in FIG.

Claims (1)

[Claims] 1. An end detection section having a function of generating a pause section detection signal indicating that a pause section of the input audio has been detected and an end detection signal indicating that the end of the input audio has been detected; a recognition section having a function of performing a recognition operation and outputting a recognition result signal by assuming that the input voice has ended when the detection signal is generated; a recognition result display section for displaying the recognition result signal; and an output section that outputs the recognition result signal at the time when the termination detection signal is generated.