JPS62211699A - Voice section detecting circuit - 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] [Object of the Invention] (Industrial Application Field) The present invention provides a speech segment detection circuit that can easily and accurately detect speech segments from input speech necessary for speech recognition processing. Regarding.

(Prior Art) Recently, various attempts have been made to perform AM processing on input speech and provide it for information processing. For example, the role played by speech recognition processing has become very important, such as as an information input means for a computer as a man-machine interface, or as an information input means for a document created in an audio word processor.

This speech recognition process is generally performed by comparing the feature data series of the input speech with preset dictionary data. Therefore, it is important to detect the input voice section (the start and end of the voice) from the input voice data with high accuracy.

Conventionally, therefore, a voice section detection circuit is generally used to detect the input voice power value as the start point when it exceeds the threshold value, and to detect it as the end point when the voice power value falls below the threshold value. I'm trying to detect it. That is, the audio power value is compared with a predetermined threshold value to detect a section in the input data that actually includes audio.

However, when performing voice section detection in this manner, there are the following problems.

That is, if there is only one threshold value for detecting the start and end points, and the threshold value is set to a low value, there is a risk that a voice section including noise components will be detected. On the other hand, if the threshold value is set high, there is a risk that audio data with a low input level will be regarded as a noise component and will be discarded. In particular, such a problem is likely to occur at the beginning of the input audio.

Therefore, two thresholds are set, and if the maximum value of the input audio power falls within the range, the lower threshold is used to detect the voice section, and if the maximum value of the audio power falls within the range. It is considered that when the threshold value is exceeded, the voice section is detected using the higher threshold value.

However, just because the maximum value of the power of the input voice is large does not necessarily mean that the power of the input voice is high overall. For this reason, if the threshold value is increased according to the maximum value of the audio power, there is a very high possibility that the audio will be truncated at the beginning.

(Problems to be Solved by the Invention) In view of the fact that in the past, it has been extremely difficult to detect speech sections, especially to detect the beginning, with high precision, the present invention It is an object of the present invention to provide a voice section detection circuit that can detect voice sections simply and with high accuracy, regardless of the magnitude of the power value.

[Structure of the Invention] (Means for Solving the Problems) The present invention provides the following features: (1) When the power value of input audio exceeds a predetermined threshold T1 over a preset first number of frames F1, the audio Detecting the frame whose power value first exceeds the threshold T1 as the starting frame of the input audio;
When the audio power value is below the predetermined threshold T1 for a preset second number of frames F2 or more after detection of the start frame, the audio power value is lower than the amT1.
The frame in which the first frame becomes lower than the threshold value T1 is detected as the end frame of the input audio; The present invention is characterized in that the frame in which the voice power value exceeds the threshold T1 immediately before the value reaches the maximum value is re-detected as the starting frame of the input voice.

(Function) Thus, according to the present invention, unless the voice power value exceeds the threshold value T1 for the number of frames F1 or more, this is not detected as the starting point, so even if the power value of the input voice is high overall, Also, noise components are no longer detected as voice sections.

Also, after the start edge is detected, the audio power value is set to the threshold T1
Since this is not detected as the end unless the number of frames is greater than or equal to the frame number F2 and less than 1, there is no possibility that a transient drop in the audio power value will be mistakenly detected as the end.

Furthermore, if the audio power value is higher than a predetermined set value after the detection of the starting frame, the detection is redone using the frame whose audio power value exceeded the threshold T1 immediately before reaching that state as the starting frame. Therefore, it is possible to detect the starting edge frame by effectively eliminating noise components that tend to cause incorrect starting edge detection when the audio power level is high.

Therefore, regardless of the magnitude of the power value of the input voice, it is possible to detect the voice section easily and with high accuracy.

(Example) Hereinafter, an example of the present invention will be described with reference to the drawings.

FIG. 1 is a schematic configuration diagram of a speech recognition device incorporating a speech section detection circuit according to the present invention.

This speech recognition device basically includes an acoustic analysis circuit 1, a speech section detection circuit 2, a matching processing circuit 3, a dictionary data memory 4, and a post-processing circuit 5.

The acoustic analysis circuit 1 is composed of, for example, a plurality of band-pass filter groups, and acoustically analyzes input speech every predetermined frame period by frequency analysis or the like to obtain its acoustic characteristics.

At the same time, the audio power of the entire band of the input audio is obtained for each of the predetermined frame periods. These analysis data (input audio data) are digitized for each frame and output from the acoustic analysis circuit 1.

The speech segment detection circuit 2 is configured to include an audio data RAM 2a that stores the audio data obtained by the acoustic analysis circuit 1, a work memory 2b used for speech segment detection processing to be described later, and the like. The acoustic feature data of the voice section detected by the voice section detection circuit 2 is selectively read out from the voice data RAM 2a and provided to the matching processing circuit 3.

The matching processing circuit 3 matches the feature data series of a given speech section with each speech feature data of a plurality of recognition target categories registered in advance in a dictionary data memory (ROM) 4, and performs a matching process, for example, between them. The input speech is recognized by calculating the composite similarity of the input speech. The recognition candidates for the input speech obtained by the matching processing circuit 3 are given to the post-processing circuit 5, and the similarity values obtained from the recognition candidates are subjected to testing processing and linguistic testing processing to obtain the recognition results. Desired.

Now, the speech section detection process in the speech section detection circuit 2 is performed as follows.

Figures 2 (a) and (b) show the processing concept,
FIG. 3 is a diagram showing the flow of the processing procedure.

This voice section detection circuit 2 basically uses the input voice power value X and a predetermined lli! [This is done by comparing with T1. First, as shown in FIG. 2(a), when the audio power value X exceeds the threshold T1 for a preset number of frames F11X, the audio power value X exceeds the threshold T1 for the first time. The frame at which this occurs is detected as the starting frame $1 of the input audio.

After the start frame S1 is detected in this way, when the audio power value X falls below the threshold T1 by a preset number of frames F2 or more, the audio power value The frame when the value falls below T1 is detected as the final frame E of the input audio. In this example, after the start frame S1 is detected, there is a period in which the audio power value X is below the threshold T1;
is less than the number of frames F2, the frame X when the audio power 1[x first falls below the threshold T1 at that time is not detected as a terminal frame.

The above process is the primary verification process for the input audio,
As a result, the start frame S1 and the end frame E of the input audio are respectively detected.

However, if the power of the input audio is high overall,
Along with this, the power of the noise component may also be high, and as described above, the detected start frame S1 may include the noise component.

Therefore, the voice section detection circuit 2 sequentially detects the voice power 111x, and if the voice power value X becomes higher than the threshold value T1 by a preset value after detecting the start frame, as shown in FIG. As shown in (1)), a second 1iilT2 is assumed. This second threshold T2 is assumed to be a value lower than the audio power value X by the set value, that is, (T2-x-D). Then, the frame at which the audio power value X exceeds the threshold T1 immediately before exceeding the threshold T2 is re-detected as the starting frame S3 of the input audio. Specifically, after the detection of the start frame, the audio power value X becomes (TI +
D) is detected, and the frame at which the audio power value X exceeds the threshold T1 immediately before that state is reached is used as a new starting frame, and the detection is redone.

This process is a secondary test process, and when the power value
This is done in consideration of the possibility that the start frame S1 obtained in the next cutoff process may contain noise components. Then, the process is repeated while updating the assumed threshold T2 as described above until, for example, the audio power value X takes the maximum value.

As a result, for example, the audio power value
[If T1 is sometimes lower than T1, the first detected start frame is corrected. Then, as shown in FIG. 2(b), the noise component of the 0 part is separated, and the speech section is detected from the original starting part (2) of the input speech.

FIG. 3 shows the flow of such voice section detection (starting and ending frame detection) processing, and will be briefly explained.

This process is performed after each data area set on the work memory 2b is initialized to "0" (
Step a).

Here, 81 is an area for storing the starting edge candidate frame obtained in the primary inspection process, and S2 is an area for storing the starting edge candidate frame obtained in the secondary inspection process. Also S3
E is an area for storing the finally determined start frame, and E is an area for storing the end frame. Further FL
AG is an area indicating that the starting edge candidate frame needs to be changed due to the secondary inspection process.

However, when the above initial settings are completed, the audio data R
step b) of reading out the voice power values xi in order from AM2a; determining whether the voice power values xi exceed the first threshold T1 (step C);
[When xi exceeds the first equivalent value T1, the frame number at that time is stored as the starting end candidate frame S1 (step d).

Thereafter, the audio power value xi is read out from the audio data RAM 2a (step e), and it is determined whether or not the audio power value xi exceeds the first threshold T1 (step f). If the power value xt is less than the first threshold T1, the process from step b described above is repeated.

At this time, if the audio power value xi exceeds the first threshold T1, it is determined whether or not this state continues for more than a preset number of frames F1. If it is less than the number F1, the process from step e is repeated.

If the audio power value xi continues to exceed the threshold value T1 by the number of frames F1 or more, the frame number obtained as the starting edge candidate frame S1 is assumed to be the starting edge frame S3 obtained for the input audio. Store (step h).

As described above, the start frame is detected by the primary cutoff process.

After that, the audio power value x 1 is read out again from the audio data RAM 2a (step 1), and it is determined whether the audio power value xi exceeds the first a@TI (step 1).
Step j). Then, the audio power value xi is the first threshold T
If it is less than 1, store the frame number at that time as the end frame E (step k), then read out the audio power value x1 (step 2), and check whether the audio power value xi exceeds the first threshold T1. Determine whether or not (
Step m).

If the audio power value xi is below the first threshold T1, then it is determined whether the condition continues for a set number of frames F2 (step n), and if the condition is satisfied. In this case, the frame number set in the start frame S3 is determined as the start frame of the input audio, and the frame number set in the end frame E is determined as the end frame, respectively, and the voice section detection process is performed in step o). finish.

Therefore, in this case, the voice section detection process ends without performing the secondary inspection process.

By the way, after determining the end frame E as described above, the audio power value xi becomes the first threshold before the audio power value xi remains below the first threshold T1 for a predetermined number of frames F2 or more. If the threshold value T1 is exceeded, this is determined in the determination process in step m. In this case, it is determined that there is a possibility that the first frame detected earlier contains a noise component. Therefore, the frame number at that time is stored as the starting end candidate frame S2 for the secondary inspection process (step p). Then, the FLAG is set to "1" to indicate that there is a possibility that re-detection of the start frame by the secondary inspection process, that is, re-detection of the start frame is necessary (step q).

After that, it is determined whether the voice power value xi is higher than the threshold T1 by a set value or more, that is, the voice power value xt
is higher than (TI +D) (step r). This process is also performed when the audio power value xi continues to be higher than the threshold T1 after the start frame S3 is detected. If this condition is not met, the process returns to step i described above and the process described above is repeated.

Note that if an end frame is detected during this iterative process, it is determined that the temporary decrease in the threshold value T1 of the audio power value xi is due to noise.

Therefore, in this case, the frame number detected as the start frame S2 is ignored.

On the other hand, during this iterative process, the audio power value x1 becomes (T
I+D), it is then determined whether the value has reached the maximum value (step S). When the maximum value of the audio power xi is detected, the FLA
It is determined whether G is "1" or not (step t).

Here, if FLAG is "0", the audio power x
This shows that i continues to exceed the threshold T1. Therefore, in this case, there is no need to redetect the start frame $3 of the input audio, and the process from step i described above is repeated, ignoring the information on the start frame S2.

On the other hand, when FLAG is "1", as mentioned above, the audio power value xi is lower than the -H threshold T1,
Thereafter, it is shown that the audio power xi is higher than the above (TI +D). Therefore, in this case, because the input level of audio power is high overall, it is determined that the start frame obtained in the first cutoff process represents a noise portion. Therefore, the starting edge candidate frame S2 detected in step m is updated as the newly found starting edge frame S3 in step u), and at the same time, the FLAG is reset to "O".

Through this process, the audio power xi becomes higher than (TI + D), and the audio power xi
Once falls below the threshold T1, a frame whose audio power x1 exceeds the threshold T1 immediately before reaching the maximum value is determined as a new starting edge candidate frame S3. In other words, the starting edge candidate frame obtained before the audio power X: once falls below the threshold T1 is updated as a new starting edge candidate frame by the frame that exceeds the threshold T1 again after the audio power xi falls below the threshold T1. It turns out.

The above process is the secondary inspection process, and after this secondary inspection process is performed, the above-mentioned end frame detection process is performed.

In this way, the voice section detection circuit 2 sets the frame that exceeds the threshold T1 for the first time when the state exceeds the threshold T1 and continues until the predetermined frame F1 as the starting frame, and also sets the frame that exceeds the threshold T1 as the starting frame. When the state in which the threshold value T1 continues to be lower than the threshold value T1 continues every predetermined frame F2, the first frame in which the threshold value T1 is lower than the threshold value T1 is detected as the terminal frame. Further, if the audio power is greater than the predetermined value after the start frame is detected, the frame that exceeded the threshold T1 immediately before the audio power exceeded the predetermined value is re-detected as the start frame.

Therefore, regardless of the power level of the input voice, it is possible to effectively eliminate noise components and detect the voice section with high precision. However, the process is not very complicated and can be executed relatively easily. Therefore, it is possible to perform the speech recognition process with high precision using the information on the speech sections detected with high precision in this manner.

Note that the present invention is not limited to the embodiments described above. For example, the threshold value T1 and the number of frames Fl.

F2 may be set according to its specifications.

Furthermore, the algorithm for detecting the start frame and the end frame can be modified in various ways. In short, the present invention can be implemented with various modifications without departing from the gist thereof.

[Effects of the Invention] As explained above, according to the present invention, it is possible to easily obtain the speech section of input speech with high accuracy and by effectively eliminating noise components, regardless of the input level. It becomes possible.

[Brief explanation of drawings]

FIG. 1 is a schematic configuration diagram of a speech recognition device configured using a speech section detection circuit according to an embodiment of the present invention, FIG. 2 is a diagram showing the processing concept of the speech section detection circuit, and FIG. It is a figure showing the processing procedure of section detection. 1... Acoustic analysis circuit, 2... Voice section detection circuit, 3
. . . matching processing circuit, 4 . . . dictionary data memory, 5 . . . invasion processing circuit.

Claims (2)

[Claims] (1) When the power value of the input audio exceeds a predetermined threshold for a preset first number of frames or more, the frame in which the audio power value first exceeds the threshold is selected from the input audio. means for detecting a starting edge frame; and when the audio power value falls below the threshold value for a preset second number of frames or more after detection of the starting edge frame, the audio power value falls below the threshold value. means for detecting the first frame that falls below as the end frame of the input audio, and when the audio power value becomes larger than the threshold by a predetermined set value or more after the detection of the start end frame, the value becomes a maximum value; 2. A speech section detection circuit comprising: means for re-detecting a frame in which the speech power value exceeds the threshold immediately before the input speech as a starting frame of the input speech. (2) The voice section detection circuit according to claim 1, wherein the power value of the input voice is digitized at a predetermined frame period, stored in the voice data memory, and used for voice section detection processing. .