JPH01285994A - Speech section detector - Google Patents

Abstract

PURPOSE:To enable setting of the optimum threshold and detecting of a speech section with high accuracy by monitoring the average power of noises by effective utilization of the idle time before a command is given and setting the threshold according to the max. value thereof. CONSTITUTION:The noise average power of every prescribed period is obtd. from the signal power detected by a characteristic detecting part 1, particularly the signal power of the noise in the period when there is no speech signal, in a threshold calculating part 2. The average level of the noise around the speech input part is monitored over a long period of time by utilizing the idle time before the command for starting the recognition processing is given from the control part in such a manner. The max. value of the noise average level is determined from the data on said level stored in the memory part and the threshold for detecting the speech section is calculated when the command is given. The detection of the speech section is thereby executed by setting the optimum threshold regardless of the length of the period of the change in the noise power.

Description

[Detailed Description of the Invention] [Objective of the Invention] (Industrial Application Field) The present invention adaptively optimally sets a threshold value used for speech segment detection of human speech used for speech recognition etc. according to the noise level. The present invention relates to a voice section detection device capable of detecting speech intervals.

(Prior Art) In a speech recognition system, it is important to detect a speech section of an input speech with high precision, extract only the input speech information of this speech section, and provide it for recognition processing.

Conventionally, this type of voice section detection has been performed, for example, at a predetermined period (5
This is performed by discriminating the voice power of human voice detected by feature extraction every 20 isec) using a predetermined threshold value. Specifically, when the audio power exceeds the threshold value for a certain period of time or more, the first point in time when the audio power exceeds the threshold value is detected as the beginning of the audio, and then when the audio power exceeds the threshold value for a certain period of time or more, it is detected as the beginning of the audio. When this occurs, the voice section detection is performed by detecting the point in time when the voice first falls below the threshold as the end of the voice. In addition, a threshold value for detecting the start end and a threshold value for detecting the end end are set respectively, and a voice section is detected from the relationship between these threshold values and the voice power.

By the way, the threshold value for detecting such voice sections is usually
It is set based on the noise power in the noise section before the start of the voice section. Generally, when a command to start recognition processing is given from the control unit, the noise power of several frames immediately after starting to capture the input signal is averaged, and a threshold is set according to this noise average power.

However, the surrounding environment surrounding the audio input section changes in various ways, and the noise that exists there also varies, with some having long and short periods of power change. For this reason, the conventional method of determining the average value of the noise power of several frames immediately after the start of recognition and setting the threshold does not work well for noise with a short period of power change of about 1 to 2 frames. Although it is possible to set a fairly regular threshold value, if there is noise with a long period of, for example, 10 to 20 frames, it becomes difficult to set the threshold value correctly.

In other words, if the average value of the noise power is calculated at a time when the noise power is low, the threshold value will be set to a smaller value than it should be, which may cause a problem in which voice section detection is performed including the noise section. arise. On the other hand, when the noise power is high, the threshold value is set to a value larger than the original value, and a problem arises in which voice section detection is performed with part of the voice section being omitted.

Such a problem can be resolved by detecting the noise average power over a reasonably long period of time. However, there is a limit to the length of time that can be set for detecting the noise average power after a command to start recognition processing is given from the control unit. Also, if you set a time like this,
Problems arise such as the waiting time until the sound f'4 is started by human power, and the response is delayed.

(Problem to be solved by the invention) In this way, in the past, in an environment where noise with a long period of power change exists, the average power of the noise is detected with high accuracy, and the threshold value for detecting voice sections is optimally set. There was a problem in that it was difficult to improve the accuracy of voice section detection.

The present invention has been made in consideration of the above circumstances, and its purpose is to accurately determine the average power of the noise, regardless of the length of the period of the noise noise change, and to calculate the noise interval. An object of the present invention is to provide a voice section detection device that can optimally set a threshold value for detection.

[Structure of the Invention] (Means for Solving the Problems) The present invention detects the noise level before the start of audio input, calculates a threshold value according to this detected noise level, and uses this threshold value to discriminate the input audio level. In a speech section detection device that detects a speech section of input speech, the noise average level is determined at every predetermined period until an instruction to detect a speech section is given, and this noise average level is stored for a predetermined number of times. At the same time, the noise average level stored in this storage section is updated one after another to always obtain the latest noise average level in the storage section. At the start of voice section detection, the maximum value of the average noise level stored in the above register is detected, and the threshold for voice section detection is optimally set based on this maximum value of the noise average level. This is a characteristic feature.

Furthermore, when storing the noise average level detected at the predetermined period in the storage section, if the noise average level detected at each predetermined period becomes significantly large, storing the noise average level in the storage section is performed. This is characterized in that the suspended noise average level is discarded if this extremely large noise average level does not continue for a certain period of time or more.

(Function) According to the present invention, the average level of noise around the voice input unit is monitored for a long time using the free time before the command to start recognition processing is given from the control unit, and the above command is executed. When given, the maximum value is calculated from the latest predetermined number of noise average level data stored in the storage unit, and the threshold for detecting voice sections is calculated according to this maximum value of the noise average level. It becomes possible to set an optimal threshold value and perform voice section detection regardless of the length of the period of power change.

In addition, if an extremely large noise average power is detected, the storage of the detected noise average power in the storage unit is suspended, and if the extremely large noise average power does not continue for a certain period of time, the stored detected noise average power is Since the power is rejected and stored in the storage unit only when it continues for a certain period of time or more, it is possible to set an optimal threshold value without being affected by large noises that occur from time to time.

Therefore, by setting a threshold according to the surrounding environment of the audio input section,
It becomes possible to perform voice section detection 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 segment detection device according to an embodiment of the present invention. In this FIG.
This is a feature extraction unit that performs acoustic analysis on a frame of ~20 rAsec) and detects various feature parameters of the input signal. One of the feature parameters detected by the feature extractor l is signal power. The threshold calculation unit 2 calculates the noise average power for each predetermined period from the signal power detected by the feature extraction unit 1, especially the signal power of noise during a period when no audio signal is input, and instructs the start of speech recognition processing. When the noise is detected, a threshold value for voice section detection is determined based on the noise average power.

The threshold value calculated by the threshold calculation unit 2 is the voice section detection unit 3.
, and the feature extraction unit 1 detects the speech section in the human voice and extracts the feature parameter series in the detected speech section. Then, the matching unit 4 compares the characteristic parameter series of the input voice extracted as described above (characteristic pattern of human voice) with each standard pattern of the recognition target voice registered in the standard pattern memory 5, and searches for similarities, for example. It is a matter of seeking degree. The comparison with this standard pattern is performed, for example, by calculating respective composite similarity values.

The degree of similarity obtained in this way with each standard pattern is compared with each other in the determination unit 6, and the category of the standard pattern that has the maximum degree of similarity is determined as a recognition result (recognition candidate). Processing takes place.

Figure 2 shows the speech recognition device configured in this way.
FIG. 3 is a diagram showing an example of the configuration of a threshold calculation unit 2 that calculates the threshold used for the voice section detection, and FIG. 3 shows the processing concept (
FIG. 2 is a diagram schematically showing functional blocks.

This threshold calculation unit 2 operates during the period until a command to start recognition processing is given from the control unit (host).
The characteristic information of the input signal given through the switch Sν
When the command is given, the feature information of the input signal given from the feature extraction section 1 is bypassed to the voice section detection section 3 via the switch Sw. In other words, using the free time until the command to start recognition processing is given, input the noise during that period, sequentially calculate the noise average power that will be the standard for calculating the threshold, and then use the noise average power as the reference for threshold calculation. The threshold value is set based on the threshold value.

The threshold calculation unit 2 sequentially stores the noise power taken in at a predetermined period in the power storage register 11 (step a). When a predetermined number of data (noise powers PI, P2. to Pk) are stored in the power storage register 11 (step b), the noise average power m is calculated according to these data (step C). The calculation of this noise average power m is performed by inputting the noise powers P I, P 2 . ~ Pk is read out sequentially, and the adder 12
Find the sum (cumulative value) using . And this noise power P1. P2. It is obtained by dividing the sum of ~Pk by the number of additions in the divider 13.

The noise average power m obtained in this manner is sequentially stored in the average value storage register 14, and the power storage register 11 is initialized to prepare for calculation of the noise average power in the next cycle. Such an average value storage register 14
The storage of the noise average power m is repeatedly executed until the above-mentioned command to start the recognition process is given (step d). The average value storage register 14 stores the noise average power m obtained for each predetermined cycle over a predetermined number of cycles (nine), and each time a new noise average power is obtained, the register 14 will be updated. As a result, the latest nine noise average powers m are always stored in the average value storage register 14.

However, when the command to start the recognition process is given,
First, the maximum value of the noise average level is extracted (step e). The maximum value of the noise average level is extracted by sequentially reading out the noise average power m stored in the average value storage register 14 to the comparator I5 and comparing it with the reference value of the noise average power stored in the register 16. While
This is done by finding the noise average power with the larger value, storing it in the register I6, and using it as a new reference value. Note that the register 16 is initially set to [0].

By successive comparison of the noise average powers m, the maximum value among the noise average powers stored in the average value storage register 14 is obtained. Then, the maximum value of the noise +i average power obtained in this way is determined by the threshold calculation unit 1.
7, and a threshold value for voice section detection is set (step f).

According to the threshold calculation unit 2 configured in this way, the noise average power is obtained over a long period of time using the idle time of the audio input, and this noise average power is always calculated over the latest plural cycles. Since the threshold value is set using the maximum value of the noise average power stored in the average value storage register 14,
Even when the period of change in noise power is long, it is possible to set an appropriate threshold value according to the noise power according to the measurement results of the noise average power over a long period of time. Then, it becomes possible to perform highly accurate speech section detection.

In other words, this is different from the conventional method, which measures the noise average power in a short period of time immediately after a command to start recognition processing is given, so the noise average power can be accurately calculated regardless of the period of noise power change. , it becomes possible to set an optimal threshold value.

By the way, some noises have values that increase over time. If the noise average power is calculated by detecting such temporally large noise as it is, there is a risk that the noise average power will become unintentionally high. Therefore, in such a case, it is preferable to calculate the noise average power by removing noise components whose power increases over time.

FIG. 4 shows a function provided in the threshold calculation unit 2 described above to reject the above-mentioned temporary peak noise.
The figure is its functional block diagram.

This peak noise rejection function is performed by detecting the noise power at a predetermined period (step A) and determining whether the value is higher than a predetermined noise level (step B), as described above. Specifically, this is performed by comparing and determining whether or not a noise power that is M times or more the noise power that has been detected so far is detected using the comparator 2.

Specifically, the comparison standard used for this comparison judgment is made by multiplying the noise power detected one cycle before by M in the multiplier 22, storing this in the register 23, and applying it to the comparator 21. be exposed.

However, when extremely large noise power is detected, it is not stored in the power storage register 11 as it is, but is stored in another power storage register 24.
(Step C). Thereafter, step D) of repeatedly performing the noise power detection described above;
Next, it is determined whether the detected noise power has become extremely small (step E). This determination is made by comparing and determining in the comparator 25 whether or not a noise power that is equal to or less than (1/M) of the previously detected noise power is detected. That is, the comparison standard used for this comparison judgment is specifically that the noise power detected one cycle ago is divided by the divider 26.
(1/M) times M), and register 27
This is done by storing the data in the data and supplying it to the comparator 25.

If the detected noise power has not become smaller, its value is sequentially stored in the power storage register 24 (step F). This process is repeatedly executed, and when a predetermined number of noise powers are stored in the power storage register 24 (step G), the data stored in the power storage register 24 is transferred to the power storage register 11. (Step H).

On the other hand, if the detected noise power becomes extremely small before a predetermined number of noise power data are stored in the power storage register 24, the detected noise with a large power will temporarily disappear as described above. It is determined that the noise power data is the same, and all of the noise power data stored in the power storage register 24 is discarded (step (2)).

Through the above processing procedure, noise components that have become temporarily large are discarded. Therefore, with this device equipped with such a peak noise rejection function, it is possible to detect the average noise average power without being affected by extreme temporal noise changes, and it is possible to detect speech intervals. It becomes possible to effectively set an appropriate threshold value for.

Note that the present invention is not limited to the embodiments described above. For example, the peak noise rejection data can also be used when storing the noise average power in the average value storage register 14. In this way, it becomes possible to deal with changes in noise power over a longer period. Then, it becomes possible to perform effective speech section detection by considering only stationary noise components. In addition, 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, the noise average power is detected in a short period of time immediately after a command to start recognition processing is given, and a threshold value for speech section detection is set. Differently, since the free time before the above command is given is effectively used to monitor the average noise power and the above threshold is set according to its maximum value, it is possible to determine the optimum value for the part without being concerned with the period of change in the noise power. Great practical effects can be achieved, such as being able to perform speech segment detection with high accuracy by setting a threshold value.

[Brief explanation of the drawing]

The figures show one embodiment of the present invention, and FIG. 1 is a schematic configuration diagram of a speech recognition device incorporating a speech segment detection device according to an embodiment of the present invention, and FIG. 2 is an embodiment of the present invention. FIG. 3 is a functional block diagram of the threshold calculation section; FIG. 4 is a diagram showing a configuration example of a peak noise rejection function incorporated in the threshold calculation section; FIG. 5 is a block diagram of a peak noise rejection function. 2...Threshold value calculation unit, 3...Pa voice section detection unit, 11.
...Power storage register, 12...Adder, 13.
...Divider, 14...Register for storing average value, ■5.
... Comparator, 16... Register, 17... Threshold calculation unit, 21... Comparator, 22... Multiplier, 23...
Register, 24...Power storage register, 25...
-Comparator, 26...Divider, 27...Register. Applicant's agent Patent attorney Takehiko Suzue

Claims (2)

[Claims] (1) In a speech section detection device that calculates a threshold according to the noise level before the start of speech input, uses this threshold to discriminate the speech level, and detects the speech section of the input speech, until an instruction to detect the speech section is given. means for determining the noise average level at each predetermined period; a storage unit that stores the noise average level determined by this means while updating the noise average level over a predetermined number of times; and 1. A voice section detection device, comprising means for setting a threshold value based on a maximum value of an average noise level stored in the section. (2) In the speech interval detection device according to claim 1, when the average noise level detected at each predetermined period becomes extremely large, storage of this average noise level in the storage unit is suspended, 1. A voice section detection device characterized by having a function of discarding the suspended noise average level if the noise average level does not continue for a certain period of time or more.