JPS63235999A - Voice initial end detector - 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] [Industrial Application Field] The present invention relates to a voice start edge detection device that detects a time point corresponding to the start edge of a voice region by performing sampling or other processing on the amplitude of a signal containing voice. .

[Prior Art] As a conventional example of detecting the start of a voice, a method using energy values for each section of a signal is often used. An example of this is the technique described in Japanese Patent Application Laid-Open No. 61-46999. This is done by extracting a sequence of a certain number of consecutive sections as the start detection section using two fixed thresholds,
The average value of the energy in the starting edge detection section is set as the third threshold, and the detection at the point in time when the energy falls below the third threshold without exceeding the lower of the two thresholds again in the starting edge detection section is performed. This is done in the backward direction.

Furthermore, there is Yasunaga Shindashi's Speech Recognition, Kyoritsu Shuppan (1979), which utilizes the energy value and the number of zero crossings for each section. This uses three threshold values for the energy value and the number of zero crossings for each section, which are obtained in advance by analyzing the non-speech parts.

Furthermore, as described in Japanese Patent Application Laid-Open No. 60-200300, there is also a method using dynamic factors such as changes in energy and changes in spectrum for each section.

[Problems to be Solved by the Invention] In general, a nasal signal called a buzz sound, which has a low fundamental frequency and relatively low energy, may be generated immediately before the occurrence of a voice, especially a plosive sound.

This buzzing sound may or may not occur even when the same person is making the same noise. Therefore, especially when performing speech recognition, it is necessary to extract the standard parameters used for it from an audio signal that does not have a buzz sound, and it is also necessary to remove that part in advance when a buzz sound occurs in the audio signal to be recognized. is. Therefore, when extracting audio signals for use in speech recognition, etc., it is desirable to automatically consider buzz sounds as non-speech and remove them in advance.

In the technique described in Japanese Patent Application Laid-Open No. 61-46999 mentioned above, when a buzz sound occurs, the start of the buzz is detected as the start of the sound, and it is impossible to remove it.

In addition, in the technique i1'Ji described by Yasunaga Anasai, Speech Recognition, Kyoritsu Shuppan (1979), the dark value is obtained by analyzing the non-speech part, so by extracting the buzz sound part as the non-speech part, Although it is possible to remove the buzz, it is difficult to extract the buzz, making it impractical.

Furthermore, the technique described in Japanese Patent Application Laid-Open No. 60-200300 requires analysis of spectra, etc., which has the disadvantage that processing is difficult and the apparatus is complicated.

[Object of the Invention] The present invention has been made in view of the above-mentioned drawbacks, and its main purpose is to extract audio signals suitable for processing such as speech recognition, which are not affected by buzz sounds. The purpose is to submit a voice start detection device.

[Means for Solving the Problems] Therefore, the present invention provides sampling means for sampling the amplitude of an input signal including audio at regular time intervals and converting the amplitude into a string of a plurality of amplitude values, and the plurality of amplitude values. calculation means for dividing a sequence of into a plurality of sections and calculating an energy-related value related to the energy value of the input signal for each section; and at least an energy-related value of the human input signal and a threshold value for the related value. extraction for extracting a continuous section that is determined as the first syllable of the speech in a speech start detection device that includes a speech start determining means that determines the start of the speech by taking into account the comparison result; means, a reference value determining means for determining a reference value that changes in relation to an increase or decrease in the energy-related value of each of the sections extracted by the extracting means, and a threshold value for determining the dark value in relation to the reference value. The invention is characterized in that it has a determining means.

[Operation] Accordingly, in the present invention, a signal including audio is converted into a sequence of multiple amplitude values by the sampling means, and the sequence of these multiple amplitude values is divided into multiple intervals by the calculation means, and the value of each interval is Energy-related values are calculated.

Thereafter, the extraction means extracts the syllable section at the beginning of the voice included in the signal. The reference value determination means extracts a reference value that changes in relation to an increase or decrease in the energy-related value in the extracted section, and the threshold value determination means calculates a threshold value in relation to the reference value.

Thereafter, the start end determining means compares the energy-related value for each section with the threshold value, and detects the start end of the audio by taking the comparison result into consideration.

In general, the energy-related value of a buzz sound generated by a single speaker is approximately constant, and the energy-related value of the voice does not change significantly. However, the level of the energy-related value of the audio signal used for speech recognition etc. is adjusted to be approximately constant, so setting a threshold value in relation to the energy-related value of the subsequent audio poses a practical problem. It becomes possible to separate the buzz sound and voice to the extent that there is no noise.

[Embodiment] Hereinafter, an embodiment of the present invention will be described in detail with reference to FIGS. 2 and 3.

Figure 2 shows a general-purpose central processing unit (hereinafter referred to as CPU) 1
5 is a block diagram configuring a voice start end determination device of the present embodiment using 5. A microphone 11 that collects acoustic information including the voice generated by a speaker and converts it into an electrical signal is connected to an input terminal of an amplifier 12. Amplifier 12 is configured to amplify the electrical signal sent by microphone 11 to a level suitable for subsequent processing. amplifier 1
An analog low-pass filter 13 is connected to the output of -7-'2. This filter 13 has a cutoff frequency of 4K
H2, and is configured to block signals with frequencies higher than the cutoff frequency. This filter 1
An A/D converter 14 corresponding to sampling means is connected to the output of No. 3. The output terminal of this A/D converter 14 is connected to the CPU 15. This CPU”+
5 is a read-only memory (hereinafter referred to as R
OM) 16 and an occasional write memory (hereinafter referred to as RAM) 17 are connected. This RA, M17 is A
An amplitude buffer 117a to which the amplitude values of the audio waveform sampled by the /D converter 14 are sequentially written; an energy-related value buffer 17b to which energy-related values are sequentially written; It includes a first pointer register 17c, a second pointer register 17d, a threshold value register 17e1 capable of storing a threshold value, and a working area for performing each process described later.

The operation of the voice start end detection device of this embodiment configured as described above will be explained in detail below with reference to the flowchart shown in FIG.

The voice generated by the speaker is collected by the microphone 11 and converted into an electrical signal. This electrical signal is amplified by an amplifier 12 to a level suitable for processing described later. The amplified electric signal is passed through the analog low-pass filter 13, which cuts off the signal valves of 4KH2 or higher. The above operation corresponds to step 20.

The output of the analog low-pass filter 13 is input to an A/D converter 14 corresponding to sampling means. A
The /D converter 14 samples the input signal at a sampling frequency of 8KH2 and outputs an amplitude value every 125 microseconds. This amplitude value includes all information up to 4KH2 of the input signal according to the sampling theorem. The i-th value of this amplitude value, that is, 12 from the start of sampling.
The amplitude value after 5X(i-1) microseconds will be hereinafter expressed as (i). This i is a value from 1 to n, where n is the total number of sampled amplitude values. These values A(1) to A(n) are sequentially written into the amplitude buffer 17a. This process is performed in step 21
It will be executed at

Next, the process proceeds to step 22, where the CPU 15 divides the n amplitude values into amplitude information for every 61!4 sections, that is, every 8 milliseconds, and calculates an energy-related value for each section. The energy-related value does not need to be an energy value in the strict sense, but may have properties relatively similar to an energy value. In this embodiment, the sum of the absolute values of every 64 amplitude values is used as the energy-related value. That is, if the j-th energy-related value is E(j>), the related value is obtained by the following formula.

4j E(j) −ΣIA(q)1 q=64j−63 Here, j is an integer, and its maximum value is the value of the quotient obtained by dividing n by 64 and rounding up the remainder. Hereinafter, this value will be referred to as the section number of that section.

This processing corresponds to the operation of the calculation means.

Next, the process proceeds to step 23, where the value of E(j> is compared with preset vowel detection thresholds in order from j-1, and E(
extract the first interval sequence in which at least two successive values of j> exceed the threshold for vowel detection, and store the interval number of the first interval in the sequence of intervals in the first pointer register 17G. The section number of the last section of the column is stored in the second pointer register 17d. This process corresponds to the operation of the extraction means.

Next, the process proceeds to step 24, where the energy-related values of each section extracted in the previous step 23 are read out from the energy-related value buffer 17b, and their maximum value is determined. This process corresponds to the operation of the reference value determining means, and the maximum value becomes the reference value.

Next, the process proceeds to step 25, where the reference value extracted in the previous step 24 is multiplied by a preset constant α, and the product is stored in the threshold register 16e.

This product corresponds to a threshold value and will be referred to as Te hereinafter. It is desirable that this threshold Te be set to a value slightly lower than the maximum value of the energy-related value of the buzz sound, and in this embodiment, this is achieved by setting it to approximately α-0.13. This process corresponds to the operation of the threshold value determining means.

Next, the process proceeds to step 30, where the value of the first pointer register 17G is read, 1 is subtracted from that value, and the value of the first pointer register 17G is subtracted from the value.
is stored in the pointer register 17G. Hereinafter, the value stored in this pointer register 17G will be expressed as k. That is, this process assigns -1 to k.

The value of k represents the section number of the section to be processed in steps 31 to 33 below. Next step 3
1, the threshold Te obtained in step 25 is compared with the energy-related value E(k) in the section indicated by the first pointer register 17G, and if the result is E(k)≧Te, step Returning to step 30, if E (k)<Te, proceed to step 32.

This step 31 corresponds to the operation of the comparison means.

In step 32, the number of zero crossings in the section indicated by the first pointer register 17C is calculated. The calculation method is described in Japanese Unexamined Patent Application Publication No. 117299/1983, so details thereof will be omitted.

Note that the number of zero crossings obtained in step 32 will be expressed as Z(k). Next, the process proceeds to step 33, where the number of zero crossings Z(k) obtained in the previous step 32 is compared with a preset threshold value for the number of zero crossings. This threshold value is set to 4 in this embodiment. Assuming that the threshold values are expressed as T and Z, if the comparison result is Z (k)≧TZ, the process returns to step 30, and if Z (k)<Tz, the process proceeds to step 34. That is, to summarize the operations from step 30 to step 33, an investigation is performed in a direction going back in time from the section immediately before the row of sections obtained at step 23, and first, E (k)<Te, force saZ (k)< The section corresponding to Tz is extracted, its value is stored in the first pointer register 17G, and the process proceeds to the next step 34. In this step 34, the first pointer register 17 at this point is
1 is added to the content of G, and the section with the section number represented by the sum is set as the start of the audio. This step 30
The processing from step 34 corresponds to the operation of the voice start edge extraction means.

In this way, in this embodiment, since the number of zero crossings, which shows a relatively large difference between the buzz sound and the voice, is also taken into account, it is possible to detect the start end more accurately.

Note that the present invention is not limited to the above-mentioned embodiments; for example, the sum of absolute values used as energy-related values in this embodiment may be a short-time energy calculated by the commonly used formula below. good.

4j E′(j) −Σ[A(q)]” ]q=64j−6 3.The reference value is the maximum value of the energy-related values in the section extracted by the extraction means, but the maximum value is taken. It may be the average value of the neighboring values of the section, or the mode value of the energy-related value of the extracted section.However, in this case, the value of the constant α is different from the value of this embodiment.Furthermore, In this embodiment, fluctuations in the number of zero crossings are used to determine the starting point, but changes in short-term spectra can also be used.

[Effects of the Invention] In the present invention, a threshold value is determined from a reference value that changes in relation to an input signal, and the start of audio is detected based on the threshold value. The signal can be accurately removed. As a result, by using the voice start edge detection device of the present invention together with an existing voice end detection device, it becomes possible to reliably extract voice information suitable for processing such as voice recognition.

[Brief explanation of drawings]

FIG. 1 is a flow chart showing the overall operation of the present invention, FIG. 2 is a block diagram showing the configuration of an embodiment of the present invention, and FIG.
The figure is a flowchart showing the operation of an embodiment of the present invention. In the figure, 14 is an A/D converter, 15 is a CPU, and 16 is an RO.
M, 17 is a RAM. Further, 21 is a step corresponding to the sampling means;
2 is a step corresponding to the calculation means, 23 is a step corresponding to the extraction means, 24 is a step corresponding to the reference value determining means, 25 is a step corresponding to the threshold value determining means, 30,
31, 32, 33, and 34 are steps corresponding to the voice start end determining means, and 31 is a step corresponding to the comparing means.

Claims (1)

[Claims] 1. Sampling means (14) that samples the amplitude of an input signal including audio at regular time intervals and converts the amplitude into a plurality of sequences of amplitude values; calculating means (22) for calculating an energy-related value related to the energy value of the input signal in each section; and comparing at least the energy-related value of the input signal with a threshold value for the related value. comprising a comparison means (31);
A speech start detection device comprising a speech start determining means (30, 31, 32, 33, 34) that determines the start of the speech by taking into account the comparison result, the continuous extraction means (23) for extracting a section; reference value determining means (24) for determining a reference value that changes in relation to an increase or decrease in the energy-related value of each of the sections extracted by the extraction means; and the reference value. a threshold determining means (25) for determining the threshold in relation to the threshold. 2. The reference value determining means (24) includes the extracting means (2).
Claim 1, characterized in that the reference value is determined as the highest value of the energy-related values of the section extracted by 3) or the average value of the energy-related values of the sections in the vicinity of the section having the highest value. The voice start end detection device described in .