JPH0383100A - Detector for voice section - 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 FIELD OF INDUSTRIAL APPLICATION The present invention relates to a device for detecting voice segments contained in an electrical signal obtained via a transducer.

Conventional technology There are cases in which a sound signal is extracted from an electrical signal and input to data processing equipment such as a microcomputer for processing.

In general, audio extraction methods include a level detection method that detects the energy, such as amplitude, of an audio signal and compares it with a predetermined threshold; a method that detects and determines the number of zero crossings or polarity bit sequence of an audio signal; , a method of combining both is used.

The level detection method uses a comparator to compare the electrical signal used to determine the presence or absence of an audio signal with a predetermined threshold, and when the level of the electrical signal is higher than the threshold level, it is determined to be an audio signal. Output the audio signal part. As this threshold, a power threshold is used, which is a determination level for determining whether there is speech or no speech by determining the power from the audio waveform during encoding and decoding.

Problems to be Solved by the Invention Although this level detection method is simple, if the threshold value is set too high, it may not be possible to detect a vocal section. There were problems with false detections when there was a lot of noise, and parts with low energy levels such as consonants were not detected.

In addition, in the conventional voice extraction method, the threshold value for detecting the level of the background sound increases. Ji
The disadvantage of erroneously recognizing d5 is that it is difficult to set an appropriate threshold level.

This I2 method for cutting out voice segments is required to have fewer malfunctions due to noise, to avoid missing beginnings and endings of words, and to be processed by a general-purpose microcomputer, so it requires less software processing time.

SUMMARY OF THE INVENTION An object of the present invention is to solve the above-mentioned problems and provide a method for cutting out a speech signal that can reliably recognize a speech section even in a signal with large background noise.

Means for Solving the Problems The present invention comprises means for dividing a signal from which a three-tone signal is to be detected into a plurality of blocks, and a circuit for calculating the sum of absolute differences in energy levels between signals in each block. , and a circuit for calculating a threshold value for determining the level of an audio signal from the sum of absolute values.

In this invention, a signal from which an audio signal is to be detected is divided into a plurality of blocks, and the sum of absolute differences in energy levels between signals in the blocks is used as a feature quantity to set a threshold for determining the level of the audio signal. The value is determined from the above feature amount. The feature values and threshold settings are calculated as follows.

Divide the feature audio signal into blocks, and calculate the sum of absolute differences between adjacent audio signals in the same block to extract the audio section! Let it be kiF.

t Block index.

(0, 1, 2, 3,...) Audio signal amplitude at time t.

Number of audio signals in the block. F is stable length of 1 bit or more so that the value is Satoshi.

80 (10m5ec at 8KHz sampling) To prevent false detection due to threshold setting 1v3λWg, threshold value T
1-( is set adaptively depending on the back 1☆ straddle level.

The background noise of the CNT NOISE block length is input, and the average feature amount F' is set using the following equation.

TH=G(F’) -・For example, CN T N0ISE length. It was set as IO block.

G(x) = 1.25*X→500 (X≧550
)=1200 (X<550).

In the embodiment shown in FIG. 1, the audio data is sampled at appropriate time intervals.

1 is the immediately preceding data register that stores the audio data from the previous sampling period, and 2 is the audio data value input at the sampling period and the audio data from the immediately preceding sampling period that is applied from the immediately preceding data register. 3 is an addition circuit that adds the absolute difference value and the sum of absolute difference values; 4 is a circuit that calculates the threshold value from the sum of absolute difference values; 5 is the constant 1 and the difference. A comparator that compares the sum of absolute values; 6 a register that holds a threshold value; 7 a comparator that extracts audio data from the signal obtained from the calculation circuit 3; the threshold value is obtained from the threshold register 6; The threshold value is m. Each bit is connected to a power line or a ground line. Even if there is no background sound, a certain value is set as the threshold. Its constant value is constant 2. Constant 1 is the value used to determine whether constant 2 is to be adopted as the threshold value.

8 is a counter that stores how many pieces of audio data have been read; 9 is a controller that controls the overall processing flow and issues signals to set values for each nostar.

11 is a counter for detecting the beginning of a word, which counts the number of blocks starting from a block with a total sum greater than TH; I2 is a counter for detecting the beginning of a word, a counter that counts the number of blocks having a sum greater than TH; 13, a counter for detecting the beginning of a word; A counter for burst error detection is a counter that counts the number of blocks after the beginning of a word is found, and unless the content of this counter exceeds a certain value, it is not recognized as a speech section.

14 is a counter for finding the end of a word, and is a counter for counting the number of blocks smaller than T)f.

FIG. 2 is a processing flowchart for calculating a threshold value for the comparator 7 for detecting a voice section.

In step S], one piece of audio data is loaded, and in step S2, the absolute difference value between the previous data and the audio data read in step Sl is calculated by the absolute difference calculation circuit 2, and the difference calculated in step S2 is calculated. The absolute value is determined in step S3,
The calculation circuit 3 is used to add 4- the sum of the absolute difference values. In step S4, it is determined whether the data counter 8 has reached a predetermined value n, and if it has reached n, the process jumps to step S6.

When the data counter is smaller than n, the process advances to step S5. Then, the data counter 8 is incremented by 1, the value of the audio data obtained in the current sampling is set in the immediately preceding data register 1i, and the process returns to step S1.

In step S6, the sum of the absolute difference values is calculated by the number of blocks (a certain number of sounds is defined as l block, and a certain number of such blocks are collected to obtain the absolute difference value. In our case, a-
128, b=8. ) and take the average.

In step S7, the comparator 5 determines whether the average calculated in step S6 is larger than a certain value α (constant 1). If α is larger than the average calculated in step S6, the process goes to step S8; otherwise, the process goes to step S9. α is larger than the average when the background noise is large, and the multiplexer 15 is switched to the threshold juice calculation circuit 4 side.

α is smaller than the average when the background noise is small,
Switch multiplexer 15 to constant 2 side.

In step S8, the threshold value is calculated by the threshold juice calculating circuit 4 based on the calculation formula, and the threshold value is set in the threshold value renostator 6. The calculation formula is, for example, threshold value - 125 x average value +
500/16) was used.

As described above, this threshold value is applied to the comparator 7 and compared with the data signal applied from the sum calculation circuit 3, and the comparator 7 converts the data signal exceeding this threshold value into an audio data signal. Output as . As a result, audio data is extracted from the input data signal.

After calculating the threshold value, the voice section is determined as follows.

(Start of word) Calculate the absolute difference value between adjacent data, and calculate the sum S of the absolute difference values for l blocks. S from threshold TH
T HD (constant) where larger blocks are consecutive
If there are T HTH (constant) blocks or more in the block, it is treated as the beginning of a word. In reality, the beginning of a word is defined as a point a certain amount of time before the beginning of the word was found.

(End of Word Detection) If there is no block with a feature amount larger than the threshold value TH in the T HEND block length, the first block is determined as the end of the word.

(Burst Detection) If the interval between the beginning and end of a word is shorter than a certain period of time, it is considered a burst error and is not considered as a speech section.

This is to reject sudden noises such as tongue clicks.

In the circuit shown in Fig. 1, the immediately preceding data register 1 and the calculation circuits 2, 3.4 are used to extract the audio section and calculate the threshold value, and these circuits are used for the audio data signal and the data signal. It can be shared by dividing it into parts.

Effects of the Invention The present invention provides a method for automatically changing and setting the threshold value for determining whether a speech section is a speech section when cutting out a speech section, taking into account the background sound at that time. Therefore, even in noisy areas, it is possible to easily cut out voice sections.

[Brief explanation of drawings]

FIG. 1 is a block diagram showing an example of a device used in the voice section extraction method of the present invention, and FIG. 2 is a diagram showing the operation of the device shown in FIG. FIG. 1. Immediately before data renostar, 2. Absolute difference calculation circuit, 3. Addition circuit, 4. Threshold juice calculation circuit, 5. Comparator, 6. Threshold value renostar, 7. Comparator. 8...Data counter, 9...Control circuit. Figure 2

Claims (1)

[Claims] (1) A means for dividing a signal from which an audio signal is to be detected into a plurality of blocks, a circuit for calculating the sum of absolute differences in energy levels between signals in each block, and a level of the audio signal based on the sum of absolute values. 1. A speech interval detection device comprising: a circuit for calculating a threshold for determination.