JP3285178B2 - Sound signal rising detection method - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a method for detecting a rising edge of an audio signal or a voice signal, which is used for encoding a tone signal or a voice signal, voice recognition, voice communication, and the like.

[0002]

2. Description of the Related Art At present, as a method of detecting a rising edge of a sound signal in consideration of auditory characteristics, for example, a detection method using auditory entropy has been proposed in an MPEG encoding system (MPEG-Audio). ing. This means that the signal-to-noise ratio (SNR) of the minimum audible noise with respect to the input signal calculated by using the auditory model for each of a plurality of frequency sub-bands (that is, SNR that needs to be secured at the time of encoding in the sub-band. ) Is calculated, and a value called auditory entropy is calculated from the required SNR, and when this value exceeds a certain threshold value, it is detected as rising. The details of the rising edge detection method in the MPPEG audio encoding method are described in I.
It is described in SO / IEC standard IS-11172-3.

[0003]

However, in this method, when calculating the required SNR, it is necessary to find the frequency characteristics of the input signal, and a large amount of computation is required, such as performing an FFT with a dimension as large as 1024 orders over the entire band. There is a lot of processing that needs to be performed, and if it is configured as a stand-alone detection device by this method, the scale is too large.

Conventionally, no smoothing process has been performed, and in the case where the input signal is a pulse train approaching, only the first pulse should be detected and detected as the rising edge of the sound signal. Conventionally, each pulse is detected as a sound signal rising. In other words, although a series of small rises is not perceived as a rise, the conventional method has a problem that each small rise is detected as a rise of a sound signal.

[0005]

According to the present invention, there is provided the following:
In step (2), the intensity corresponding to the input sound signal is obtained. In the second step, the intensity is smoothed for a short time, that is, about 5 to 40 milliseconds to obtain a short-time smoothed signal. In the stage, the short-time smoothed signal takes a long time,
A long-time smoothed signal is obtained by performing smoothing for about 150 milliseconds, and a discontinuity of the long-time smoothed signal is calculated in a fourth stage, and the obtained discontinuous point is set as a rising point.

As the strength corresponding to the input audio signal, a linear prediction analysis of the input audio signal is performed, and the residual signal is obtained through the input audio signal to an inverse filter having a filter coefficient corresponding to the resulting linear prediction coefficient. , The intensity of the residual signal may be used. The smoothing in the third step is performed by convolving the short-time smoothed signal with the impulse response, or by passing the short-time smoothed signal through a digital detection circuit.

In the fourth step, a current long-time smoothed signal is predicted and synthesized in the fifth step from the past long-time smoothed signal, and the predicted synthesized long-time smoothed signal is subjected to the sixth step. In comparison with the actual present long-time smoothed signal, when the latter is significantly larger than the former, the discontinuous point, that is, the rising is determined. The prediction synthesis is performed by linearly summing the past long-time smoothed signals or by extrapolating the past long-time smoothed signals with a function that approximates the long-time smoothed signal.

[0008]

[Operation] In human hearing, there is a function of recognizing a fine structure and a whole structure of a sound signal pattern. For this reason, for example, a single impulse is recognized as a rising sound,
If this impulse is repeated in a fast cycle, it will not be recognized as a rising edge. According to the present invention, a nerve excitation pattern having such auditory characteristics is simulated by an auditory model having short-time smoothing and long-time smoothing, and a rise is detected by determining discontinuity of the pattern. Therefore, according to the present invention, the detection of the rising according to the auditory characteristics is performed.

When a linear prediction analysis filter is passed through an input signal, a residual signal is obtained in which discontinuous points, such as rising sounds, which cannot be predicted from past signals, are emphasized. Therefore, the detection sensitivity is higher than when the strength of the residual signal is used as the strength corresponding to the input audio signal.

[0010]

FIG. 1 shows a processing procedure in an embodiment of the present invention. In the first stage 11, the input sound signal is 32 KHz
Alternatively, it is input as a sampled digital signal of 48 KHz or the like, and as the sound intensity of the input audio signal, for example, the absolute value or square value of the input audio signal can be used. Alternatively, the input audio signal is subjected to linear prediction analysis, and a filter having a filter coefficient corresponding to the linear prediction coefficient, a so-called inverse filter, is used to obtain a residual signal through the input audio signal, and the absolute value, square value, or A signal obtained by normalizing this with the average value of the amplitude of the input signal may be used.

In a second step 12, the intensity signal from the first step 11 is windowed at a frequency of about 5 ms, for example, for about 20 ms in the past, and the average in the window is calculated and smoothed for a short time. Obtain a short-term smoothed signal. The window shape of the time window is
For example, a rectangular window, a Hamming window, a Hanning window, an exponential function window, or the like is used, and the time window is set to about 5 to 40 ms. In the third step 13, long-time smoothing is performed by convolving the short-time smoothing signal from the second step 12 with an impulse response having a fast rise and a slow fall. This smoothing is performed continuously for 20 to 150 ms. As shown in FIG. 2A, the waveform of the impulse response suddenly rises and gradually falls, and the rise and fall are linear, exponential and trigonometric, respectively. Alternatively, as shown by the equivalent circuit in FIG. 2B, the rising time constant is small and the falling time constant is large (the rising time constant may be 0). May be performed. The processing of this digital detection circuit is performed, for example, by inputting each of the short-time smoothed signal sequences x ₀ , x ₁ ,..., X _n−1 , x _n , x _{n + 1} ,. The long-term smoothed signal sequence y ₀ , y ₁ ,..., Y _n-1 , y _n , y _{n + 1} ... Of the output obtained immediately before is compared with (21), and if x _n > y _n-1 , x the _n output as _{y n (22), x n} > to y _n-1 unless _{y n-1 R (0 <} R
<Outputs a value obtained by multiplying the 1) as y _n (23).

In a fourth step 14, a prediction synthesizing step 15 for predicting the present long-time smoothed signal from the long-time smoothed signal from the past third step 13 is performed. And a comparing step 16 for comparing the output of step 13 with the output of step 13. In the prediction synthesis step 15, the prediction value of the current signal is synthesized from the past signal using a technique such as prediction synthesis using a linear sum of past signals, function fitting by a least square method or a maximum likelihood estimation method, or the like. That is, a sequence of long-smoothed signal y _n, input to the series connection of the unit delay circuit 25 only the signal time, multiplied by the constant by the multiplier 26 to the output of the respective delay circuits 25, the multiplication result The result is added by the adder 27 and output as a predicted value y _{n + 1} . Alternatively, as shown in FIG. 3B, the long-time smoothed signal y _n
Is approximated as a function 28 with respect to time, and extrapolated by the function to obtain a predicted value y _{n + 1} .

In the comparison step 16, the ratio of the current long-term smoothed signal y _{n + 1} to its predicted combined signal y _{n + 1} 'is represented by y
_{n + 1} / y _{n + 1} 'is calculated, and if this value exceeds a certain threshold value, it is determined that the signal rises, or the difference between the current signal y _{n + 1} and the predicted combined signal y _{n + 1} ' (y _{n + 1} -y
_{n + 1} ') =. DELTA.y, and if this value .DELTA.y exceeds a certain threshold value, it is determined as rising. Alternatively, the current signal yn _{+ 1} and the predicted synthesized signal are multiplied by a constant A
The difference from y _{n + 1} ′ is calculated, and if this value exceeds a certain threshold value, a rising method is used.

[0014]

As described above, according to the present invention, the excitation pattern of the nerve is simulated by the short-time smoothing and the long-time smoothing, and this signal is used for the judgment of the rising edge. It is possible to detect a rising edge.

[Brief description of the drawings]

FIG. 1 is a flowchart showing a procedure of an embodiment of the present invention.

2A is a diagram showing an example of the shape of an impulse response used when performing convolution in long-time smoothing, FIG. 2B is a diagram showing an equivalent circuit of a digital detection circuit used for long-time smoothing, and FIG. 6 is a flowchart showing a procedure for applying to a discrete signal.

FIG. 3A is a diagram illustrating a configuration example when predictive synthesis in continuity detection is obtained by a linear sum of past long-time smoothed signals, and FIG. 3B is an example of a function that approximates past long-time smoothed signals; FIG.

Continuation of the front page (56) References JP-A-61-46999 (JP, A) JP-A-3-253899 (JP, A) JP-A-4-67200 (JP, A) JP-A-6-19498 (JP) , A) JP-A-6-27995 (JP, A) (58) Fields investigated (Int. Cl. ⁷ , DB name) G10L 11/02, 21/02

Claims (9)

(57) [Claims] 1. A first step of obtaining an intensity corresponding to an input audio signal, and a second step of performing short-time smoothing on the intensity obtained in the first step to obtain a short-time smoothed signal. And a third step of performing a long-time smoothing on the short-time smoothed signal obtained in the second step to obtain a long-time smoothed signal, and a third step obtained in the third step. Calculating the discontinuity of the long-time smoothed signal, and setting the obtained discontinuous point as a rising edge. 2. A linear prediction analysis of the input audio signal, a residual signal is obtained through the input audio signal through a filter having a filter coefficient corresponding to a linear prediction coefficient of the result, and the intensity of the residual signal is calculated by 2. The sound signal rising detection method according to claim 1, wherein the intensity is used as an intensity corresponding to the input audio signal in the first stage. 3. The method according to claim 1, wherein the third step is to convolve an impulse response with the short-time smoothed signal obtained in the second step to perform smoothing. The described sound signal rising detection method. 4. The method according to claim 1, wherein the third step is to smooth the short-time smoothed signal obtained in the second step by passing it through a digital detection circuit. Or the sound signal rising detection method according to 2. 5. The fourth step is a fifth step of predictively synthesizing a current long-time smoothed signal from the past long-time smoothed signal obtained in the third step, and the fifth step thereof. A sixth step of comparing the predicted and synthesized long-term smoothed signal obtained in the step with the actual long-term smoothed signal at the present time, and judging that the signal rises when the latter is significantly larger than the former. The sound signal rising detection method according to any one of claims 1 to 4, wherein: 6. The sound signal rising detection method according to claim 5, wherein in the fifth step, prediction synthesis is performed by linearly summing past long-time smoothed signals. 7. The sound signal rising detection method according to claim 5, wherein in the fifth step, prediction synthesis is performed by extrapolating a past long-time smoothed signal with a function that approximates the signal. 8. The sixth step comprises: combining the current long-term smoothed signal obtained in the third step with the predicted and synthesized long-time smoothed signal obtained in the fifth step. 8. The sound signal rising detection method according to claim 5, wherein a ratio is calculated, and when the ratio exceeds a certain threshold value, it is determined that the sound signal rises. 9. The sixth step is to convert the current long-term smoothed signal obtained in the third step and the predicted and synthesized long-time smoothed signal obtained in the fifth step into a constant. 8. The sound signal rising detection method according to claim 5, wherein a difference from the multiplied signal is calculated, and if the difference exceeds a certain threshold value, it is determined that the sound signal rises.