JPS62227200A - Zero crossing number detection - 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 method for detecting the number of zero crossings, which is a characteristic parameter in a speech recognition device.

(Prior Art) Conventionally, an audio power value and a frequency analysis are calculated from a digital audio signal obtained by A/D conversion of an analog audio signal, and a frame of a predetermined length of this digital audio signal is calculated. A method has been proposed in which the number of zero crossings between this digital audio signal and a predetermined reference value is detected as a feature parameter, and speech recognition processing is performed based on these audio power values, frequency analysis results, and the number of zero crossings. There is.

Before explaining the present invention, this conventional method for detecting the number of zero crossings will be explained.

FIG. 7 is a block diagram showing an example of the configuration of a conventional speech recognition device, in which 11 is a speech input terminal, 12 is an amplifier, and 13 is a low-pass filter (LPF). 14 is an A/D converter in which analog input audio is converted into a digital audio signal. 15 is a frequency analysis section that performs frequency analysis of the human voice signal from the A/D converter 14. Reference numeral 16 is a zero-crossing number calculation unit that counts the number of times the digitized human voice signal crosses a zero level, which is a reference value, that is, the number of zero-crossings. Reference numeral 17 denotes an audio power calculation unit that calculates audio power. Reference numeral 18 denotes a voice section detection unit that detects a section in which a voice exists in a manually input signal (hereinafter referred to as a voice section) from the frequency analysis result and the voice power. Reference numeral 19 denotes a recognition unit which receives the frequency analysis result from the frequency analysis unit 15; the number of zero crossings from the crossing number calculation unit 16, the voice power from the voice power calculation unit 17, and the voice interval detection signal from the voice interval detection unit 18. Each receiver is configured to perform recognition processing on the input audio signal and output the recognition result to the output terminal 20.

FIG. 8 is a block diagram showing an example of the configuration of the conventional zero-crossing number calculation section 16 in the speech recognition device of FIG. 7, and this conventional zero-crossing number detection method will be explained.

The number of zero crossings is the number of times the audio signal crosses the zero level within a frame period. Whether or not the audio signal crosses the zero level at a certain point in time can be detected by calculating the value S1 (where i is the frame number) obtained by the following equation.

S,=x, ■Xi-However, x,=0: The value of xH is positive or 0=1: The value of xl is negative. Here, ■ is an operation called exclusive OR. This calculation yields 0 if the two human power values match, and 1 if they do not match. Therefore, the number of zero crossings is calculated by counting the number of times the value S1 becomes 1 during a predetermined frame period.

FIG. 8 is a block diagram showing an example of a structure in which the principle of calculating the number of zero crossings is applied to the zero crossing number calculation section of FIG. 7. In the figure, xl is a sign bit corresponding to the sample value X+ of the digital audio signal from the A/D converter 14 (hereinafter sometimes simply referred to as audio signal Xi). Also, C is the sampling clock of the audio signal X1,
F is a frame clock for synchronizing with a predetermined frame length and frame period. 21 is a shift register, 22 is an anti-coincidence circuit, 23 is a counter, and 24 is a register.

According to the configuration shown in FIG. 8, it is assumed that the counter 23 is reset at the start of a certain frame by the frame clock F. The output Z i − of the counter 23 immediately before that time
1 is stored in register 24. In the anti-coincidence circuit 22,
An exclusive OR operation S1 is performed between the sign bit x1 and the sign bit X1-1 stored in the shift register 21 at the previous sampling time. The counter 23 counts the number of sampling clocks C while this value S remains at 1, and the value immediately before the next frame clock F is input is given as the number Z of zero crossings in that frame period. This value Z is stored in the register 24 and becomes an input to the recognition unit 19 in FIG.

Figures 9(A) and (B) and Figures 10(A) and (B)
) is the case where no noise is added to the silent section, fricative S and vowel I using this conventional zero crossing number detection method, and
It is an explanatory diagram of the number of zero crossings Z obtained when low-level noise is added, and (
A) shows how the audio signal X of the A/D converter 14 changes, and (B) shows the number of zero crossings 2 corresponding thereto.

As can be understood from FIGS. 9(A) and (B), the number of zero crossings Z shows a small value in the silent section and the vowel I section, but it shows a very large value in the fricative S section.

However, in the examples shown in FIGS. 10(A) and 10(B), in which it seems that some noise is added to the human signal, the number of zero crossings Z shows a large value even in the silent section. Thus, it can be seen that this conventional detection method with a zero crossing count of 2 is greatly affected by even low level noise.

In order to remove the influence of this noise, instead of measuring the number of times the human voice signal crosses the 7 level, which is the reference value in this case,
A method has been proposed for determining the number of times a given reference value is crossed.

For this detection method, the zero crossing number calculation unit 16 shown in FIG.
Explain with reference to. In this example, the A/D converter 14
A level obtained by subtracting a certain constant from the zero level of the audio signal X+ from is set as a reference level for zero crossing. This setting process can be realized by subtracting the offset value from the audio signal x1. This subtraction process is performed by the subtraction circuit 25. Let the value of the audio signal after the subtraction be X'i. In this case, the level that serves as a reference for zero crossing is called an offset. and,
As in the case of the conventional calculation method described above, the value S1 of the exclusive OR of the current sign bit x1 of the value x'1 and the sign bit x1-1 at the previous sampling clock.
seek. Also in this case, this value S.

Count the number of intervals where Z is 1 and calculate the value immediately before the next frame clock input as the zero crossing number Z.
shall be.

This offset audio signal
Examples of ' are shown in FIGS. 12(A) and 12(B), respectively. The broken line in the figure indicates an offset value that serves as a reference for intersection. As is clear from this example, the value of the number of zero crossings 2 in the silent section is indeed a value where the influence of noise has been removed, but the magnitude of the characteristic value in the part of the fricative S is also low. Therefore, the sensitivity as a parameter indicating voice characteristics has decreased.

(Problems to be Solved by the Invention) As described above, one detection method of the prior art is susceptible to noise, and the other detection method has a problem in that the sensitivity of voice characteristic parameters is reduced.

An object of the present invention is to provide a method for detecting the number of zero crossings that eliminates the influence of noise in the conventional method for detecting the number of zero crossings and has sufficient sensitivity as a characteristic parameter of speech.

(Means for solving the problem) In order to achieve this objective, according to the zero crossing number detection method according to the present invention, the predicted voice power value of the current frame is calculated based on the voice power value of one frame before. and setting a reference value for detecting the number of zero crossings based on the obtained voice power predicted value.

In carrying out the present invention, the relationship between the predicted voice power value and the offset value from the zero level value of the digital voice signal is stored in advance in a table ROM, and the predicted voice power value is used as an address to correspond from the table ROM. It is preferable to read out an offset value and use the read offset value as a reference value for detecting the number of zero crossings.

(Function) As described above, the present invention adds a process of setting a standard value (also referred to as an IJI value), which is a reference for crossing, to an appropriate value in advance according to the audio power level as a method for calculating the number of zero crossings. It is something.

In this way, when the voice power is relatively small and the noise proportion is large, the reference value is set high to remove the influence of noise, while when the voice power is large and the noise proportion is small, the reference value is set high and the noise effect is removed. Since the reference value is set small, the number of zero crossings detected by the present invention is not affected by noise and has a sufficiently high sensitivity as a voice characteristic parameter.

(Example) Hereinafter, an example of the zero crossing number detection method of the present invention will be described with reference to the drawings.

FIG. 1 is a block diagram showing an embodiment of the present invention,
Components that are the same as those shown in FIG. 7 are denoted by the same reference numerals, and detailed explanation thereof will be omitted.

The difference between the zero-crossing number detection method of the present invention and the conventional one is that the zero-crossing calculation unit 16 sets a reference value (threshold) for calculating the number of zero-crossings in correspondence with the audio power value generated from the audio power calculation unit 17. It is at the point where the value is taken. For example, a predicted voice power value for the current frame is determined based on the voice power value up to one frame before, and a reference value for detecting the number of zero crossings is set based on the determined voice power predicted value.

Therefore, as shown in FIG. 1, in the speech recognition device for carrying out the present invention, the input of the zero-crossing number calculation unit 16 includes the audio signal from the A/D converter 14, the sampling clock, the frame clock, and other signals. The configuration is such that the audio power signal generated from the audio power calculation section 17 can be supplied to the audio power calculating section 17.

FIG. 2 shows a block diagram of an example of the configuration of the zero crossing number calculating section 16 for carrying out the present invention. In addition, in Figure 2,
Components that are the same as those shown in FIG. 11 are designated by the same reference numerals. In the figure, the audio signal number 8 Xi, the level value Pj of the audio power signal (referred to as audio power level), the sampling clock C, and the frame clock F are shown. In this example, the audio power level P
Let J (j is the frame number) be the audio power level of the current frame predicted from the audio power level up to one frame before, that is, the audio power predicted value.

In this case, the value of the audio power level pJ is a simple method of using the average of the audio power of the previous few frames as the audio power prediction value, as given by the following equation (1), because the audio power does not change rapidly. However, there is no problem in implementation.

However, Pk is the value of the audio power in the second frame.

Such a predicted value can be obtained by a conventionally known method, and a means for obtaining this predicted value may be provided in the audio power calculation section 17 or within the zero crossing number calculation section 16.

Next, a method for determining the offset value from the zero level of the digital audio signal from the audio power predicted value obtained in this manner will be explained. In the illustrated embodiment, a table ROM (read only memory hereinafter referred to as ROM) 26
Use. This table ROM 26 has previously learned and stored the relationship between the predicted voice power value and the offset value. Then, the predicted voice power value Pj obtained from the voice power Pj of the voice power calculation unit 17 is stored in the table ROM 26.
A corresponding appropriate offset value θ is read from the table ROM 26 as an address for the subtraction circuit 25, which is configured as a subtracter in this case.
Output to.

In this subtraction circuit 25, an offset value θ is calculated from the digital audio signal Xi from the A/D converter 14 in FIG.
1 is subtracted. Therefore, the anti-coincidence circuit 22 has (Xi
The sign bit X+ of -〇J) and the sign bit x, -1 of (Xi-1-〇j) delayed by one clock of the sampling clock C by the shift register 21 are input. The output S″1 of the anti-coincidence circuit 22 has sign bits X, and x
, -1, that is, when the digital audio signal Xi and the digital audio signal x, -1 one frame before have different signs, the reference offset value θ4 is set at this point. Since this indicates that the digital audio signal crosses the , the value is 1 only in this case.

The counter 23 counts the number of times ZhI that the value S'1 becomes 1 during the interval determined by the frame clock F. The register 24 holds the count value Z''1 of the counter 23 and the value immediately before the frame is input, that is, the number of zero crossings Z'' in that frame period, and is input to the recognition unit 19 in FIG.

In addition, in this embodiment, the table ROM 26 shown in FIG.
A simple two-stage threshold value is used as the offset value θ4 as a reference value to be stored in . FIG. 3 shows a table RO that stores the relationship between this predicted voice power value PJ and the reference value θ1.
It is a diagram showing the contents of M26. Also, Figure 4 shows table R
A voice power predicted value P stored in the OM26,
FIG. 3 is a graph showing the relationship between the reference value θ. Here, the reference value θLO when the voice power predicted value pj is less than the value P0
W is set lower than the reference value θHIGH when the predicted voice power value PJ is greater than the value P0.

These values can be set to appropriate values by learning in advance. The reason for this setting is that when the audio power is relatively small, the influence of noise that is included in the input waveform is large, so it is necessary to set the reference value high.On the other hand, when the audio power is large, the percentage of noise that is included in the input waveform is This is because the reference value can be set close to the ideal zero level since the value is relatively low.

Next, examples obtained by calculating the number of zero crossings based on this example are shown in FIGS. 5(A)-(C) and FIGS. 6(A)-(C). Figures 5 (A), (B), and (C) are average power predicted values PJ for a certain input voice without noise.
(solid line diagram), input audio sample value x+ (dashed line diagram)
and the number of zero crossings Z'' (shown as a solid line).Furthermore, FIGS. 6(A), (B), and (C) show the PJ for another input voice with low noise (shown as a solid line).
, Also, Z indicated by a dotted line is the number of zero crossings calculated by a conventional method without using the present invention.

As can be understood from these experimental results, in any case, the shadow of noise can be effectively suppressed in the silent section, and the characteristic value of the number of zero crossings in the voiced section of the fricative S and the vowel I can be suppressed. is large, and therefore has high sensitivity as a parameter indicating voice characteristics.

This invention is not limited only to the embodiments described above, and many modifications and changes can be made without departing from the scope of this invention. For example, in the above-described embodiment, a case was explained in which the value θ4, which is a reference for crossing, is obtained from the audio power value PJ using the table ROM2fi, but instead of using this table ROM, a RAM (random access・Memory) may be used, and appropriate values may be written in this RAM at the time of initial setting or when updating recognition parameters, etc.

Further, in the above-described embodiment, a two-stage threshold value is used as the offset value θ1, but it is clear that the present invention is not limited to this, and three or more stages of threshold values may be used.

(Effects of the Invention) As is clear from the embodiments described above, according to the method for detecting the number of zero crossings of the present invention, it is possible to eliminate the influence of noise, and also to detect zeros that have a sufficient sensitivity as a characteristic parameter. The number of intersections can be detected.

[Brief explanation of drawings]

FIG. 1 is a block diagram showing a speech recognition device to which the method for detecting the number of zero crossings of the present invention is applied, and FIG. 3 and 4 are diagrams respectively showing the contents of the table ROM and offset values used to explain the method for detecting the number of zero crossings of the present invention. FIGS. 5 and 6 are diagrams showing the number of zero crossings of the present invention An explanatory diagram for explaining the number of zero crossings obtained by the detection method. Figure 7 is a block diagram showing a conventional speech recognition device to which the conventional method of detecting the number of zero crossings is applied. Figure 8 is a diagram showing the conventional method of calculating the number of zero crossings. FIG. 9 and FIG. 10 are explanatory diagrams for explaining the number of zero crossings obtained using the conventional zero crossing number calculating section shown in FIG. 8. FIG. FIG. 12 is a block diagram showing the zero-crossing number calculation section. FIG. 12 is an explanatory diagram for explaining the number of zero-crossings obtained using the zero-crossing number calculation section of FIG. 11-...Input terminal, 12-...Amplifier 13-...Low pass filter, 14...A/D converter 15-...
Frequency analysis section, 16-...Zero crossing number calculation section 17...
・Voice power calculation unit, l 8-...Voice section detection unit 19
---Recognition unit, 20... Output terminal 21--
・Shift register, 22-... anti-match circuit 23-...
Counter, 24--Register 25--Subtraction circuit, 26-Table ROM. itanshi θ ' ``j ' PMAxθ
H1 (yH < θLOVJ - is based on the voice power 13a) Fig. 4

Claims (2)

[Claims] (1) Detect the audio power value from the digital audio signal obtained by A/D converting the analog audio signal, and convert the digital audio signal and the reference value for each frame of a predetermined length of the digital audio signal. In order to detect the number of zero crossings between the voice power values and the number of zero crossings, and to perform voice recognition based on the voice power values and the number of zero crossings, A method for detecting the number of zero crossings, comprising: obtaining a predicted voice power value of a frame, and setting the reference value based on the predicted voice power value. (2) The relationship between the predicted voice power value and the offset value from the zero level value of the digital voice signal is stored in a table ROM in advance, and the corresponding offset is stored from the table ROM using the predicted voice power value as an address. 2. The method for detecting the number of zero crossings according to claim 1, characterized in that the offset value is read out and the offset value is used as the reference value.