JPH0432899A - Pitch detector for sound signal - Google Patents

Abstract

PURPOSE:To efficiently detect an accurate reference pitch free from misdeetction by detecting the feature elements of waveforms in respective parts while successively detecting an intermediate point and an end point corresponding to the intermediate point and detecting the reference pitch of a sound signal from a fuzzy output corresponding to the detection of the feature element. CONSTITUTION:The starting point of waveform pitches of a sound signal is detected by a pitch starting point detecting means 2, the feature elements of waveforms in respective parts between the start point and an intermediate point and between the intermediate point and an end point are detected by a waveform detecting means 3 and fuzzy inference processing is developed by a fuzzy inference processing means 4 by using the feature elements of the waveforms detected by the means 3 as inputs. In this case, the feature elements of the waveforms of respective parts are detected while successively moving the intermediate point and the end point corresponding to the intermediate point and the reference pitch of the sound signal is detected by the fuzzy output of the means 4 corresponding to these feature elements. Consequently, an accurate pitch free from misdetection can be efficiently detected.

Description

DETAILED DESCRIPTION OF THE INVENTION [Field of Industrial Application] The present invention relates to an audio signal pitch detection device for detecting the fundamental period (pitch) of an audio signal, which is used for audio compression and the like.

[Prior Art] For example, voice mail is known as a system for sending and distributing a person's voice as a message. By the way, since voice messages in such systems have a large amount of information, the voice signals are compressed and encoded, stored, and then transferred.

Therefore, when compressing and encoding an audio signal, it may be necessary to detect the basic pitch of the audio signal every predetermined section. In such a case, find the pitch start point of the audio signal for each section, shift one sample point at a time based on this start point, calculate the autocorrelation relationship, and detect the pitch at the point where this relationship is minimum. A method is being considered.

[Problem to be solved by the invention] However, calculating the autocorrelation for each sample point based on the starting point in this way requires a huge amount of calculation, which is time consuming and inefficient. There was a tendency for detection errors to occur.

The present invention has been made in view of the above circumstances, and an object of the present invention is to provide an audio signal pitch detection device that can accurately and efficiently detect the basic pitch of an audio signal.

[Means for Solving the Problems] The present invention detects a pitch start point on the waveform of an audio signal using pitch start point detection means, and detects a pitch start point at a predetermined distance from the start point based on the pitch start point detected by the pitch start point detection means. The change point is set as the intermediate point, and the end point is set based on the distance from the starting point to the intermediate point, and characteristic elements of the waveform of each part are detected from the starting point to the intermediate point and from the intermediate point to the ending point using a waveform detection means. Then, the fuzzy inference processing means develops fuzzy inference processing using the characteristic elements of the waveform detected by the waveform detection means as input, and each of the above points is sequentially moved between the intermediate point and the end point corresponding to the intermediate point. In addition to detecting the characteristic elements of the waveform of the body part, the basic pitch of the audio signal is detected based on the fuzzy output of the fuzzy inference processing means corresponding to these characteristic elements.

[Function] As a result, according to the present invention, while sequentially moving the intermediate point and end point set with respect to the pitch start point on the waveform of the audio signal, the intermediate point from the start point and the end point from the intermediate point are detected respectively. Since the fuzzy inference is developed based on the characteristic elements of the waveform of each part, the basic pitch determined from the fuzzy output can be efficiently obtained without any detection errors.

[Example] An embodiment of the present invention will be described below with reference to the drawings.

FIG. 1 shows the circuit configuration of the same embodiment. In the figure, reference numeral 1 denotes an audio signal capture unit, and the second audio signal capture unit 1 captures an audio signal of a predetermined length.

The audio signal captured by the audio signal capture section 1 is sent to the pitch start point detection section 2. This pitch start point detection section 2 is
Detecting a starting point for basic pitch extraction of an audio signal.

The outputs of the audio signal acquisition section 1 and pitch start point detection section 2 are as follows.
The signal is sent to the waveform detection section 3.

The waveform detection unit 3 searches for a first valley separated from the pitch start point detected by the pitch start point detection unit 2 by the necessary minimum distance and a last valley separated by the necessary maximum distance from the start point, and sets the first valley as the intermediate point. The last valley is set as the final point, and a distance equal to the distance from the starting point to the midpoint is taken from the midpoint first, and this is set as the final point. Then, the number of peaks and the number of dog peaks, which are characteristic elements of the waveform, are detected for each section from the start point to the middle point and from the middle point to the end point, and the value of the most dog peak between the start point and the middle point, and the starting point The value of the peak located at the distance from the middle point to the end point is detected, and the value of the valley at the starting point and the valley value at the middle point are further detected. Here, Inuyama originally indicates the envelope of the mountain after removing small mountains through LPF, but this time, for the sake of simplification, it is defined as the height from the valley or between 0 and 8C+p (maximum - minimum ) 73 or higher is considered to be an inuyama, and the location where the maximum value is reached is considered to be the mountain position.

The output of the waveform detection section 3 is arranged into the input form necessary for fuzzy inference and sent to the fuzzy inference processing section 4. The fuzzy inference processing unit 4 uses the production master rule as follows: (1) If the valley value at the starting point and the valley value at the first intermediate point are about the same, then they are very similar. (2) If the number of mountains in the headquarters and the number of mountains in the comparison section are about the same, they are similar. (3
) If the cycle is long, I can't say anything. (4) If the number of Inuyama at the headquarters is different from the number of Inuyama at the comparison department,
Does not resemble. (5) If the peak of the maximum value at the headquarters and the peak of the comparison section at the same position are different, it is not a lie. The fuzzy inference is developed according to these rules.

Then, the output of the fuzzy inference processing section 4 is sent to the bisochi output section 5. The pitch output section 5 detects and outputs the basic pitch of the audio signal from the number of points from the starting point to the intermediate point when the output of the fuzzy inference processing section 4 is maximum.

Next, the operation of the embodiment configured in this way will be explained.

Now, when an audio signal of a predetermined length as shown in FIG. Becomes sampled.

In this state, the flowchart shown in FIG. 3 is executed.

First, in step A1, a pitch starting point is detected.

In this case, in the pitch start point detection section 2, as shown in FIG.
As shown in the figure, each point of 79p from the beginning Op of the audio signal is sampled, and the deepest valley is set as the pitch starting point S for pitch extraction. Here, the reason why the sample points from Op to 79p are designated to detect the deepest valley is that if this range is set, the deepest valley can be detected reliably.

Next, the process proceeds to step A2, and as shown in FIG. 2(b), the waveform detection unit 3 detects the first trough at the minimum necessary distance from the pitch starting point S, for example, 20p or more, and the maximum necessary distance. For example, find the last valley within 144p and use the first valley as the intermediate point SI.

Set the last valley to the final point E. In this case, the reason we searched for the first valley was 20p or more because there was no pitch shorter than 20p, and the reason we searched for the last valley was within 144p because there was no pitch longer than 144p. It is. Further, the process proceeds through step A3 to step A4, and as shown in FIG. 2(C), a distance equal to the distance from the starting point S to the intermediate point Sl is taken from the intermediate point Sl first to set the ending point El.

Then, the process proceeds to step A5, and the number of mountains and the number of dog mountains are detected for each section from the starting point S to the intermediate point Sl and from the intermediate point Sl to the ending point EI.

Next, proceed to step A6, and calculate the value of the closest mountain between the starting point S and the intermediate point Sl, and the value of the mountain located at the distance from the starting point S to the closest mountain in the force direction from the intermediate point Sl to the final point EI. , and further detect the valley value at the starting point S and the valley value at the intermediate point SI.

Next, the process proceeds to step A7, where an input normalized to "1", which is an input for fuzzy inference, is calculated for the output of the waveform detection section 3, and this is provided to the fuzzy inference processing section 4. As a result, in step A8, fuzzy inference processing is developed in accordance with the production master rule described above. In this case, if we assume that the area between the starting point S and the intermediate point Sl is the headquarters, and the area between the intermediate point Sl and the ending point El is the comparison area, then
), the number of peaks in the headquarters and comparison section is the same, but the value of the valley at the starting point S, the value of the valley at the intermediate point Sl, the number of peaks at the headquarters and comparison section, and the maximum value at the headquarters The values of the peaks of the comparing section located at the same position as the peaks of are all different, and fuzzy outputs corresponding to these values can be obtained. Note that fuzzy inference processing in accordance with the production master rule is a well-known technique, so a description thereof will be omitted here.

Then, the output from the fuzzy inference processing section 4 is sent to the pitch output section 5. In the pitch output section 5, step A9
Determine whether the fuzzy output is the maximum ever. If it is determined to be YES here, the process proceeds to step AIO, where the fuzzy output maximum value is rewritten, and further, the pitch value is rewritten in step All, and the process returns to step A2. On the other hand, if it is determined NO in step A9, immediately proceed to step A2.
will return to .

In step A2, as shown in FIG. 2(d), the valley next to the valley of the intermediate point SI is set as a new intermediate point SI+1. As a result, a new end point EI+1 is set in step A4 according to the intermediate point SI+1, and the operations from step A5 onwards are subsequently executed. In this case, if the fuzzy output is the maximum ever, the fuzzy output maximum value is rewritten in step AIO, and the pitch value is rewritten in step All.

Thereafter, the above operation is repeated while moving the intermediate point one unit at a time, and then the end point E+
When n reaches the final point E, the answer is YES in step A3, and the process proceeds to step AI2.

In step A12, a pitch value corresponding to the maximum fuzzy output value up to now is extracted, and this value is output as the basic pitch of the audio signal.

FIG. 2(e) shows a detection state in which the fuzzy output is maximum, and the intermediate point is set at SI+m and the end point is set at E I 10 m. In this case, the number of mountains in the headquarters and comparison section,
The trough value at the starting point S and the trough value at the intermediate point SI, the number of peaks at the headquarters and the comparison section, and the value at the peak at the comparison section at the same position as the peak at the headquarters are all the same.

Therefore, in this way, the number of peaks and dog peaks of the waveform are detected for each section from the start point to the middle point and from the middle point to the end point, based on the first detected start point, and Detect the value of the maximum weight between the points and the value of the peak located at the distance from the start point to the maximum weight from the middle point to the end point, and then calculate the value of the valley at the start point and the value of the valley at the middle point. , and develops fuzzy inference using these feature elements.Furthermore, since this kind of fuzzy inference is executed while moving the intermediate points from valley to valley, the basics required from the fuzzy output Accurate pitches without detection errors can be efficiently obtained. Furthermore, since the starting point is determined at the trough of the waveform, there is consistency with other parts, which also has the advantage of making it easier to analyze the waveform.

Note that the present invention is not limited to the above-mentioned embodiments, but can be implemented with appropriate modifications within the scope without changing the gist.

For example, in the above embodiment, the pitch starting point is detected by detecting the valley, but the pitch starting point may be determined by detecting the peak. In addition, the feature elements detected in each section from the start point to the middle point and from the middle point to the end point are - examples, and some of these may be used, or feature elements other than these may be detected. It may also be used as

In addition, the pitch starting point does not necessarily have to be the maximum depth or a nearby valley. If so, you can say, ``They are similar.''

[Effects of the Invention] The present invention detects a pitch start point on the waveform of an audio signal using a pitch start point detection means, and detects a change point a predetermined distance away from the start point based on the pitch start point detected by the pitch start point detection means. At the same time, the end point is set based on the distance from the start point to the midpoint, and the waveform detecting means detects characteristic elements of the waveform of each part from the start point to the midpoint and from the midpoint to the end point. The fuzzy inference processing means uses the characteristic elements of the waveform detected by the detection means as input to develop fuzzy inference processing, and the intermediate point and the end point corresponding to the intermediate point are sequentially moved while each of the above parts is In addition to detecting the characteristic elements of the waveform, the basic pitch of the audio signal is detected by the fuzzy output of the fuzzy inference processing means according to these characteristic elements, so that the pitch starting point on the waveform of the audio signal is While sequentially moving the intermediate point and end point set for the target, fuzzy inference can be developed based on the characteristic elements of the waveform of each region detected from the starting point to the intermediate point and from the intermediate point to the end point, and the fuzzy inference can be This means that accurate basic pitches without detection errors can be efficiently obtained.

[Brief explanation of the drawing]

FIG. 1 is a block diagram showing the circuit configuration of an embodiment of the present invention, FIG. 2 is a waveform diagram for explaining the operation of the embodiment, and FIG. 3 is a diagram for explaining the operation of the embodiment. This is a flowchart for 1... Audio signal capture unit, 2... Pitch start point detection unit, 3... Waveform detection unit, 4... Fuzzy inference processing unit,
5...Pitch output section. Applicant's agent Patent attorney Takehiko Suzue

Claims (1)

[Claims] Pitch start point detection means for detecting a pitch start point on a waveform of an audio signal; and a pitch start point detection means for detecting a pitch start point on a waveform of an audio signal; Waveform detection means for setting an end point based on the distance to the point and detecting characteristic elements of the waveform of each part from the start point to the intermediate point and from the intermediate point to the end point, and the characteristics of the waveform detected by the waveform detection means fuzzy inference processing means that develops fuzzy inference processing using elements as input; detects characteristic elements of the waveform of each region while sequentially moving the intermediate point and the end point corresponding to the intermediate point; A pitch detection device for an audio signal, characterized in that the basic pitch of the audio signal is detected by the fuzzy output of the fuzzy inference processing means according to the element.