JPS5931079B2 - Pituchi extractor - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION Field of Industrial Application The present invention relates to a device for removing harmonic components from an audio signal and extracting the fundamental frequency component (also referred to as pitch) of the audio signal. It can be used as a visual display device for intonation correction.

Conventional configuration and its problems Conventionally, as a method for extracting the fundamental frequency from an audio signal in real time, the Hasey method, which uses multiple low-pass filters, and repeated rectification, integration, and differentiation of the audio signal, have been used. There is a method (Gruen 2 type pitch extraction device) that removes harmonic components caused by vocal tract resonance and obtains a pulse train synchronized with pitch, but the fundamental frequency of the voice normally fluctuates by 4% per period even for the same person, and can vary by up to 10%. It can vary by as much as ~20%.

Furthermore, the fundamental frequency for men, women, and children is 80H2 to 5.
It is distributed over a wide range up to 00H2, and the above method has the inconvenience of having to manually switch the filter passband and integration time constant depending on the vocal range of men, women, and children, and the speed of change in the fundamental frequency. Detection of harmonic components other than the fundamental frequency that the circuit cannot follow (referred to as double pitch)
It has the disadvantage that it often causes Purpose of the Invention The purpose of the present invention is to eliminate the need for manual switching of time constants and bands over a wide range of vocal ranges, and to minimize the occurrence of double pitches even with rapid fluctuations in the fundamental frequency.
An object of the present invention is to provide a pitch extraction device capable of extracting fundamental frequencies in real time.

Structure of the Invention The pitch extracting device of the present invention, when the voice start/silence detecting section detects the rise of the voice, passes the voice signal through a plurality of overlapping low-pass filters with mutually shifted passbands.
The cycle of each output is measured for a certain period of time by the cycle detection unit, the maximum cycle is detected from the plurality of cycles by the maximum cycle detection unit, this is regarded as the pitch cycle of the input voice, and the pitch (basic frequency).

Each filter control section generates a control signal based on the detected pitch value, and feeds it back to each low-pass filter, thereby controlling the pass band of the filter to be optimal for pitch extraction. After that, until the audio input ends, the pitch of the audio is detected one after another and the band of each of the above filters is changed in response to the change, so that the band of each filter is always maintained at the optimal level even if the pitch changes significantly. , Accurate pitch extraction with less occurrence of double pitches can be performed in real time, and even if the speakers are different, pitches can be extracted without changing constants or the like. DESCRIPTION OF EMBODIMENTS Hereinafter, embodiments of the present invention will be described with reference to the drawings.

FIG. 1 is a block diagram of a pitch extraction apparatus in one embodiment of the present invention.

In Fig. 1, 1 is a low-pass filter 1 for removing harmonic components from the audio and extracting only the pitch.
Its passband is variable depending on external signals. Such a low-pass filter is, for example, a voltage-controlled low-pass filter that changes the forward transfer conductance of the amplifier by an external voltage input, or a voltage-controlled low-pass filter that changes the passband by changing the clock frequency from an external source. It can be easily constructed with a switched capacitor type low-pass filter. 2 and 3 are low-pass filters similar to the low-pass filter 1, and when there is no sound, the passbands are different as shown in FIG. 2, and the cut-off frequencies are FCl, fC2, and f.
It has become C3.

Reference numeral 4 denotes a filter control unit that controls the passband of the low-pass filter 1. For example, if the low-pass filter 1 is a voltage control type, a DC amplifier that generates a DC voltage proportional to the pitch output is used. If the low-pass filter is a switched capacitor type, it is composed of a voltage controlled oscillator that generates a clock pulse with a frequency proportional to the pitch output.

Reference numerals 5 and 6 are filter control units similar to 4, which control low-pass filters 2 and 3, respectively.

Reference numeral 7 denotes a period detection section for detecting the period of the output waveform of the low-pass filter 1, and is composed of, for example, a counter circuit and a digital-to-analog conversion circuit.

Reference numerals 8 and 9 denote period detection sections similar to the period detection section 7, which detect the period from the output waveforms of the low-pass filters 2 and 3.

Reference numeral 10 denotes a maximum period detection section which detects the maximum period among the periods obtained from the period detection sections 7, 8, and 9.

Reference numeral 11 denotes a reciprocal calculation section, which performs reciprocal calculation on the output of the maximum period detection section 10 and converts the period into a frequency value.

For example, if the output of the maximum period detection section 10 is T, it is configured with a divider that calculates 1/T. Reference numeral 12 denotes a speech/non-speech detection section, which is composed of, for example, a rectifier/integrator circuit and a threshold circuit, and detects the start and end of speech to control the start/stop of the pitch extraction algorithm.

Figure 2 shows the low-pass filters 1, 2, and
FIG. 3 is a characteristic diagram showing the pass band during silence of No. 3;

Here, FCl, fC2, and fC3 indicate the cutoff frequencies of low-pass filters 1, 2, and 3, respectively, and in order to accurately detect the initial pitch at the rise of the voice, FC2
There is a relationship of C3, fC1<2fc2, and the configuration is such that the second and third harmonics are removed and only the fundamental frequency is extracted. FIG. 3 shows the signals of each part of the pitch extraction device shown in FIG. indicates the output of the low-pass filters 2 and 3.

e is the output of the reciprocal calculation unit 11 and indicates the pitch output. f is the cutoff frequency FCl, f of the low-pass filters 1, 2, and 3;
It shows the trajectories of C2 and fC3. In the above configuration, for example, FCl=400Hz, fc2-250
Hz, fc3=150Hz, pitch is FC2 and FC
3 (for example, 200 Hz), first, the sound/silence detection unit 12 detects the rise of the sound, and for a certain period of time t, the low-pass filters 1, 2, The cutoff frequencies of 3 are respectively FC, , fC2
, fC3 to search for the initial pitch value. As a result, the low-pass filters 1, 2, and 3 have the following effects during time t:
Waveforms as shown in FIGS. 3B, c, and d are produced, respectively.

Since the pitch of the input audio is between FC2 and FC3, the output of the low-pass filter 1 with the highest cut-off frequency is a waveform b mixed with harmonics for a certain period of time t.
The output of the low-pass filter 12 is a waveform c from which harmonics have been removed, and the input sound to the low-pass filter 13 is greatly attenuated because it is out of the passband of the filter 13. The output will appear.

The cycle detecting sections 7, 8, 9 include low-pass filters 1, 2.
, 3, the period of each waveform is measured for a certain period of time t, and the values are sequentially stored. From waveform b, Tll, T, 2, T, 3, Tl4, Tl5...
is obtained, and T2l, T22, and T23 are found from the waveform c. Since the level of the waveform is small, the period cannot be determined from the waveform d. From these cycle data, the maximum cycle detection unit 10 searches for data with the longest cycle, and sets this as the pitch cycle. If this pitch period is T22, the reciprocal calculation section 11 performs a calculation of 1/T22 to convert the period to a frequency value, which is output as a DC voltage proportional to the pitch. Further, the output of the reciprocal calculation section 11 is transmitted to the filter control section 4.
. The cutoff frequency Fc of the pass-through filter 1 is Fc
The filter control unit 6 sets the cutoff frequency FC3 of the low-pass filter 3 to Fc3-0.8×F.
Change to O. Setting FCl, fC2, and fC3 as close to each other as possible will increase the attenuation of harmonics and allow accurate pitch extraction, but on the other hand, if the pitch fluctuates greatly, it will go out of the passband and the pitch will be lost. extraction may become impossible. According to experiments, 1 in the same person
The pitch variation per cycle reaches a maximum of 10-20%.
Therefore, in this embodiment, FC2 is set to FO to perform accurate pitch extraction, and FCl is set to 20% higher than FO.
Even if the pitch changes to 20% higher than FO by setting the pitch higher than FO, the pitch can be extracted by the low-pass filter 1. If the pitch changes lower than FO, the low-pass filter 1 will be able to extract the pitch. Since it is used, it will not go out of the passband, and if the pitch is sufficiently large from Fc3/2,
Pitch extraction is possible using low-pass filters 1 and 2. Thereafter, until the end of the voice is detected by the voice/silence detector, FC,,f is performed based on pitch detection and feedback.
By repeatedly changing C2 and fC3, the characteristics of filters 1, 2, and 3 are always maintained in a state that takes into account changes in pitch. This enables pitch extraction with less occurrence of double pitches.

FIG. 3e shows the trajectory of the pitch extracted in this way, and the pitch rises with a delay of time t from the rise of the voice due to the determination of the initial pitch value.

FIG. 3f shows changes in the cutoff frequencies FCl, fC2, and fC3 of the low-pass filters 1, 2, and 3 at this time. Effects of the Invention Since the pitch extraction device of the present invention is configured as described above,
By optimally changing the characteristics of multiple low-pass filters for pitch extraction in response to pitch changes in the audio, it is possible to sufficiently follow sudden changes in pitch and suppress the occurrence of double pitch. In addition, there is no need to change the internal constant of the filter for each speaker, and accurate pitch extraction is possible.

[Brief explanation of the drawing]

Fig. 1 is a block diagram showing an embodiment of the pitch extraction device of the present invention, Fig. 2 is a characteristic diagram of a low-pass filter in an embodiment of the present invention, and Fig. 3 a to f are diagrams of the same device. It is a signal waveform diagram of each component part. 1, 2, 3...Low pass filter 4, 5
, 6... filter control section, 7, 8, 9...
...Synchronization detection section, 10...Maximum period detection section,
11...Reciprocal calculation section, 12...Sound/
Silence detection section.

Claims (1)

[Claims] 1. A plurality of low-pass filters with variable passbands that remove harmonic components from audio input and extract only fundamental frequency components, and a plurality of low-pass filters that detect the period of the fundamental frequency components extracted by the low-pass filters. A period detection section, a maximum period detection section that detects the maximum period from the periods of the plurality of fundamental frequency components, a reciprocal calculation section that calculates the reciprocal of the period and converts it into a frequency value, and the frequency value. It consists of multiple filter control sections that control the pass band of the pass-pass filter, and a sound/silence detection section that detects the rise and fall of the audio signal. A pitch extraction device characterized by adapting a passband.