JPS6239754B2 - - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION The present invention relates to an apparatus for detecting the pitch of an audio signal waveform.

Pitch detection in voiced sounds is one of the necessary processes when recording, reproducing, or transmitting audio signal waveforms at low information density. As a method for pitch detection, for example, in a speech analysis and synthesis system using the LPC (linear prediction) method, the residual signal is generally low-passed after extracting other necessary speech parameters such as the spectral envelope using a whitening filter of about 10 stages. The method used is to take autocorrelation after passing through a filter and perform moving average processing to obtain pitch information, but the amount of calculation required is enormous and the hardware is also large-scale. In addition, in LPC-based speech analysis and synthesis systems, pitch detection errors during analysis are one of the factors that impair the quality of reproduced speech, but since error correction is done manually, analysis costs are extremely high. It's getting expensive.

On the other hand, systems that record, playback, and transmit audio signal waveforms using delta modulation (including adaptive delta modulation, the same applies hereinafter) have a low information density.
Although it is not very low at 16Kb to 32Kb/ses, it is advantageous over the LPC method in terms of hardware simplicity and real-time performance. In such a delta modulation recording/reproducing or transmission system, the inventors
We are considering reducing the information density by recording or transmitting the repeated waveform of the voiced part of the audio signal waveform in real time, for example, with one wave representing every four waves. In that case, the operation to detect whether voiced parts are repeated is pitch detection.
If a method similar to that used in the LPC method is used for pitch detection, it would be very wasteful in terms of structure, so it is required to perform pitch detection with a simple structure.

This invention was made in view of the above points,
The object is to provide an audio pitch detection device that can accurately detect the pitch of an audio signal waveform with a simple configuration.

In this invention, the average slope of the audio signal waveform per predetermined time period is sequentially detected, and a pulse corresponding to a period in which the slope exceeds a certain value is obtained. Then, the leading edge of this pulse is detected to obtain a relatively thin pulse, and only those pulses after a certain period of time have elapsed since the previous pulse are extracted, and intervals corresponding to the pitch of the audio signal waveform are extracted. obtain a pulse train of

According to the present invention, it is possible not only to detect pitches in speech with a significantly simpler configuration than pitch detection in the LPC method, but also to greatly reduce detection errors.

Embodiments of the present invention will be described below with reference to the drawings.

FIG. 1 shows an embodiment of the present invention. In the figure, an audio signal input to an input terminal 1 is applied to an encoder 3 using a delta modulation system via an input filter 2, and becomes encoded data synchronized with a clock signal from a clock oscillator 4. This encoded data is then input to the audio pitch detection section 5. This audio pitch detection section 5 is constructed as follows.

The output data of the encoder 3 is sent to the pitch detection section 5.
A clock signal from the clock oscillator 4 is serially input to the shift register 11 in the clock oscillator 4. The contents of each stage of the shift register 11 are output in parallel and sent to an adder 13 via a ladder resistor 12.
is added.

FIG. 2 shows a typical waveform example of the audio sound portion of the audio signal, where T ₀ represents the pitch period and T ₁ represents the period of the first formant. Here, if the number of stages of the shift register 11 is N and the period of the clock signal is τ, then the time τ・N=T is set to an appropriate value less than half of the period _T1 of the first formant as shown in Figure 2. be done. The optimum value of this time T varies depending on the gender of the speaker, etc., but is within the range of about 0.5 to 2 msec. Adding the inputs of each stage of the shift register 11 by the adder 13 via the ladder resistor 12 means that the values of the ladder resistor 12 are approximately equal, that is, the outputs of each stage of the shift register 11 are added with approximately the same weight. Then, T of the input audio signal waveform
This corresponds to detecting the average slope per unit time.

For example, if the audio signal waveform to the input terminal 1 is shown as a in FIG. . That is, when the shift register 11 receives one piece of data, it outputs one piece of data N bits before, and the adder 13
outputs an output according to the difference between "1" and "0" of the N-bit data in the shift register 11, so
The output of the adder 13 remains unchanged as "0" if the shift register 11 outputs "1" when "1" is input.
If "0" is entered in the shift register 11, the shift register 11 will remain unchanged or decrease by one step.

The output of the adder 13 is input to a comparator 14 and compared with a constant level indicated by Vth in FIG. 3d. As a result, the output of the comparator 14 is
As shown in FIG. 3e, a pulse is obtained that corresponds to a period in which the output of the adder 13 exceeds Vth, in other words, a period in which the average slope per unit time T of the input audio signal waveform exceeds a certain value.

The output of comparator 14 is applied to a first monostable multivibrator 15. This multivibrator 15 is connected to the comparator 1 shown in FIG. 3e.
It is triggered at the rising edge of the output pulse No. 4, that is, the leading edge, and generates a thin pulse as shown in the figure f.
The output of the first monostable multivibrator 15 is applied directly to one input terminal of the AND gate 17 and also to the second monostable multivibrator 16 . The second monostable multivibrator 16 is the first monostable multivibrator 1
It has a time constant (for example, 2.5 to 10 msec) sufficiently longer than 5, and is triggered by the falling edge, that is, the trailing edge, of the output pulse of the first monostable multivibrator 15 shown in FIG. A pulse having a width sufficiently wider than that of the pulse f is generated as an inverted output. The output of the second monostable multivibrator 16 is then applied to the other input terminal of the AND gate 17. Therefore, as shown in FIG. 3g, the ADN gate 17 outputs only the pulse corresponding to P ₁ of the pulses P ₁ and P ₂ of f.

In other words, the driving source of voiced sounds is an impulse, and the mechanism is to radiate this impulse through the acoustic tube formed by the throat, palate, and lips, so the waveform is attenuated from the excitation point as shown in Figure 2. However, when looking at the output of the first monostable multivibrator 14, as shown in FIG. 3, a pitch may be erroneously detected even at the second peak of the voiced part of the input audio signal waveform. addition to this
It's called error. Therefore, in the present invention, among the output pulses of the first monostable multivibrator 15,
The AND gate 17 prohibits pulses within a certain time period corresponding to the width of the output pulse of the second monostable multivibrator 16 from the previous pulse, and outputs only pulses after this certain period of time has elapsed. As a result, a pulse train having an interval corresponding to the pitch period of the input audio signal waveform is obtained at the output of the AND gate 17.

According to the above configuration, the pitch of the audio signal waveform can be detected satisfactorily. For example, the clock signal frequency 1/τ is 16 kHz, and the number of stages N of the shift register 11 is 12, that is, T = 0.75.
msec, and the threshold level _Vth of the comparator 14 is set so that when 8 or more bits of the 12-bit data in the shift register 11 are "1", the output of the comparator 14 is inverted, and the output of the comparator 14 is inverted. ``OYASUMI NO''
When we input the audio signal of about 1 second when uttering "JIKAN" and actually detected the pitches, we found that out of the actual pitch of 244 waves excluding the weak part at the end of the word,
3 at minimum 3msec, maximum 4.2msec, average 3.3msec
98.75 with just one omit error.
A high correct answer rate of % was obtained. Moreover, the position where this error occurred was also a part where there was no difference in sound quality even if it was regarded as silence between low-level phonemes.

As described above, according to the present invention, pitches in speech can be detected accurately with a low error rate. In addition, conventional LPC performs pitch detection from the residual signal after other audio parameters are extracted using a multi-stage white filter.
Compared to the pitch detection means in the system, the hardware is significantly simpler, and in this respect it is extremely suitable for audio recording and reproducing systems using the delta modulation system.

Furthermore, according to the embodiment shown in FIG. 1, since the pitch of the audio signal waveform is detected from the output data of the encoder 3, the circuit can be constructed entirely of digital circuits, making it suitable for IC implementation. Note that the processing in the pitch detection section can also be executed by a microcomputer.

Next, an application example of the present invention will be explained with reference to FIGS. 4 and 5. FIG. 4 shows an example in which the present invention is applied to an automatic waveform compression device for an audio recording device, and parts common to those in FIG. 1 are given the same reference numerals. Moreover, FIG. 5 is a waveform diagram of each part of FIG. 4. A pulse train having an interval corresponding to the pitch period of the audio signal waveform as shown in FIG.
For example, if N = 4, the N-ary counter 41 is composed of a two-stage binary counter, and every time four pulses are inputted from the pitch detection section 5, one pitch is inputted until the next one pulse input is applied. only for a period
The AND gate 43 generates the gate signal shown in FIG. 5b. AND gate 44 is AND gate 43
The data for one pitch from the encoder 3 corresponding to the period specified by the gate signal is shown in Fig. 5c.
With respect to the waveform of the voiced part, which is extracted as shown in FIG. That is, the waveform of the voiced sound part is compressed into I/N and stored.

On the other hand, for waveforms of portions where the pitch cannot be detected by the pitch detection section 5, such as unvoiced parts and word endings, the output data of the encoder 3 is stored in the memory system 45 as is through the AND gate 49. In this operation, the level detector 46 detects the level of the audio signal to obtain the voiced/unvoiced discrimination signal shown in FIG. The window comparator 47 checks whether it is within a predetermined range to obtain the limit signal shown in FIG.
Detects a matching signal between and e, and performs AND with this signal f.
This can be achieved by controlling the gate 49.

With this configuration, when the same voice signal waveform of a woman saying "OYASUMI NO JIKAN" as described above was stored in the memory system 45, the portion of the output data of the encoder 3 that was stored as is without being compressed was approximately 3% of the total. , the average information density is about 4Kb/
It was also reduced to sec. In addition, in this case, the omitted parts tend to be limited to the beginnings of words or between phonemes, except for voiced parts where the level is high and constant, so this omit error information is also effective for phoneme segmentation. Can be seen.

[Brief explanation of the drawing]

FIG. 1 is a diagram showing the configuration of an audio pitch detection device according to an embodiment of the present invention, FIG. 2 is a diagram showing a typical waveform example of a voiced part of an audio signal, and FIG.
FIG. 4 is a diagram showing an application example of the present invention, and FIG. 5 is a waveform diagram of each part to explain the operation of FIG. 4. 1... Audio signal input terminal, 2... Input filter, 3... Encoder, 4... Clock oscillator,
5... Pitch detection unit, 11... Shift register,
12...Ladder resistor, 13...Adder, 14...
Comparator, 15, 16... Monostable multivibrator, 17... AND gate, 41... N-ary counter, 42... Interval counter, 45...
...Memory system, 46...Level detection section, 47
...window comparator.

Claims (1)

[Claims] 1 Consists of a shift register that stores a data string obtained by encoding an audio signal waveform by delta modulation for a certain length of time, and an adder that adds the outputs of each stage of this shift register, and calculates the average of the audio signal waveform per predetermined time. a first means for sequentially detecting the slope; and a second means comprising a comparator for detecting a period in which the average slope detected by the first means exceeds a certain value and generating a pulse corresponding to that period. , a third means for generating a pulse by detecting the leading edge of the output pulse of the second means;
It is characterized by comprising a fourth means for extracting only the pulses after a certain period of time has elapsed since the previous pulse from among the output pulses of the means to obtain a pulse train having an interval corresponding to the pitch period of the audio signal waveform. Audio pitch detection device.