JPH07244498A - Basic frequency extractor - Google Patents

Abstract

(57) [Abstract] [Purpose] The A-B-S method is introduced to extract an accurate fundamental frequency. An error calculator 1 calculates an autocorrelation function of input speech calculated by the input speech autocorrelation function calculator 15 and an autocorrelation function of synthesized speech calculated by the synthesized speech autocorrelation function calculator 17.
6, and the fundamental frequency is adjusted by the frequency adjuster 13 to extract the fundamental frequency.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a fundamental frequency extraction device which is used in digital mobile communication or the like and is required when analyzing and synthesizing a voice signal.

[0002]

2. Description of the Related Art In recent years, a technique of analyzing and synthesizing a voice signal has been attracting attention as a basic technique of information compression which efficiently utilizes a data transmission / storage medium. Above all, a fundamental frequency extraction device that extracts a fundamental frequency, which is an important parameter in the analysis and synthesis processing of a speech signal, without error is widely required in order to improve the quality of synthesized speech.

A conventional fundamental frequency extraction device will be described below. FIG. 17 shows a block diagram of a conventional fundamental frequency extraction device. In FIG. 17, reference numeral 41 is a voice input device for inputting voice. An autocorrelation calculator 42 calculates the autocorrelation function of the input voice. Reference numeral 43 denotes an error function calculator, which obtains the certainty as a fundamental frequency from the obtained autocorrelation. The error function obtained by the error function calculator 43 is accumulated in the buffer 44 for n frames (n is arbitrary). Reference numeral 45 is a fundamental frequency extractor, which extracts the frequency at which the sum of the accumulated error function values is the minimum.

In the fundamental frequency extraction device configured as described above, the autocorrelation calculator 42 obtains the autocorrelation function from the voice input by the voice input device 41. The error function calculator 43 calculates the error function shown in FIG. 18 based on the calculated autocorrelation function. In addition, FIG.
As shown in (a), the error function obtained for each frame is accumulated in the buffer 44 for n frames, and the fundamental frequency extracted from the fundamental frequency shown in FIG. 19 (b) is changed so that the sum of these functions is minimized. It is extracted by the device 45.

[0005]

However, in the conventional fundamental frequency extraction device as described above, in the case of a system for analyzing and synthesizing an audio signal without considering the phase information of the input audio signal, the input audio signal Since the waveform of and the waveform of the synthesized voice are completely different as shown in FIGS. 20 and 21,
Since it is not possible to directly compare these waveforms, it is difficult to confirm the validity of the extracted fundamental frequency. Especially, the frequency that is an integral multiple of the true fundamental frequency (double pitch) or the frequency that is an integral fraction (half pitch) is the basic frequency,
It had a problem that it was often mistakenly extracted.

The present invention solves the above-mentioned conventional problems, and an object of the present invention is to provide a fundamental frequency extraction device for extracting an accurate fundamental frequency in a speech synthesizer using any speech analysis and synthesis method.

Another object of the present invention is to provide an excellent fundamental frequency extraction device capable of avoiding double-pitch and half-pitch errors in a system that synthesizes voice for each harmonic frequency.

[0008]

In order to achieve the above object, the invention of claim 1 is a voice input device for inputting voice,
An input speech autocorrelation calculator that calculates an autocorrelation function of the input speech, an initial fundamental frequency analyzer that analyzes the period of the input speech to obtain an initial fundamental frequency, and an error calculator that calculates the initial fundamental frequency. A fundamental frequency adjusting device that determines a fundamental frequency by adjusting the output of the device, a speech analysis and synthesis device that analyzes and synthesizes speech according to the fundamental frequency, a synthesized speech autocorrelation calculator that calculates an autocorrelation function of the synthesized speech, An error calculator for calculating an error of the autocorrelation function of the input voice and the synthesized voice is provided.

[0009] According to a second aspect of the present invention, the voice synthesizing means divides a frequency into harmonic bands based on a fundamental frequency, and a harmonic amplitude for each of the divided harmonics. A harmonic amplitude calculator to be obtained, and a harmonic sound source generator that generates a signal for each harmonic according to the calculated amplitude, the input speech autocorrelation calculator and the synthesized speech autocorrelation calculator, A first harmonic sound source synthesizer that generates a first synthesized voice by adding all the harmonics generated by the harmonic sound source generator, and a fundamental frequency of the obtained harmonic sound source is removed. This is replaced with a second harmonic sound source synthesizer that generates a second synthesized voice by adding harmonic sound sources.

[0010]

According to the structure of claim 1, even in a system in which the input voice and the synthesized voice cannot be directly compared to analyze and synthesize the voice without considering the phase information of the input voice signal, the autocorrelation function is The extracted fundamental frequency can be corrected by comparing. Then, the fundamental frequency closest to the input voice can be extracted by using a closed loop structure in which such an A-B-S method (analysis method by synthesis) is introduced.

According to the structure of claim 2, when the input voice signal is filtered by Takashiro, only the synthesized voice having the frequency characteristic analyzed with the extracted fundamental frequency as a reference and the fundamental frequency component from this synthesized voice. By comparing the autocorrelation functions of the synthesized speech having the frequency characteristics with the removed, it is possible to detect double pitch and half pitch errors, and it is possible to extract a more accurate fundamental frequency.

[0012]

DESCRIPTION OF THE PREFERRED EMBODIMENTS An embodiment of the present invention will be described below with reference to the drawings.

In FIG. 1, 11 is a voice input device for inputting voice, and 12 is a fundamental frequency analyzer, which analyzes the input voice and obtains an initial value of a fundamental frequency in the subsequent processing. Reference numeral 13 is a frequency adjuster, which is a fundamental frequency analyzer 12
Error calculator 1 with the fundamental frequency obtained from
Adjust according to the output from 6. Reference numeral 14 is a voice analysis / synthesis device, which analyzes a voice signal based on the fundamental frequency determined by the frequency adjuster 13 to generate a synthetic voice. 15
Is an input speech autocorrelation function calculator, which calculates an autocorrelation function of the input speech.

The error calculator 16 calculates the input speech autocorrelation function obtained from the input speech autocorrelation function calculator 15 by the error calculation function and the synthesized speech autocorrelation function calculator 1.
The error from the autocorrelation function of the synthetic speech obtained from 7 is calculated. The synthetic speech autocorrelation function calculator 17 obtains the autocorrelation function of the synthetic speech generated by the speech analysis / synthesis unit 14.

FIG. 2 is a block diagram showing the arrangement of a voice analysis / synthesis device that enables detection of double-pitch errors by using a system for analyzing / synthesizing voice signals for each harmonic frequency band.

In FIG. 2, reference numeral 13 is a frequency adjuster, which adjusts the fundamental frequency when analyzing the input voice signal. 22 is a harmonic band divider, which is a frequency adjuster 13
The frequency band is divided based on the fundamental frequency defined in. Reference numeral 23 is a harmonic amplitude calculator, which calculates the amplitude of the harmonic spectrum for each frequency band divided by the harmonic band divider 22, and based on the amplitude value and the harmonic frequency, the sound source for each harmonic frequency. The harmonic source generator 24
Generate with. Reference numeral 25 denotes a first harmonic sound source synthesizer, which adds all the sound sources obtained by the harmonic sound source generator 24 to generate a synthetic speech a. Obtained with the vessel 26.

Reference numeral 27 denotes a second harmonic sound source synthesizer, which adds all the sound sources except the fundamental frequency sound source among the sound sources obtained by the harmonic sound source generator 24 to generate a synthesized voice b,
The autocorrelation function b of the synthesized voice b is obtained by the second autocorrelation function calculator 28. An error calculator 29 compares the autocorrelation function a and the autocorrelation function b, and the frequency adjuster 13
It is determined whether or not the fundamental frequency defined in step 2 is a double pitch.

The operation of the fundamental frequency extraction device of this embodiment constructed as described above will be described.

In FIG. 1, the fundamental frequency analyzer 12 obtains the initial value of the fundamental frequency in the subsequent processing from the speech signal shown in FIG. The frequency adjuster 13 determines the basic frequency used when the voice synthesizer 11 analyzes the voice signal based on the obtained initial value. The voice analysis / synthesis unit 14 analyzes the voice signal based on the determined fundamental frequency to generate a synthetic voice as shown in FIG. The input speech autocorrelation function calculator 15 calculates the input speech autocorrelation function shown in FIG. 4, and the synthetic speech autocorrelation function calculator 17 calculates the synthesized speech autocorrelation function (see FIG. 6) shown in FIG. Ask. The error of the autocorrelation function of the input voice and the synthesized voice is obtained by the error calculator 16, and the fundamental frequency is adjusted by the frequency adjuster 13 to the fundamental frequency at which the error is minimized.

Next, the operation in the case of detecting the double pitch error by using the system for analyzing and synthesizing the voice signal for each harmonic frequency band will be described with reference to the flowchart of FIG. 7 and FIGS. 8 to 12. To do.

First, in step 31 of FIG. 7, the end condition (flag ← 0, threshold value Em) of the fundamental frequency extraction algorithm is set. In the next step 32, the frequency adjuster 13 sets the fundamental frequency for analyzing the input audio signal. FIG. 8 shows the power spectrum and fundamental frequency of the input voice at this time. Based on the frequency determined here, the frequency band is divided by the fundamental harmonic band divider 22 as shown in FIG. 9 (step 33). The harmonic amplitude calculator 23 calculates the amplitude of each harmonic spectrum for each of the divided frequency bands (step 3).
4). Based on this amplitude value and the harmonic frequency, the harmonic sound source generator 24 generates a sound source for each harmonic frequency as shown in FIG. (Step 36). The following two types of synthesized speech are obtained from the obtained harmonic sound source.

First, in the first harmonic sound source synthesizer 25,
Speech is synthesized from all the harmonic sound sources as shown in FIG. 11 (step 38), and in the second harmonic sound source synthesizer 27,
Speech is synthesized from all sound sources except the fundamental frequency sound source as shown in FIG. 12 (step 39). 2 obtained here
The autocorrelation function of each synthetic speech is obtained by the autocorrelation function calculator 26, 28. (Steps 40, 4
1).

In the error calculator 29, first, step 42
In, the difference e between the two autocorrelation functions is obtained, and in the next step 43, the difference e is compared with the threshold value Em. Finally, in step 44, it is determined whether the fundamental frequency extraction algorithm completion flag is set.
When it is determined in step 43 that e <Em is not satisfied, the fundamental frequency extraction algorithm completion flag is set to "1".
, The fundamental frequency adjuster 13 halves the fundamental frequency (step 37), and the above process is repeated. When it is determined in step 44 that flag = 1 is not satisfied, the fundamental frequency adjuster 13 doubles the fundamental frequency (step 37), and the above process is repeated. If step 44 is true, the fundamental frequency is doubled in step 45 to obtain the extracted fundamental frequency.

FIGS. 13, 14 and 15, and 16 show the output waveforms obtained by the autocorrelation function calculators 26 and 28 in steps 40 and 41, respectively, and FIGS. 13 and 14 show the fundamental frequency. Represents the autocorrelation function obtained when is correctly extracted, and the autocorrelation function obtained when the fundamental frequency is erroneous in double pitch. As is apparent from FIGS. 15 and 16, the autocorrelation function when a double pitch error occurs is significantly different from the others.

In the fundamental frequency extracting apparatus according to the present embodiment, even in the voice analysis / synthesizer which cannot directly compare the input voice and the synthesized voice, a closed loop is formed.
An excellent effect is obtained in that the -bS method can be introduced.

Further, by comparing the autocorrelation functions of the two types of synthesized speech, the double pitch extraction error of the fundamental frequency can be detected, and an excellent effect is obtained in that the accurate fundamental frequency can be extracted.

As described above, according to the present embodiment, by providing the autocorrelation function calculator, it is possible to introduce a closed loop configuration that enables the A-B-S method without depending on the voice analysis / synthesis device. Therefore, the fundamental frequency can be accurately extracted.

[0028]

As described above, in the present invention, by providing the autocorrelation function calculator, it is possible to introduce a closed loop configuration that enables the A-B-S method without depending on the speech synthesizer. Therefore, it is possible to realize an excellent fundamental frequency extracting device that can accurately extract the fundamental frequency.

Further, in the present invention, by providing two kinds of harmonic sound source synthesizers and an autocorrelation function calculator, a double pitch extraction error of the fundamental frequency can be detected and the fundamental frequency can be accurately extracted. An excellent fundamental frequency extraction device that can be realized can be realized.

[Brief description of drawings]

FIG. 1 is a block diagram of a fundamental frequency extraction device according to an embodiment of the present invention.

FIG. 2 is a block diagram showing a configuration of a voice analysis / synthesis device according to an exemplary embodiment of the present invention.

FIG. 3 is a waveform diagram of an input voice in this embodiment.

FIG. 4 is a diagram showing an autocorrelation function of input speech in the present embodiment.

FIG. 5 is a waveform diagram of synthesized speech in this embodiment.

FIG. 6 is a diagram showing an autocorrelation function of synthesized speech in the present embodiment.

FIG. 7 is a flowchart showing an operation procedure of the fundamental frequency extraction device in the present embodiment.

FIG. 8 is a diagram showing a power spectrum and a fundamental frequency of input speech in the present embodiment.

FIG. 9 is a diagram showing division of a harmonic band based on a fundamental frequency in the present embodiment.

FIG. 10 is a waveform diagram of each harmonic sound source in this embodiment.

FIG. 11 is a waveform diagram of synthesized speech by the harmonic sound source synthesizer 25 in the present embodiment.

FIG. 12 is a waveform diagram of synthesized speech by the harmonic sound source synthesizer 27 in the present embodiment.

FIG. 13 is an output waveform diagram of the autocorrelation function calculator 26 when the fundamental frequency is correct in the present embodiment.

FIG. 14 is an output waveform diagram of the autocorrelation function calculator 28 when the fundamental frequency is correct in the present embodiment.

FIG. 15 is an output waveform diagram of the autocorrelation function calculator 26 when the fundamental frequency is incorrect in the present embodiment.

FIG. 16 is an output waveform diagram of the autocorrelation function calculator 28 when the fundamental frequency is incorrect in the present embodiment.

FIG. 17 is a block diagram of a conventional fundamental frequency extraction device.

FIG. 18 is an explanatory diagram showing a conventional error function.

FIG. 19 is an explanatory diagram showing a conventional accumulated error and a fundamental frequency.

FIG. 20 is a diagram showing a waveform of a conventional input signal.

FIG. 21 is a diagram showing a waveform of a conventional synthesized speech signal.

[Explanation of symbols]

 11 voice input device 12 fundamental frequency analyzer 13 frequency adjuster 14 voice analysis synthesizer 15 input voice autocorrelation function calculator 16 error calculator 17 synthetic voice autocorrelation function calculator 22 harmonic band divider 23 harmonic amplitude calculator 24 Harmonic Source Generator 25 First Harmonic Source Synthesizer 26 First Autocorrelation Function Calculator 27 Second Harmonic Source Synthesizer 28 Second Autocorrelation Function Calculator 29 Error Calculator 41 Voice Input Device 42 autocorrelation calculator 43 error function calculator 44 buffer 45 fundamental frequency extractor

Claims (2)

[Claims] 1. A voice input means for inputting voice, an input voice autocorrelation calculating means for calculating an autocorrelation function of the input voice, and an initial stage for obtaining an initial fundamental frequency by analyzing a cycle of the input voice. A fundamental frequency analyzing means, a fundamental frequency adjusting device for determining the fundamental frequency by adjusting the initial fundamental frequency by the output of the error calculating means, a voice analysis / synthesizing device for analyzing and synthesizing a voice according to the fundamental frequency, and a self-synthesizing voice A fundamental frequency extraction device comprising: synthetic speech autocorrelation calculation means for calculating a correlation function; and error calculation means for calculating an error of the autocorrelation function of the input speech and the synthesized speech. 2. A harmonic band dividing means for dividing the frequency into harmonic bands based on a fundamental frequency by the voice synthesizing means, and a harmonic amplitude calculating means for obtaining an amplitude of the harmonic for each of the divided harmonics. A harmonic sound source generating means for generating a signal for each harmonic according to the calculated amplitude, and the input sound autocorrelation calculating means and the synthesized sound autocorrelation calculating means are First harmonic sound source synthesizing means for generating a first synthesized voice by adding all generated harmonic waves, and adding a harmonic sound source excluding the fundamental frequency of the obtained harmonic sound sources The fundamental frequency extraction device according to claim 1, wherein the fundamental frequency extraction device is replaced with a second harmonic sound source synthesizing means for generating a second synthetic voice.