JPH11305797A - Voice analyzing synthesizer - Google Patents

Abstract

PROBLEM TO BE SOLVED: To improve voice quality by providing a pulsative analyzer analyzing a residual signal and performing voice decision when a pulsative rate is higher than a set threshold value in a voice analyzer. SOLUTION: Voice analyzer A side processing inputs a voice signal from an input terminal 1. A linear predictive analyzer 2 calculates a linear predictive coefficient (b) to output it to a linear predictive analytic filter 3 and to transmit it to the linear predictive synthetic filter 13 of the voice synthesizer B also. The linear predictive analytic filter 3 inputs the voice signal (a) and the linear predictive coefficient (b) to output the residual signal (c). An amplifier 12 amplifies the residual signal (f) outputted from a residual signal generator 11 so as to become the same as the power of the residual signal based on information of residual power (g) sent from the voice analyzer A side. The linear predictive synthetic filter 13 synthesizes the voice signal (i) from the information of the linear predictive coefficient (b) sent from the voice analyzer A side and the amplified residual signal (h) to output it to an output terminal 14.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an apparatus for compressing an audio signal for encoding and decoding.

[0002]

2. Description of the Related Art There is a speech analysis / synthesis technique generally called a vocoder (for example, "Basics of speech information processing", Saito, Nakata, Ohmsha, 1981). FIG. 3 is a configuration diagram of a conventional vocoder, wherein A 'above the dotted line is a speech analyzer, and B' below the dotted line is a speech synthesizer. The voice analyzer A 'has 101
Is an input terminal, 102 is a linear prediction analyzer that receives the audio signal a and calculates a linear prediction coefficient b, and 103 is an input of the audio signal a and the linear prediction coefficient b and calculates a first residual signal c. Linear prediction analysis filter 104 performs the first residual signal c
And a residual power analyzer 105 for calculating a residual power g. A pitch analyzer 105 for inputting the first residual signal c and calculating a voiced / unvoiced decision e ′ and a pitch frequency d. It is.

[0003] A voice synthesizer B 'includes a residual signal generator 111 which receives a voiced / unvoiced decision e' and a pitch frequency d and generates a second residual signal f '. An amplifier that receives the residual signal f 'and the residual power g and amplifies the second residual signal f'. An amplifier 113 receives the amplified second residual signal h 'and the linear prediction coefficient b. And the audio signal i '
, And 114 is an output terminal.

Next, the operation of the vocoder shown in FIG. 3 will be briefly described. The following processing is performed for a certain length (for example, 5 mse
c.) It is performed for each frame. In the processing on the voice analyzer A 'side, a voice signal a is input from the input terminal 101. FIG. 2A shows an example of an audio waveform of the audio signal a.

[0005] The linear prediction coefficient b is calculated by the linear prediction analyzer 102, output to the linear prediction analysis filter 103, and transmitted to the linear prediction synthesis filter 113 of the speech synthesizer B ′. The linear prediction analysis filter 103 receives the audio signal a and the linear prediction coefficient b, and outputs a first residual signal c. FIG. 2 (2) is an example of the first residual signal c obtained from the audio signal a of FIG. 2 (1). Residual power analyzer 1
04 calculates the residual power g of the first residual signal c and sends it to the amplifier 112 of the voice synthesizer B '. The pitch analyzer 105 receives the first residual signal c, and determines whether or not the pitch periodicity is high based on a correlation value of the first residual signal c (hereinafter, referred to as a pitch correlation value). , A voiced result if high, a voiceless result if low, and a pitch frequency d to the residual signal generator 111 of the voice synthesizer B '.

As a method of calculating the pitch frequency and the pitch correlation value, a pitch search technique which is generally widely used can be used. The search method uses a residual signal as an analysis target and a time interval (time shift amount) corresponding to a pitch frequency.
Is a variable, and the pitch frequency is determined from the time interval that maximizes the normalized autocorrelation value. Further, a method of using the normalized autocorrelation value at that time as a pitch correlation value can be used.

[0007] In the processing on the voice synthesizer B 'side, the residual signal generator 111 uses the voiced / unvoiced determination result e' sent from the voice analyzer A 'side and the pitch frequency d information to determine whether the voice signal is voiced. In this case, a periodic signal (for example, a pulse train) determined by the pitch frequency d is generated, and in the case of no voice, a noise signal (for example, white noise) is generated. This is shown in FIG.

FIG. 4 schematically shows a residual signal generator, which comprises a periodic signal generator and a noise signal generator. When the voice is voiced, a periodic signal is output according to the information of the pitch periodic wave d. When the voice is unvoiced, a noise signal is switched and output.

The amplifier 112 outputs a signal from the residual signal generator 111 based on the information on the residual power g sent from the voice analyzer A 'so that the power becomes equal to the power of the first residual signal c. The output second residual signal f 'is amplified. This situation is schematically shown in FIG. Linear prediction synthesis filter 11
3 synthesizes an audio signal i 'from the information of the linear prediction coefficient b sent from the audio analyzer A' side and the amplified second residual signal h ', and outputs it to the output terminal 114.

[0010] In addition to the method of modeling a sound source signal by switching between a periodic signal and a noise signal as described above, by modeling a signal in which a voiced signal and an unvoiced signal are mixed, the quality of an analysis-synthesized speech is improved. There is also a technology for improving (for example, “High-Quality Harmonic”
Coding At Very Low Bit Ra
tes ", G. Yang, H. Leich, ICASS
P, 1994). The control of the mixing ratio is performed based on, for example, a pitch correlation value. When the degree of periodicity is high, a large amount of periodic signals are mixed, and when the degree of periodicity is low, a large amount of noise signals are mixed.

[0011]

In the above-mentioned speech analysis / synthesizer, it is important to appropriately model a sound source signal. However, in the process of determining voiced / unvoiced according to the pitch correlation value, the pitch correlation value is low in a rising section of a periodic signal where a pitch frequency exists, and the voice is determined to be unvoiced, and the synthesized voice is noise-like. There was a problem of becoming.

[0012]

In order to solve the above-mentioned problems, a speech analysis / synthesizer according to the present invention comprises a speech analyzer / speech synthesizer composed of a speech analyzer and a speech analyzer. Calculating a linear prediction coefficient, outputting a linear prediction analysis filter and a linear prediction analyzer to the speech synthesizer, and inputting the speech signal and the linear prediction coefficient to calculate a first residual signal; A linear predictive analysis filter for outputting to subsequent processing means in the voice analyzer, a residual power analyzer for calculating the residual power by inputting the first residual signal, and outputting the residual power to the voice synthesizer; , A voice analyzer having a pitch analyzer that receives the first residual signal, calculates a pitch frequency, and outputs the pitch frequency to a pulse analyzer, and a voiced / unvoiced determination result from the pulse analyzer. And the pitch calculated by the pitch analyzer A second residual signal generator for generating a second residual signal by inputting a frequency, and outputting the second residual signal to an amplifier; the second residual signal; and a residual power from the residual power analyzer. And an amplifier for amplifying a second residual signal, the amplified second residual signal, and the linear prediction coefficient output from the linear prediction analyzer to generate an audio signal. And a speech synthesizer configured to output a linear prediction synthesis filter, and analyzing the first residual signal. If the pulse characteristic is higher than a set threshold,
The speech analyzer is provided with the pulse analyzer that performs voiced determination.

Further, the linear prediction analyzer, the linear prediction analysis filter, the residual power analyzer, and the first
, A voice analyzer comprising a voiced / unvoiced mixing ratio and a pitch analyzer for calculating the pitch frequency and outputting the pitch frequency to the pulse analyzer, and an output from the pulse analyzer. The voiced / unvoiced mixing ratio and the pitch frequency are input to generate a third residual signal, the residual signal generator, the amplifier, and the linear prediction synthesis filter. A voice synthesizer for analyzing the first residual signal and changing a voiced / unvoiced mixing ratio according to a pulsed degree output from the pitch analyzer.

[0014] The linear prediction analyzer calculates a linear prediction coefficient of the input speech signal. The linear prediction analysis filter inputs a speech signal and a linear prediction coefficient, and outputs a residual signal.
The residual power analyzer calculates the power of the residual signal. The pitch analyzer receives the residual signal and outputs a pitch frequency based on the correlation value of the residual signal. The pulse nature analyzer determines the pulse nature of the waveform of the residual signal, and changes the voiced / unvoiced mixture ratio according to the pulse nature.

[0015]

[First Embodiment] FIG. 1 is a block diagram showing an embodiment of a speech analysis / synthesizer according to the present invention. The difference from the prior art shown in FIG. 3 is that a pulse analyzer 6 is added to the voice analyzer A, and the output of the pitch analyzer 5 is input to the pulse analyzer 6.

1 is an input terminal of the speech analyzer A, 2 is a linear prediction analyzer that receives a speech signal a and calculates a linear prediction coefficient b, and 3 is a speech signal that receives a speech signal a and a linear prediction coefficient b. 1
A linear prediction analysis filter for calculating the residual signal c, a residual power analyzer 4 for receiving the first residual signal c and calculating a residual power g, and a linear power analyzer 5 for converting the first residual signal c. A pitch analyzer for inputting and calculating a pitch frequency d, a pulse analyzer 6 for inputting the first residual signal c and the pitch frequency d and performing a voiced / unvoiced determination e based on the pulse level. It is a vessel.

A voice synthesizer B 11 receives a voiced / unvoiced determination result e and a pitch frequency d to generate a second residual signal f, and a second residual signal generator 12 generates a second residual signal f. Difference signal f
And the residual power g, and an amplifier 13 for amplifying the second residual signal f. The amplifier 13 receives the amplified second residual signal h and the linear prediction coefficient b to convert the audio signal i. The linear prediction synthesis filter to be generated, 14 is an output terminal.

Next, the operation of the vocoder shown in FIG. 1 will be described. The following processing is performed for each frame of a certain fixed length (for example, 5 msec.). In the processing on the audio analyzer A side, the audio signal a is input from the input terminal 1. FIG. 2A shows an example of the audio waveform of the audio signal a. The linear prediction coefficient b is calculated by the linear prediction analyzer 2 and output to the linear prediction analysis filter 3, and the linear prediction coefficient b of the speech synthesizer B is calculated.
3 is also transmitted. The linear prediction analysis filter 3 receives the audio signal a and the linear prediction coefficient b, and outputs a first residual signal c.

FIG. 2B shows an example of the first residual signal c obtained from the audio signal a in FIG. 2A. The residual power analyzer 4 calculates the residual power g of the residual signal c, and transmits it to the amplifier 12 of the speech synthesizer B. The pitch analyzer 5 receives the first residual signal c, calculates a pitch frequency d, and outputs it to the pulse analyzer 6. The pulse analyzer 6 determines whether or not the pitch periodicity is high based on the correlation value of the first residual signal c (hereinafter referred to as pitch correlation value). Transmits the unvoiced determination result e and the pitch frequency d to the residual signal generator 11 of the voice synthesizer B.

As shown in FIG. 2 (2), generally, the residual signal in a voice section in which a pitch frequency exists has a property that the waveform becomes pulse-like compared to the original voice signal. The pulse analyzer 6 receives the first residual signal c and outputs the first residual signal c
Is calculated. As a method of calculating the degree of pulse, the ratio between the maximum absolute value and the average absolute value of the in-frame residual signal waveform represented by the equation (1) or the in-frame residual signal represented by the equation (2) is used. The ratio between the root mean square value of the waveform and the maximum absolute value can be used. Alternatively, in order to normalize the value calculated by the formulas (1) and (2) to a value of 1 or less to facilitate handling, the value can be normalized by the square root of the frame length as in the formula (3).

Ymax = Max (| Y1 |, | Y2 |, ..., | YN |) Yave = (| Y1 | + | Y2 | + ... + | YN |) / N Yrms = Sqrt ((Y1 * Y1 + Y2 * Y2 + ... + YN * YN) / N) Pulse = Ymax / Yave (1) Pulse = Ymax / Yrms (2) Pulse = Ymax / (Yrms * Sqrt (N)) (3) where N is The number of samples in the frame (for example, 40) Yi (i = 1, 2,..., N) is the peak value | x | of the residual signal in the frame is the absolute value of x Max (Y1, Y2,. , YN) is the maximum value of Y1, Y2,..., YN. Sqrt (x) represents the square root of x.

When equation (3) is used, for example, when only one pulse exists in a frame as an example of the strongest pulse, the calculated value is 1, and conversely, the weakest pulse. As an example, when there are all signals of equal amplitude in a frame, the calculated value is 1 / Sqrt (N).

Next, the calculated pulse index value is compared with a certain judgment threshold value (for example, 0.5). If the pulse index value is larger, voiced / unvoiced input from the pitch analyzer 5 is obtained. Is set to voiced determination. In this case, as the pitch frequency, a method of setting the pitch correlation value to a value with the highest possibility of the pitch frequency (higher pitch correlation value) or a method of determining a certain fixed value can be considered.

In the processing on the voice synthesizer B side, the residual signal generator 11 determines whether the voice signal is voiced from the voiced / unvoiced determination result e sent from the voice analyzer A side and the pitch frequency d information.
A periodic signal (for example, a pulse train) determined by the pitch frequency d is generated, and in the case of no voice, a noise signal (for example, white noise) is generated. This is shown in FIG.

The amplifier 12 outputs the signal from the residual signal generator 11 based on the information on the residual power g sent from the voice analyzer A so that the power becomes the same as the power of the first residual signal c. The amplified second residual signal f is amplified. Figure 2 shows this situation.
This is schematically shown in (4). The scales on the vertical axis and the horizontal axis in FIG. 2D are the same as those in FIGS. 2A to 2C. However, in this method, since there is no information on the position where the pulse is generated, where the position of the pulse stands in the analysis frame depends on FIGS. 2 (3) and (4) and FIGS. 2 (1) and (2). ). The linear predictive synthesis filter 13 includes a speech analyzer A
The audio signal i is synthesized from the information of the linear prediction coefficient b sent from the side and the amplified second residual signal h, and the output terminal 1
4 is output.

If the processing content for the audio signal is exemplified in the case of FIG. 2, in the prior art, the pitch correlation value is calculated to be low in the rising section of the pitch, and FIG.
As shown in (3), a section from 3.880 seconds to 3.885 seconds is determined to be unvoiced, and a noise signal is generated. In the present invention, however, the degree of pulse property of the residual signal is high, and FIG. ), Voiced determination is performed, and a periodic signal (here, a pulse train) is generated.

[Embodiment 2] In the above embodiment,
Although the residual signal, which is the sound source signal, is represented by a periodic signal or a noise signal, as described in the related art, the present invention can be applied to a system in which a voiced signal and an unvoiced signal are represented by a mixed signal. it can. In this case, as in FIG. 1, the only processing block newly added is the pulse analyzer 6.
As in the processing content, the pulse content of the residual signal is calculated by using, for example, Equation (3), as in the first embodiment. The pitch correlation value input from the pitch analyzer 5 is changed according to the calculated index value of the pulse characteristic, and is output to the voice synthesizer B side.
The process of changing the pitch correlation value is such that the higher the degree of pulse property, the higher the pitch correlation value. For example, a process such as Expression (4) can be considered.

In the case of Pcor = Pulse, Pulse> Pcor Pcor = Pcor, Pluse <= Pcor (4) Here, Pcor is a pitch correlation value normalized to a value of 1 or less. Pulse represents a pulse degree. For example, it can be obtained by equation (3).

As described above, by changing the pitch correlation value according to the pulse degree of the residual signal waveform, it is possible to generate more periodic signal (eg, pulse train) components in the rising section of the pitch earlier than in the conventional method. And sound quality can be improved.

[0030]

According to the present invention, a voice analysis / synthesizer with improved sound quality can be realized by performing voiced judgment even in a section where pitch extraction is difficult, such as a rising section of a voice pitch.

[Brief description of the drawings]

FIG. 1 is a block diagram showing an embodiment of a speech analyzer / synthesizer according to the present invention.

FIG. 2 is a waveform diagram of an audio signal according to the prior art and the processing of the audio analysis / synthesizer according to the present invention.

FIG. 3 is a block diagram showing an embodiment of a conventional speech analyzer / synthesizer.

FIG. 4 is an explanatory diagram of an operation of a residual signal generator of a conventional speech analysis / synthesizer.

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 1,101 Input terminal 2,102 Linear prediction analyzer 3,103 Linear prediction analysis filter 4,104 Residual power analyzer 5,105 Pitch analyzer 6 Pulse analyzer 11,111 Residual signal generator 12,112 Amplifier 13,113 Linear prediction synthesis filter 14,114 Output terminal

Claims (2)

[Claims] 1. A speech analysis / synthesizer composed of a speech analyzer and a speech synthesizer, wherein a speech signal is input, a linear prediction coefficient is calculated, and a linear prediction coefficient is output to a linear prediction analysis filter and the speech synthesizer. A prediction analyzer, the linear prediction analysis filter that receives the audio signal and the linear prediction coefficient, calculates a first residual signal, and outputs the first residual signal to a subsequent processing unit in the audio analyzer. A residual power analyzer that receives the first residual signal to calculate a residual power, and outputs the residual power to the speech synthesizer; and inputs the first residual signal, calculates a pitch frequency, and outputs a pulse. A voice analyzer having a pitch analyzer for outputting to the sex analyzer; a voiced / unvoiced determination result from the pulsed sex analyzer; and a pitch frequency calculated by the pitch analyzer. And output it to the amplifier. A residual signal generator; an amplifier that receives the second residual signal and the residual power from the residual power analyzer and amplifies the second residual signal; And a speech synthesizer comprising a linear prediction synthesis filter that receives the linear prediction coefficient output from the linear prediction analyzer, generates a speech signal, and outputs the speech signal. 1. A speech analysis / synthesizer characterized in that the speech analyzer is provided with the pulse analyzer for performing a voiced judgment when the pulse signal degree is higher than a set threshold value by analyzing the residual signal of No. 1. . 2. The linear prediction analyzer, the linear prediction analysis filter, the residual power analyzer, and a voiced / unvoiced mixing ratio by inputting the first residual signal and the pitch frequency. A voice analyzer having a pitch analyzer that calculates and outputs the pitch to the pulse analyzer; a voiced / unvoiced mixing ratio output from the pulse analyzer and the pitch frequency; 3; a residual signal generator that generates the residual signal of No. 3; an amplifier that amplifies the third residual signal; and a speech synthesizer including the linear prediction synthesis filter. 2. The speech analysis / synthesis according to claim 1, further comprising the pulse analyzer for analyzing a residual signal and changing a voiced / unvoiced mixture ratio according to a pulse output from the pitch analyzer. vessel.