JP2711737B2 - Linear predictive analysis / synthesis decoder - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION (Technical Field to Which the Invention Belongs) The present invention relates to a method of synthesizing a digital signal encoded by a linear predictive coding (LPC) method among speech coding methods. The present invention relates to a linear prediction analysis / synthesis type decoder that decodes and converts into a speech signal.

(Prior Art and Problems Thereof) FIG. 1 is a principle system diagram of a speech codec according to a conventional linear prediction analysis / synthesis coding method, and FIG.
(B) is a decoder. In the figure, an input audio signal A
Is subjected to linear prediction analysis. That is, the linear prediction analysis coefficient and the sound source strength information a are extracted by the linear prediction analyzer and the sound source strength extractor 11, the voiced unvoiced discrimination signal c is extracted by the voiced unvoiced discriminator 12, and the pitch period b at the time of voiced sound is The parameters a, b, and c extracted by the pitch period extractor 13 are encoded by the encoder 14 and transmitted as the transmission signal B. The decoder separates the input signal C into signals d, e, and f of the respective parameters by the separation circuit 21 and controls the cycle of the pitch pulse generator 22 according to the pitch cycle e.
After the output of the pitch pulse generator 22 and the output of the noise generator 23 are selected by the switch 24 according to the voiced / unvoiced discrimination signal and the sound source strength information f, the strength is determined, and the output signal is used as a drive signal to perform linear prediction. The audio signal D restored by driving the synthesis filter 25 based on the analysis coefficient d is extracted. In this case, as a limit of the audio quality of the audio signal D, the drive signal of the synthesis filter 25 is converted to a pitch pulse generator 22.
It has been pointed out that either one of the output of the noise generator 23 and the output of the noise generator 23 is selected, and a mixed signal of both is required as the third drive signal.

That is, in the conventional method, the output of the pitch pulse generator 22 is selected for voiced sound, and the output of the noise generator 23 is selected for unvoiced sound.
In the model that selects the output of the above as a drive signal, there is a problem that it is not possible to model a drive signal in which both are mixed at the transition from unvoiced sound to voiced sound, and the voice quality of the decoded voice is low. There was a disadvantage that it was bad.

(Object of the Invention) An object of the present invention is to increase the information to be transmitted, for example, without performing any special operation on the encoder side in the speech codec using the linear prediction analysis / synthesis technique as described above. A signal processing method capable of producing a signal obtained by mixing a pitch pulse generator output and a noise signal as a third signal for driving a synthesis filter for restoring audio only on the decoder side without using the decoder. To provide equipment.

FIG. 2 is a circuit block diagram showing an embodiment of the present invention. The encoder on the transmitting side is the same as the conventional one, and is omitted here. Data E from the encoder is separated by a separation circuit.
31 separates them into respective parameters g, h, i.
Among the parameters, the linear prediction coefficient g is input to the processing control circuit 34 and the synthesis filter 38, the voiced / unvoiced identification signal and the sound source intensity information h are input to the processing control circuit 34, and the pitch period i is input to the pitch pulse generator 32. Is entered.

Based on the linear prediction coefficient g, the processing control circuit 34 converts the drive signal into a frequency range using the output from the pitch pulse generator 32 on the frequency axis and a frequency using the output of the noise generator 33. The control signals j and k in these frequency ranges are applied to output control circuits 35 and 36 provided at the outputs of the pitch pulse generator 32 and the noise generator 33, respectively. These output control circuits 35 and 36 are provided with respective generators 32 and 3
3 is divided into a desired frequency range and output.
After being synthesized in step 7, the driving signal of the synthesis filter 38 is used. The output F of the synthesis filter 38 is a restored audio signal.

Next, the processing control circuit 34 determines how the linear prediction coefficient g
It will be described whether the drive signal can be divided on the frequency axis into a pitch component output from the pitch pulse generator 32 and a noise component output from the noise generator 33.

Audio signals are generally divided into voiced and unvoiced sounds.
The frequency spectrum of a voiced sound is composed of an envelope and a fine structure that is a harmonic of pitch. FIG. 3 is a waveform diagram showing these spectra. The solid line in FIG. 3 (a) is the harmonic component of the pitch, and the dotted line is the spectral envelope.
The spectral envelope consists of a plurality of resonance peaks called formants. Although the fine structure is modeled by harmonic components of the pitch, it is not necessarily composed only of harmonic components, and a method of modeling a portion where the level of the spectral envelope is small by white noise is appropriate. The present invention provides a method for determining the frequency range of the part modeled by the white noise. The frequency range modeled by this white noise is the range indicated by G and H in FIG. 3A, and it can be seen that this portion is the valley portion of the spectral envelope. Therefore, when the spectrum envelope is obtained, it is compared with the threshold value, and the frequency range below the threshold value is the part that replaces the harmonic component of the desired pitch with white noise.

Next, a method for obtaining the spectrum envelope will be described. The spectral envelope is given by the formant synthesis, as described above. It is well known both theoretically and experimentally that one formant can be represented by a single-tuned resonant circuit. Therefore, the formant center frequency and 3d
If the B bandwidth (or resonance Q) and the formant amplitude are determined, the spectral envelope can be described as a combination of the frequency characteristics of the single-tuned resonance circuit based on these.
This is shown in FIG. 3 (b). Here, the portions indicated by the dotted lines of J and K are the frequency characteristics of the single-tuned resonance circuit, and the solid line is the combined frequency characteristics. Therefore, a common part obtained by superimposing the results of comparing these respective resonance frequency characteristics with the threshold value is the range to be obtained.
It is shown in FIG. In FIG. 3 (c), 1 indicates a portion larger than the threshold, and 0 indicates a portion smaller than the threshold.

The method of obtaining a signal in which the harmonic component of the pitch and the white noise are mixed based on the function of the portion larger than the threshold value as a function of the frequency obtained in this way, firstly generates the harmonic component of the pitch. 3 (c) in the frequency domain. That is, the second
This is the product of the output of the pitch pulse generator 32 and the output j of the processing control circuit 34, that is, the function signal j of FIG. 3 (c).
The processing to multiply in the frequency domain is a band elimination filter,
It can be realized by a combination of a low-pass filter, a high-pass filter or a band-pass filter. Similarly, the band-limited white noise filter is obtained by replacing the function 1 in FIG. 3 (c) with 0 and replacing 0 with 1 (signal k in FIG. 2) and the white noise generator 33 in FIG. And the output is multiplied by the output control circuit 36 in the frequency domain.

According to the above method, a drive signal having a fine structure in which pitch harmonics and band-limited white noise are mixed can be obtained.

Next, the center frequency of the formant, the 3 dB bandwidth, and the amplitude of the formant are closely related to the parameters obtained as a result of the linear prediction analysis, and are all obtained from the parameters. Details will be described later as (Note).

Therefore, according to the method of the present invention, in the linear predictive speech synthesizer, the driving sound source which was approximated only by the pitch harmonic at the time of the voiced sound was synthesized with the pitch harmonic and the band-limited white noise. It can be a driving sound source, and
It is not necessary to perform any processing on the analyzer (encoder side for transmission), and all the processing is performed only on the synthesizer side (decoder side on reception). It is not necessary to increase the amount of information to be transmitted.

(Note) The relationship between the linear prediction analysis parameters and the physical quantities of formants (center frequency, bandwidth, amplitude) is shown here. The parameters obtained by linear prediction analysis are not transmitted as they are from the encoder to the decoder, but are converted to reflection coefficients, PARCOR coefficients or LAR coefficients, etc. It can be easily converted.

All-pole system function H identified by linear predictive analysis
(Z) is expressed as 1 / A (z), and the root of A (z) = 0 gives a pole, which corresponds to a formant. Assuming that the linear prediction parameters are {α _i } _{i = 1, 2,..., P} , the root of A (z) = 1 + α ₁ Z ⁻¹ + α ₂ Z ⁻² +... + Α _p Z ^−p = 0 Therefore, the center frequency and the bandwidth of the formant are obtained as follows.

Also, the formant amplitude is determined by the PARCOR coefficient ｛k _i ｝
_{By making} the loss at the glottis of the pseudovocal tract model determined by _{i = 1, 2,..., p} lossless, the all-pole system function H
It is obtained by converting (z) into a line spectrum. The root that gives the pole of the line-spectralized all-pole system function H _{p + 1} (z) is And λ _i gives the line spectral frequency. The amplitude can be obtained from the residue calculation at λ _i of the all-pole system function H _{p + 1} (z). As a result of the residue calculation, the all-pole system function can be expressed by the following equation.

Since the inverse Fourier transform of the power transfer function P _{p + 1} (λ) = H _{p + 1} (λ) H _{p + 1} ^* (λ) becomes the autocorrelation function Pτ, it can be expanded as the following equation. .

Than this, by solving the simultaneous equations using measured values of Ptau, it is possible to obtain a line spectral amplitude m _i.

(Effects of the Invention) As described in detail above, the present invention is applied to linear prediction analysis
By applying the present invention as a decoder for an encoded speech signal by a synthesis method, the naturalness of the synthesized speech can be enhanced.

[Brief description of the drawings]

FIG. 1 is a block diagram of a conventional linear prediction analysis / synthesis type encoder and decoder, FIG. 2 is a circuit block diagram of a decoder showing an embodiment of the present invention, and FIG. 3 is a spectrum characteristic diagram. 11 ………………………………………………………………………………………………………………………………………………………………………………………………………………………………… Pitch period extractor
21,31 ... Separation circuit, 22,32 ... Pitch pulse generator, 2
3, 33 noise generator, 24 switcher, 25, 38 synthesis filter, 34 processing control circuit, 35, 36 output control circuit, 37 synthesizer.

Claims (1)

(57) [Claims] 1. A method for separating a linear prediction coefficient, a voiced / unvoiced discrimination signal, sound source intensity information, and pitch period information contained in a received digital signal in which a speech signal is analyzed and coded by a linear prediction analysis / synthesis type encoder. An output separation circuit, a pitch pulse generator that outputs a pitch pulse controlled by the pitch period information, a noise generator that outputs white noise, an output of the pitch pulse generator, and an output of the noise generator. And a synthesis filter that outputs a speech signal decoded according to the linear prediction coefficient as a driving sound source. A linear prediction coefficient and a voiced unvoiced identification signal and a sound source intensity from the separation circuit, Information is input, and a spectrum envelope on the frequency axis is obtained by synthesizing a voiced formant, and is compared with a predetermined threshold. A processing control circuit that outputs a pitch component function signal representing a frequency domain in which the level of the spectrum envelope is larger than a threshold, and a noise component function signal representing a frequency domain in which the level of the spectrum envelope is smaller than the threshold; A first output control circuit for multiplying an output of a pulse generator by the pitch component function signal to output a pitch pulse in a frequency region larger than the threshold, a white noise from the noise generator and the noise component function signal And a second output control circuit that outputs white noise in a frequency region smaller than the threshold value by multiplying the output of the first output control circuit and the output of the second output control circuit. A linear predictive analysis / synthesis type decoder, comprising: a synthesizer that outputs a driving sound source to the synthesis filter.