KR100628170B1 - Apparatus and method of speech coding - Google Patents

Abstract

The present invention relates to a speech compression apparatus and method comprising dividing a speech spectrum into a plurality of frames, assigning a frame classification to the plurality of frames, and determining speech modeling parameters based on the assigned frame classification. The voiced portion of the speech spectrum and the unvoiced portion of the speech spectrum are synthesized separately using an analysis by synthesis that enables accurate correspondence between the voiced and unvoiced portions of the reconstructed signal. In particular, the frequency response of a particular simulated signal based on previous and current frames is used as an approximating function. The simulated signal is synthesized at the encoder side in a manner that can be generated at the decoder side. In addition, two good encoding methods are selected for encoding the spectral magnitude. Therefore, the present invention can efficiently and accurately reconstruct synthesized digital speech.

Voice coding, coding device

Description

Apparatus and method for coding speech {APPARATUS AND METHOD OF SPEECH CODING}

1 is a block diagram showing an encoder of the present invention

2 is a diagram showing a Hamming window response scaling according to the present invention.

3 shows a direct method of speech model parameter determination according to the invention.

4 is a diagram illustrating an analysis method by synthesis for determining a speech model parameter according to the present invention.

5 is a block diagram showing a method for analysis by synthesis for spectral size determination according to the present invention.

6 is a block diagram showing hybrid encoding of spectral magnitude vectors in accordance with the present invention.

7 is a block diagram showing a decoder according to the present invention.

8 is a block diagram showing voiced speech synthesis in accordance with the present invention.

9 illustrates an example of band / frequency correspondence in voiced sound synthesis according to the prior art;

10 is a view showing frequency response of some bands in voiced sound synthesis according to the prior art.

11 shows an example of an excitation spectrum in voiced sound synthesis according to the prior art.

12 shows an example of band / frequency correspondence in the voiced sound synthesis method according to the present invention.

13 is a view showing frequency response of some bands for the voiced sound synthesis method according to the present invention.

14 shows an example of a planetary excitation spectrum obtained using the planetary synthesis procedure according to the present invention.

15 is a block diagram showing unvoiced speech synthesis

TECHNICAL FIELD The present invention relates to a communication system, and more particularly to a speech compression method of a communication system.

Many voice compression systems are known. In general, these systems can be divided into three types: time domain, frequency domain, and hybrid codec. However, for low bit rate coding, multi-band excitation (MBE) compression techniques provide the highest quality decoded speech.

The MBE vocoder encodes the obtained speech signal by first dividing the input speech into a limited frame. These frames are converted from the time domain to the frequency domain. The frequency spectrum of the framed and windowed signals is then calculated and analyzed. In addition, a pitch value, a set of voiced / unvoiced decisions for a frequency band, a spectral magnitude group, and corresponding phase values and The same voice model parameters are needed for speech synthesis in the MBE vocoder. In general, phase coding is not transmitted in low bit rate coding.

There are many methods of spectral estimation, all of which are based on the estimation of frequency bands by excitation functions. The most traditional type of excitation function is the frequency response of the Hamming window. However, Hamming windows only goodly estimate the initial spectrum for a stationary speech signal. For non-stationary speech signals, some kind of excitation function does not sufficiently match the actual form of the spectrum for accurate estimation. For example, a change in pitch frequency during the analysis can lead to widening of the peak in the spectral magnitude envelope. Thus, the peak width of a given excitation function no longer matches the width of the actual peak. In addition, if the analyzed speech frame is a mixture of two different processes, the spectrum takes a very complex form, making it difficult to accurately estimate it using some simple excitation function.

There are several techniques for encoding MBE parameters. Typically, simple scalar quantization is used to encode pitch values and band grouping methods are used to encode voiced / unvoiced crystals. The most difficult task is to encode the spectral magnitude, which is why Vector Quantization (VQ) and Linear Prediction are used. Many high efficiency compression methods have been proposed based on VQ, one of which is a hierarchical structured codebook method used to encode spectral size.

Although the VQ technique allows for accurate quantification in some problematic areas, it is generally effective for data similar to the data that was included in the "learning sequence". Other effective ways of encoding spectral magnitudes include intra-frame and inter-frame linear prediction. The intra-frame method is capable of properly encoding the spectral size, but the effect is significantly lower in low bit rate coding. The inter-frame prediction method is also a good method, but has a limitation that its use is only suitable for stationary speech signals.

Speech synthesis according to the prior art is performed according to an accepted speech model. In general, in an MBE vocoder, two components, the voiced and unvoiced portions of speech, are synthesized separately and then combined to create a complete speech signal.

The unvoiced component of speech is produced in a frequency band, which is determined to be unvoiced. In each voice frame, an arbitrary block of noise is windowed and transformed into a frequency domain, where the spectral portion corresponding to meteorological harmonics is set to zero. The remaining spectral components corresponding to the unvoiced part of the voice are normalized to the unvoiced harmonic magnitude.

Other techniques are used to generate negative oily components in the MBE method. Since voiced sound is modeled by individual harmonics in the frequency domain, the voiced sound can be implemented at the decoder as a bank of tuned oscillators. An oscillator is defined by its amplitude, frequency, and phase, and the oscillator is assigned to each harmonic in the planetary region of the frame.

However, variations in the measured parameters in adjacent frames can lead to discontinuities at the frame edges, which can significantly degrade speech quality. Thus, during compositing, the parameters of the current and previous frames are interpolated to smooth the transition at the frame boundary resulting in continuous voiced sound at the frame boundary.

You can vary the interpolation method (for amplitude, frequency, and phase). However, interpolation is generally satisfactory only at a constant pitch. If the pitch changes drastically, the implementation process rules do not yield satisfactory results due to traditional harmonic lacing associated with the same number of frequencies in neighboring speech frames. When the pitch frequency changes, a frequency difference occurs in the laced harmonics, and in the conventional band corresponding to the harmonic band, the difference becomes larger when the number of bands is larger and when the pitch change is larger. As a result, undesirable artifacts occur in the decoded speech.

Accordingly, an object of the present invention is to provide a method capable of improving the quality of speech spectrum estimation for both voiced and unvoiced bands.

Another object of the present invention is to improve the encoding efficiency of spectral size groups regardless of the bit rate in encoding.

Another object of the present invention is to improve the quality of speech synthesis.

Other advantages, objects, and features of the invention will be set forth in the description which follows, and in part will be obvious to those skilled in the art upon reviewing the invention, or may be understood by practice of the invention. The objects and advantages of the invention will be embodied and obtained in particular as pointed out in the claims.

To achieve the object of the present invention as described herein, speech spectral estimation is performed on a spectrum divided into a plurality of bands, depending on the pitch frequency of the speech frame. The pitch frequency of the audio signal is determined, the frequency band is configured, and the voiced / unvoiced determination of the frequency band is performed. Next, a synthesis analysis method called speech spectral estimation is used for the calculation of the spectral magnitude.

By more precise evaluation of the harmonic magnitude on the encoder side, the quality of the voiced portion of the reconstruction signal on the decoder side is improved. In addition, by more precisely calculating the magnitude in the unvoiced band of the spectrum, the quality of the noise portion of the reconstructed signal is improved. Using the synthetic analysis method in both the voiced and unvoiced bands, an accurate correspondence is made between the voiced and unvoiced portions of the reconstructed signal.

The present invention also improves the encoding efficiency of spectral size groups. For low bit rate encoding, the problem is to represent spectral size data with a certain number of bits. The present invention with respect to spectral size encoding presents two main challenges: reducing the initial amount of spectral size to a certain number and encoding the reduced group. The method according to the present invention effectively solves the first problem by using a Wavelet Transform (WT). And, if the speech signal is stationary, the second problem can be effectively solved using inter-frame prediction.

However, at time intervals involving signals that are not stopped, it is rather effective to not use prediction. In this case, wavelet transform technology can effectively solve the encoding problem. As the encoding efficiency increases, the quality of the reconstructed speech signal is improved at the same bit rate, or the bit rate required for the same quality level is reduced.

The present invention also improves the quality of speech synthesis. Speech synthesis is performed sequentially for each frame. Since the fundamental frequency becomes the reference for the whole band division of the spectrum to be estimated, a difference occurs in the frequency of the raced harmonics when the pitch changes. The present invention uses the frequency correspondence between the raced bands of the current and previous frame. This provides accurate and reliable speech synthesis processing with pitch frequency shifting and jumping. Obvious problems (errors) seen in pitch determination do not reach dramatic results with conventional methods.

Preferred embodiments of the present invention will be described with reference to the MBE encoding method. MBE vocoder is described in "Multiband Excitation Vocoder", IEEE Transactions on Acoustics, Speech, and Signal Processing, vol. 36, No. 8, 1988, pp. D.W.Griffin and J.S.Lim at 1223-35, incorporated herein in their entirety. Typically, the vocoder operates at a speech signal sampling rate of 8 kHz. The use of a speech signal encoder and decoder according to the present invention will be described.

, 유성 결정 그룹

, 스펙트럼 크기 그룹

, 그리고 위상값(비트전송률이 매우 낮기 때문에, 본 실시예에서는 위상값을 전송하지 않는다)과 같은 MBE 모델 파라미터를 결정하는데 이용된다.1 shows an embodiment of an encoder according to the invention, which comprises a voice model parameter determination apparatus 1 and a parameter encoding apparatus 2. The speech model parameter determination apparatus 1 includes a rectangular windowing unit 10, a hamming windowing unit 20, a frame classification unit 30, and a fast Fourier transform (FFT). A device 40, a pitch detection device 3, a V / UV discrimination unit 80, and a spectrum size determination device 90. These components are the pitch frequency

Meteor determination group

Spectral size group

And MBE model parameters such as phase values (since the present embodiment does not transmit phase values because the bit rate is very low).

The parameter encoding apparatus 2 includes a scalar quantification apparatus 100, a spectral size wavelet reducing apparatus 110, a spectral size hybrid encoding apparatus 120, and a multiplexer unit 130. These components are used to encode MBE model parameters into a plurality of bits. The pitch detection apparatus 3 also includes a pitch candidate group determining apparatus 50, an optimum candidate selecting apparatus 60, and a best candidate refinement unit 70.

_R에 의해 멀티플렉싱된다. 해밍윈도우장치(20)에서는, FFT 장치(40)에 의해 스펙트럼 계산이 수행되기 위해 신호가 또한 해밍윈도우 함수

_H에 의해 멀티플렉싱된다. 주파수 해상도(frequency resolution)를 증가시키기 위해, 처리된 프레임에 일련의 제로가 가산되는데, 상기 가산은 FFT를 수행하여 FFT_LENGTH 배열(array)을 만들어 내기 전에 이루어진다. 배열 FFT_LENGTH = 2048 로써 양호한 주파수 해상도를 얻을 수 있으나, 실시간(real-time) 적용에서는 FFT_LENGTH = 512의 배열값을 이용하였다.For parameter measurement, the speech signal is first divided into a 30ms overlapping segment with an advance of 20-24ms. In the rectangular window device 10, the signal is rectangular window function so that the frame classification is performed by the frame classification device 30.

Multiplexed by _R. In Hamming window device 20, the signal is also a Hamming window function for spectral calculation to be performed by FFT device 40.

Multiplexed by _H. To increase frequency resolution, a series of zeros are added to the processed frame, which is done before performing the FFT to produce an FFT_LENGTH array. Good frequency resolution can be obtained with the array FFT_LENGTH = 2048, but an array value of FFT_LENGTH = 512 was used for real-time applications.

_f를 생성한다. 프레임들은 두 가지 방법으로 분류된다. 첫 번째는 프레임을 따라 변화하는 신호값의 특징 및 범위에 의한 분류이고, 두 번째는 프레임 내의 신호 발진(oscillation) 특징에 의한 분류이다.The frame classification device 30 may be referred to as an auxiliary device to the MBE model of the prior art. The apparatus processes the speech frames in the time domain and uses the frame classification characteristics used for more reliable and reliable signal processing in the pitch candidate group determining apparatus 50 and the V / UV discriminating apparatus 80.

Create _f Frames are classified in two ways. The first is classification based on the characteristics and ranges of signal values varying along the frame, and the second is classification based on signal oscillation features within the frame.

In the first classification, the signal types in the frame were defined as shown in Table 1 below. This classification method is based on the simultaneous examination of feature changes and signal sample values within a frame. In addition, the signal oscillation type in the frame was defined as shown in Table 2 below. The second classification method is based on the zero value crossing the first and second portions of the current frame.

Table 1. (First way to classify speech frames)

SILENCE Very low amplitude frame with accidental single short noise peak FIZZLE Rather low amplitude frame with systematic noise peaks VOWEL Pure collection frames VOWEL_FADING Fading collection frame VOWEL_RISING Rising collection frames PAST_VOWEL Frames where fading bars appear only at the beginning of the frame BEFORE_VOWEL Frame where rising vowel appears only at end of frame CHAOS All other types of frames

Table 2. (Second way to classify speech frames)

WELK Frames That Don't Contain Oscillations VIBRATION Frame containing rash at the beginning and end PAST_VIBRATION Frame containing oscillation only at start BEFORE_VIBRATION Frame containing rash only at the end

As a result, frames are derived and defined with 32 (ie 8 × 4) combination types. The derived frame is embodied from the combination type through a logical operation. An example of a derived type of frame is as follows:

SOME_OF_VOWEL 'AND' NOT 'SOME_OF_VIBRATION, here

SOME_OF_VOWEL = VOWEL

'OR' VOWEL_FADING

'OR' VOWEL-RISING

'OR' PAST_VOWEL

'OR' BEFORE_VOWEL, and

SOME_OF_VIBRATION = VIBRATION

'OR' PAST_VIBRATION

'OR' BEFORE_VIBRATION.

The frame classification characteristics thus obtained are used for pitch detection and meteor determination. The operation of the pitch detection device 3 using the frame classification characteristic will be described.

Especially in real-time applications, one of the most important and difficult challenges is the reliable pitch frequency detection problem. The pitch detection method according to the present invention presents an effective and reliable solution based on analysis in time and frequency domain. Pitch frequency detection consists of three steps. First, the pitch candidate group determining apparatus 50 determines a pitch candidate group by using an auto-correlation function (ACF) analysis in the time domain. Second, the optimum candidate selecting device 60 selects an optimal candidate by measuring all candidates in the frequency domain. Then, the best candidate improving apparatus 70 improves the best candidate value in the frequency domain.

By using short-time center clipping before calculating the ACF, more reliable pitch detection can be achieved. After central clipping, low-pass filtering of the processed frame is performed. Depending on the frame type determined by the frame classifier 30, a direct or inverse order is used for the ACF calculation. Even if the same formula is used for direct and reverse order for the ACF calculation, the direct order is related to the sample pairs located at the beginning of the frame, while the reverse order is related to the sample pairs located at the end of the frame. For example, the reverse order of ACF calculation is applied to frame types such as VOWEL_RISING and BEFORE_VOWEL, and the direct order of ACF calculation is used to frame types such as VOWEL_FADING and PAST_VOWEL.

The pitch candidate group determining device 50 determines a pitch candidate group, where the candidate group is configured to obtain local maximums of the ACF located to the left of the time lag corresponding to the global maximum ACF value. It includes everything. The group may include a plurality of candidates for different frames. The precise search range in the frequency domain is defined by the frequency corresponding to the global maximum ACF value.

In the best candidate selection device 60, measurements are performed on all candidates of the obtained group to select the best candidate. In particular, the optimum pitch frequency is found near the pitch candidate value by using a weighted value and using the minimum value criterion of the summed squared error (SSE). Estimation is performed in the frequency domain, and measurement of quality estimation is performed as follows. According to the investigated pitch frequency p , the entire frequency range is divided into n frequency kk frequency bands having a width of p Hz. The speech spectrum of each band is approximated by a scaled Hamming window response and the SSE of the approximation is calculated by the following equation (1):

From here

The amplitude value A _i for the i -th band is calculated as follows.

Where S is the speech spectrum, W is the scaled Hamming window response, and a _i and b _i are the number of harmonics corresponding to the beginning and end of the i -th band.

Traditionally, Hamming window responses with constant width are used in the MBE method. However, extensive experiments have shown that using a fixed Hamming window response results in unjustified odds at low pitch frequencies (e.g., sub-harmonics with true pitch values). Turned out.

In a preferred embodiment, as shown in Fig. 2, a special scale factor corresponding to the irradiated pitch value is used to scale the Hamming window response. _Scaling is performed for frequencies below a certain frequency F _scale . The frequency value F _scale = 140 Hz was determined experimentally. In particular, resizing is performed as follows. At a given FFT_LENGTH value of the total number of harmonics of the spectrum obtained by the FFT transform, the first Hamming window response has a N _orig component that differs significantly from zero. At FFT_LENGTH = 2048, N _orig was 3.1. At the low fundamental frequency F _0exam < F _scale , the _scaled Hamming window response used as the excitation function should be sharper in shape. Thus, the array of response values should have N ₀ < N _orig components that differ significantly from zero. The number N ₀ of these components is calculated as follows. N ₀ = int [ N _orig ( FFT_LENGTH / 2048)] ( F _0exam / F _scale )]. To obtain a scaled response, a proportional sharpening procedure based on linear interpolation is applied to the original Hamming window response.

According to the present invention, the approximation of the frequency band corresponding to the low pitch frequency is further improved. The proper sharpness of the approximation function in each band allows the selection of true pitch candidates. In real-time applications, the _scaled Hamming window responses corresponding to different F _0exam < F _scale can all be tabulated and used as look-up tables. On the other hand, in order to avoid occurrence of unalignment at high pitch frequencies (e.g., multiple harmonics where the pitch value is true), the first band is enlarged under the SSE ₁ calculation in comparison with the other bands such that a ₁ = 1.

However, while selecting the best candidate, different parts of the spectrum may not be equally important. In view of this problem, the weighting factor Q is introduced into the SSE calculation. A piecewise-linear weight Q was used as in equation (3) below.

Here, bf is the number of harmonics corresponding to the start of fading of the weighted value and ef is the number of harmonics corresponding to the end of fading of the weighted value, where 0 ≦ bf < ef . The obtained optimum pitch frequency value is improved in the optimum candidate improvement apparatus 70 by finding the optimum pitch frequency value in the range close to the pitch candidate value using the minimum value of the approximation without applying the weight value.

An important feature of the MBE method is that voiced / unvoiced decisions are generated for each frequency band of the original spectrum rather than the entire frame. In addition, the harmonic component may exist within a frictional frame and the meteor frame may include some noise band. The generation of voiced / unvoiced crystals for each frequency band of the original spectrum is performed in the V / UV discriminator 80. In low bit rate performance in the MBE vocoder, the generation of the decision is performed in groups of frequency bands adjacent to each other. In a preferred embodiment adaptive band division is used (in proportion to the value of the number of bands).

The planetary discrimination process starts when a predetermined pitch value is obtained. For discrimination, the first spectrum is divided into frequency bands according to a predetermined pitch value and each frequency band is estimated by the scaled frequency response of the Hamming window. In addition, frequency response scaling is performed by the same technique for the same reason as described in the optimum candidate selection device 60. Scaling provides an accurate relationship between the width of the frequency band and the approximation window. In addition, it is very important to precisely adjust the position of the approximation window and the position of the frequency band peaks.

에 의해 증가되며, 명백한 유성 프레임에 대해서는 유성 결정을 하게 된다. 하지만, 프레임의 진폭 특성이 VOWEL 유형에 속하지는 않지만 VIBRATION 유형의 하나에 속할 때는 명백한 자음 프레임에 있어서, 문턱값은 나눔수

에 의해 감소되어 명백한 자음 프레임에 대해 무성 결정을 하게 된다. 분류가 명확하지 않다면, 문턱값은 변하지 않으며 미리 정의된 값을 갖는다.The Noise to Signal Ratio (NSR) value of the approximation quality defines the voiced / unvoiced properties of a group of frequency bands. The threshold value of the NSR value depends on the classification characteristics of the current frame. For example, if the amplitude characteristics of a frame belong to both VOWEL types but not the VIBRATION type, then the threshold is a divisor.

It is increased by, and the meteor determination is made for the apparent meteor frame. However, for consonant frames that are apparent when the amplitude characteristics of the frame do not belong to the VOWEL type but belong to one of the VIBRATION types, the threshold is divided by

Is reduced to make an unvoiced decision for the apparent consonant frame. If the classification is not clear, the threshold does not change and has a predefined value.

The measurement of approximation quality including NSR is calculated by the following equation.

Where NSR _i is the noise-to-signal ratio of the i -th band group, and includes bands from n _i to n _{i + 1} −1, where n _i is the number of bands in the first frequency band of the i -th group; Err _m is the summary squared error of approximation for the m -th band; S (k) is the magnitude of the k -th wave of the estimated spectrum; a _i and b _i are the number of harmonics corresponding to the beginning and end of the i -th band group, where a _i is the number of harmonics of the first harmonic of the n _i -th band and b _i is ( n _{i + 1-} It is the number of harmonics of the last harmonic of 1-th band.

Determination of Err _m is performed separately for each band included in the group. For this determination, position tuning of the scaled Hamming window response to the frequency band peak is performed in the meteor frame. This produces an accurate planetary / unvoiced crystal, which is done in the following way. The Err _m value is calculated at various locations in the approximation window with respect to the center of the frequency band. Then, the position of the approximation window corresponding to the minimum Err _m value is selected so that an optimal NSR _i value for the entire band group is obtained from the minimum Err _m values for each band included in the group. Thus, a voiced / unvoiced decision is generated using the NSR criteria as described above.

Spectral sizing using the AbS method will now be described. In general, the purpose of the analysis and coding function is to obtain at the transmitting side the data necessary for speech generation at the receiving side. According to the MBE model, speech generation is performed using speech model parameters including pitch values, the pitch values being a harmonic band system; Voiced / unvoiced decision groups for frequency bands; Spectral amplitude groups; The phase value is then derived.

이 측정된다. 모델 파라미터의 최적 벡터는 인코딩되어 합성장치에 전송된다.As shown in FIG. 3, the speech model parameters can be easily calculated directly or explicitly and output to the encoder. However, analysis by synthesis implicitly defines all or part of the speech model parameters before outputting the parameters to the encoder. Referring to FIG. 4, the conceptual construction method of the AbS method for parameter determination includes a quality evaluation device, a synthesis device, and a search device. The same synthesizing apparatus is used to generate speech at the transmitting side and the receiving side. To provide the model parameter group P is a synthetic signal s ~ search (t), the synthesized signal s ~ (t) is the closest to the actual speech signal s (t) by a predetermined evaluation standard. The search for the optimal group P is performed as an iterative process, where the vector value P is different for each iteration and the object function value

This is measured. The optimal vector of model parameters is encoded and sent to the synthesizer.

In one embodiment of the present invention, the spectral amplitude is measured using the AbS method based on the planetary / voiceless determination and the pitch frequency value measured by direct calculation. The AbS method for measuring the spectral amplitude will be interpreted as defined above. On the transmitting side, the same synthesizing apparatus as that of the receiving side is used for voice generation. Thus, on the transmitting side and the receiving side, the same rules are used to interpolate the amplitude, phase, and frequency from the previous frame to the current frame.

와 유성/무성 결정 그룹

에서, 스펙트럼 진폭 그룹

이 탐색되어, 스펙트럼 을 갖는 합성 신호

가 제공되는데, 상기 스펙트럼은 실제의 음성 신호 스펙트럼 S에 가장 가까운 것이다. 평가기준은 스펙트럼 에 의한 어림셈 스펙트럼 S의 최소 SSE 이다. 최적 스펙트럼 크기를 탐색하는 것은 반복처리로서 수행될 수 있다. 벡터 값 M은 각 반복마다 달라지며, 객체함수 값

이 측 정된다. 발견된 최적 값은 인코딩되어 전송된다.Specific pitch value

And meteor / voice decision groups

In, spectral amplitude group

Is searched, the synthesized signal having the spectrum

Is provided, the spectrum closest to the actual speech signal spectrum S. The evaluation criterion is the minimum SSE of the speculation approximation spectrum S. Searching for the optimal spectral size can be performed as an iterative process. The vector value M is different for each iteration, and the object function value

This is measured. The optimal value found is encoded and sent.

In a preferred embodiment of the present invention, one-iteration of sizing is presented because this method is suitable for real-time performance. The size determination according to the invention is based on the linearity of the Fourier transform and the linearization of speech signal processing.

하에서, 모델 음성 신호

가 스펙트럼 진폭의 배정 에탈론(단위) 값에 합성된다. 합성 신호의 스펙트럼

가 계산되어 실제 음성 신호의 스펙트럼 S에 비교된다. 상기 비교는 각 대역에서 별개로 이루어진다.A group of spectral magnitudes is formed by assigning a constant etalon value M _m = e (in particular, the etalon value is equal to the unit value: M _m = 1) to each size to be determined. Planetary decision groups V ^p , V ^c and specific pitch values for previous and current frames

Under the model voice signal

Is synthesized to the assigned etalon (unit) value of the spectral amplitude. Spectrum of Synthetic Signal

Is calculated and compared to the spectrum S of the actual speech signal. The comparison is made separately in each band.

Similar to the response analysis of unit disturbance in linear system theory, part of the spectrum _m related to the m-th band under the activity of the m-th spectral component of the unit amplitude is interpreted as the response of the linearized system. A portion of the actual spectrum S _m related to the m-th band can be estimated as in equation (5) below.

Where E _m is the error of approximation. The value μ _m is the factor when S _m is estimated in the best way. Therefore, μ _m values for all the bands is calculated using the least square method (Least Square Method).

에 의한 스펙트럼 S 어림셈의 요약 제곱 오차를 최소화하는 어림셈 계수가 결정된다. 선형성(또는 준-선형성(quasi-linearity))으로 인해, 곱셈 계수(multiplicative coefficient)가 스펙트럼 크기(M _m = e·μ _m 또는 M _m =l·μ _m )의 값으로 간주될 수 있으며, 이 때문에 합성된 신호는 실제 음성 신호의 스펙트럼 S에 가장 가까운 스펙트럼

를 갖는다. 이들 값 μ _m 는 인코딩되어 전송된다. 수신측에서는, 출력 음성 신호의 합성에 스펙트럼 진폭값을 배정하는데 상기 값들이 이용된다.As a result, the spectrum

An approximation coefficient that minimizes the sum of squared errors of the spectral S estimates is determined. Linear (or quasi-linear (quasi-linearity)) due to, and multiplication coefficient (multiplicative coefficient) can be regarded as the value of the spectral amplitude (M _m = e · μ _m or M _m = l · μ _m), the Therefore, the synthesized signal is the spectrum closest to the spectrum S of the actual speech signal.

Has These values μ _m are encoded and transmitted. On the receiving side, these values are used to assign spectral amplitude values to the synthesis of the output speech signal.

A block diagram showing spectral sizing according to the present invention is shown in FIG. The calculation of the meteor and voiceless size is performed separately. In particular, the calculation of the meteor spectral magnitude is performed by the Bands' Correspondence Map Building unit 91, the planetary synthesis device 92, the Hamming window device 93, the FFT device 94, and the meteor size evaluation device. Performed by 95. The band compatible image forming apparatus 91 and the voice synthesizer 92 used to generate the voiced excitation spectrum are combined with the band compatible image forming apparatus 160 and the voice synthesizer 170 used for voiced sound synthesis at the decoder side shown in FIG. Same thing. Since the excitation signal is synthesized at the encoder side in a manner generated at the decoder side, the frequency response is well suited to be used as an approximation function.

현재 프레임에 대한 유성 결정 벡터

현재 프레임에 대한 스펙트럼 크기 벡터

그리고 대역대응사상구성장치에 의해 작성된 대역대응사상을 포함한다. 대역대응사상구성장치(91) 및 유성합성장치(92)의 상세한 작동이 디코더 측을 참조하여 후에 기술될 것이다(대역대응사상구성장치(160) 및 유성합성장치(170)를 참조). 그러나 상기 장치들은, 현재 프레임에 대한 소정의 입력 파라미터 그룹 하에서 그리고 유성합성장치에 내장된 이전프레임파라미터누산기(Previous Frame Parameters Accumulator)에 저장된 이전 프레임에 대한 유사 파라미터 그룹

하에서 출력 음성 신호를 시간영역에서 합성한다는 사실을 주목할 필요가 있다.As described above, the input parameter group for the planetary synthesizer is the pitch frequency for the current frame.

Meteor determination vector for current frame

Spectral magnitude vector for the current frame

And the band-corresponding phase created by the band-corresponding phase forming apparatus. The detailed operation of the band-compatible phase composing device 91 and the planetary synthesis device 92 will be described later with reference to the decoder side (see the band-compatible phase configuring device 160 and the planetary synthesis device 170). However, the devices are similar parameter groups for the previous frame under the predetermined input parameter group for the current frame and stored in the previous frame parameter accumulator built in the synthesizer.

It is worth noting that the output speech signal is synthesized in the time domain under the following conditions.

To make the planetary excitation spectrum, the spectral amplitude for the meteor band is determined by assigning a constant etalon value, wherein the etalon value is one. Assuming that the components of the meteor determination vector are 1 in the meteor band and 0 in other cases, the assignment can be expressed as follows.

The signal output by the planetary synthesis device 92 is windowed by the hamming window device 93 and processed by the FFT device 94. After the conversion is made, the output signal exhibits the planetary excitation spectrum S ^v_e . An example of the planetary excitation spectrum obtained by the planetary synthesis procedure according to the present invention is shown in FIG. 14. The voiced portion of the spectrum has a constant structure, while the unvoiced portion of the spectrum is near zero. The spectrum has similar characteristics even under varying pitch frequencies and planetary crystals, which are important for accurate spectral estimation.

The obtained planetary excitation spectrum is used for planetary size evaluation in the planetary size evaluation device 95. The planetary size evaluation device 95 separately estimates the planetary band of the actual spectrum S by the excitation spectrum S ^v_e by performing the size measurement using the least square method. By moving the spectrum to both sides with respect to the band center, the clip position of the excitation spectrum relative to the frequency band is adjusted in the planetary frame. Then, the position of the excitation spectral clip that provides the best NSR in the approximation is selected for magnitude estimation, which is performed using the least square method.

The resulting planetary magnitude group M ^{( v)} is only part of the M spectral magnitude vector. The unvoiced spectral magnitude group is another part of the M spectral magnitude vector. The calculation of the unvoiced spectral magnitude is performed by the synchronous noise generation unit 96, the hamming window device 97, the FFT device 98, and the unvoiced size evaluation device 99 shown in FIG. The synchronous noise generating device 96 produces a white noise signal having a unit amplitude range. As with the processing to obtain the meteor magnitude value, the noise is processed in the same way on the encoder and decoder side. In addition, on the encoding side, synchronization characteristics are provided to better estimate the unvoiced spectrum.

The signal obtained by the synchronous noise generating device 96 is windowed by the hamming window device 97 and processed by the FFT device 98. In the unvoiced size evaluation device 99, the spectral magnitude for each unvoiced band is calculated using the least square method. The unvoiced spectral magnitude group M ^(uv) thus obtained is combined with the planetary magnitude group M ^(v) to obtain a spectral magnitude vector M.

[50,400]으로 제한되어 256 레벨(8비트)로 정량화된다. 이 경우의 피치 주파수 표현(representation)의 최대 오차는 0.684 Hz이다. 결정된 정량화 값은 멀티플렉싱장치(130)로 전달된다. 또, 유성/무성 결정 그룹의 벡터 V는 멀티플렉싱장치(130)로 전달만 된다.Referring to FIG. 1, encoding of a speech model parameter according to the present invention includes three types. The encoding of the pitch frequency is performed by the scalar quantifier 100. The pitch frequency value is in the frequency range, for example

Limited to [50,400] and quantified to 256 levels (8 bits). In this case, the maximum error of the pitch frequency representation is 0.684 Hz. The determined quantification value is transferred to the multiplexing device 130. In addition, the vector V of the voiced / unvoiced decision group is only transmitted to the multiplexing device 130.

The spectral magnitude vector M is encoded in two steps. First, the spectral magnitude vector is reduced by the spectral size wavelet reducing apparatus 110. Next, hybrid encoding is performed by the spectral size hybrid encoding apparatus 120 on the reduced spectral magnitude vector. The reduction of the spectral magnitude vector according to the present invention will be described in detail.

First, the logic of the vector M element is represented by the following formula:

가 일정한 차원 r의 벡터

로 변환된다. 벡터

를 인코딩하는데 WT를 더 이용하는 것을 고려하면, r=1·2 ⁿ 이 되도록 r을 선택할 수 있으며, 여기에서 l은 양의 정수이고 n은 상기 WT의 예상되는 단계 횟수이다. 바람직한 실시예에서, 벡터

의 차원수(dimensionality)는 n=3, l=2을 갖는 r=16 의 값으로 감소된다.Where m defines the dimension of the vector M. The value of m depends on the pitch frequency and changes over time. Next, dimension m

Vector of constant dimension r

Is converted to. vector

Considering further using WT to encode, r = 1 · 2 ⁿ R can be selected, where l is a positive integer and n is the expected number of steps in the WT. In a preferred embodiment, the vector

The dimensionality of is reduced to the value of r = 16 with n = 3 and l = 2 .

이 재구성될 수 있다. 벡터

의 차원이 r 이라면, 감소 연산을 할 필요가 없다. 다른 경우에는, 다음의 단계를 포함하는 절차가 수행된다:Decrease operations are irreversible. However, if you follow the described method, you can

This can be reconstructed. vector

If the dimension of is r, then there is no need to perform a decrement operation. In other cases, a procedure is performed that includes the following steps:

요소에 기초한 삼차스플라인(cubic spline)이 구성된다;-Vector

A cubic spline based on urea is constructed;

the smallest number S is calculated such that s = r.2k ≧ m, k = 0,1,2 , ...;

A new uniform grid with S nodes is constructed and the value of the cubic spline is calculated at these nodes;

K number of wavelet transform steps are applied to the obtained s value group;

요소이고, 반면 하이-패스(high-pass) 계수는 버려진다.The resulting r low-pass wavelet coefficient is a vector

Element, while the high-pass coefficient is discarded.

The number k of the WT steps in the above process is not fixed and may be different in different signal frames.

는 두 번 인코딩된다. 즉, 웨이브렛인코딩장치(121)에서 웨이브렛 구성방법을 따라, 그리고 인터-프레임예측인코딩장치 (Inter-Frame Prediction Encoder unit)(122)에서 인터-프레임 예측 구성방법을 따라 인코딩된다. 두 번의 인코딩이 이루어지는 동안, NSR 판정기준을 이용하여 각 구성법의 효율이 측정되어 더 우수한 구성법이 콤퍼레이터(comparator unit)(123)에 의한 벡터

에 대한 기준 인코딩 구성법으로 선택된다.With reference to FIG. 6, hybrid encoding of spectral magnitude vectors according to the present invention will be described in detail. vector

Is encoded twice. That is, the wavelet encoding apparatus 121 is encoded according to the wavelet construction method and the inter-frame prediction encoding unit 122 according to the inter-frame prediction construction method. During two encodings, the efficiency of each construct is measured using the NSR criterion so that a better construct is obtained by the vector by the comparator unit 123.

Is selected as the reference encoding scheme for.

에 적용된다. l 로우-패스와 r-l 하이-패스 웨이브렛 계수 모두 정량화된다. 격자정량화기술(lattice quantization technique)이 로우-패스 웨이브렛 계수를 인코딩하는데 이용되고, 적응스칼라정량화(adaptive scalar quantization)가 하이-패스 웨이브렛 계수에 적용된다. 하이-패스 웨이브렛 계수의 속성 때문에, 제로를 중심으로 대칭을 이룬 스칼라정량기(scalar quantizer)가 구성된다. 바람직한 실시예에서, WT 단계의 횟수 n=3이고 로우-패스 웨이브렛 계수의 개수 l=2이다. 배직교필터(biorthogonal filter)(5,3)가 감소 단계와 인코딩 단계에서 WT 필터로 이용된다.In the wavelet encoding apparatus 121, n number of steps of WT are vectors

Applies to Both low-pass and rl high-pass wavelet coefficients are quantified. Lattice quantization techniques are used to encode low-pass wavelet coefficients, and adaptive scalar quantization is applied to high-pass wavelet coefficients. Because of the nature of the high-pass wavelet coefficients, a scalar quantizer is constructed that is symmetric about zero. In a preferred embodiment, the number of WT steps n = 3 and the number of low-pass wavelet coefficients l = 2. A biorthogonal filter (5,3) is used as the WT filter in the reduction and encoding steps.

In the inter-frame prediction encoding apparatus 122, inter-frame prediction for encoding the spectral size is used as a competitive encoding scheme. Inter-frame prediction takes advantage of the similarity of the spectral magnitudes of neighboring frames and, for fixed signals, is highly efficient. The prediction error is encoded using adaptive scalar quantification.

Simultaneously with encoding, a decoding process is performed. This is necessary for both inter-frame prediction configuration operation and quality measurement of the test encoding apparatus. The simultaneous use of competing encoding schemes such as wavelets and inter-frame prediction can further increase the efficiency of the method of the present invention. Accordingly, the comparator 123 compares the efficiency of the two configuration methods and sends the data and the decision bit corresponding to the optimum to the multiplexing device 130. The multiplexing device 130 forms a bitstream by combining all coded parameter values into a plurality of output bits.

7 is a block diagram illustrating a decoder, which decodes input bits to synthesize synthesized digital speech. Demultiplexing unit 140 separates a plurality of input bits according to an accepted data structure. The model parameter decoding apparatus 150 preliminarily decodes the parameter, thereby determining the output voice. The model parameter decoding apparatus 150 operates in a manner opposite to the model parameter encoding apparatus (see the scalar quantification apparatus 100, the spectral size wavelet reducing apparatus 110, and the spectral size hybrid encoding apparatus 120).

Band-correspondence configuration device 160 constructs a map, which maps the race pairs of frequency bands by using the pitch frequency values for the current and previous frames. The voiced sound portion is generated by the voice synthesizer 170 and the unvoiced sound portion is generated by the voice synthesizer 180. A summing unit 190 combines the outputs of the voiced and unsynthesized devices 170 and 180 to produce a synthesized digital voice.

The planetary part of the synthesized signal S ^v (n) is made as the sum of the appropriate harmonic components, which is expressed as

,n=0,..., L-1은 m-번째 주파수대역에 해당하는 조파 성분 신호이고, L은 음성 프레임에서 겹치지 않는 부분의 길이이고, 그리고 I ^v는 유성 대역으로 결정된 주파수대역 그룹이다.From here,

, n = 0, ..., L-1 is the harmonic component signal corresponding to the m-th frequency band, L is the length of the non-overlapping portion of the speech frame, and I ^v is a frequency band group determined as a meteor band .

는 다음과 같이 표현될 수 있다.In addition, a harmonic component signal using a time index (number of samples) n in a frame.

Can be expressed as follows.

Where A _m (n) represents the amplitude of the m-th harmonic interpolated between the beginning and the end of the frame, and θ _m (n) represents the phase of the harmonic signal.

Although there are important problems in speech synthesis, such as interpolation of harmonic amplitudes, interpolation of harmonic angular frequencies, and continuity of harmonic phases, one of the most important problems is the interworking of inter-frame frequency bands. It can be caused by an interaction. In a vocoder similar to the MBE vocoder, the harmonic components of the current frame are raced together with the harmonic components of the previous frame to perform synthesis. In the prior art, harmonics relating to the same number of frequency bands in neighboring speech frames are raced.

In a preferred embodiment of speech synthesis according to the invention, the harmonics related to approximately the same frequency are raced on the basis of the configured frequency band correspondence phase. A block diagram showing voiced sound synthesis according to the present invention is shown in FIG.

유성 결정 벡터

그리고 스펙트럼 크기 벡터

를 포함하고, 또 대역대응사상구성장치(160)에 의해 구성된 대역대응사상을 포함한다. 이전프레임파라미터누산기(171)에 저장된 이전 프레임의 파라미터 그룹

이 또한 음성 합성을 위해 이용된다. 레이싱제어장치(Lacing Controller unit)(172)는 레이싱된 대역의 유성 상태에 따라 어림셈 유형을 선택함으로써 위상보간장치(Phase Interpolator unit)(173), 각주파수보간장치(Angular Frequency Interpolator unit)(174), 진폭보간장치(Amplitude Interpolator unit)(175)의 작동을 조절한다. 제어오실레이터뱅크장치(Bank of Controlled Oscillators unit)(176)는 방정식 (7)을 이용하여 유성음 합성을 제공한다.The input parameter group for voiced synthesis is the pitch frequency for the current frame.

Meteor crystal vector

And spectral magnitude vector

It includes, and also includes a band-compatible corresponding image configured by the band-compatible image forming apparatus 160. Parameter group of previous frame stored in previous frame parameter accumulator 171

This is also used for speech synthesis. The racing controller unit 172 selects an approximation type according to the meteor state of the race band, thereby allowing a phase interpolator unit 173 and an angular frequency interpolator unit 174 to be selected. Adjust the operation of the amplitude interpolator unit 175. The Bank of Controlled Oscillators unit 176 provides voiced sound synthesis using Equation (7).

A prominent feature of the present invention is the fact that there is a band-corresponding phase configuring device 160 that determines the manner of harmonic racing. In the prior art, harmonics relating to the same number of frequency bands in neighboring speech frames are raced. An example of band / frequency correspondence in prior art harmonic synthesis is shown in FIG. The pitch frequency of the previous frame is 100 Hz, the pitch frequency of the current frame is 83.7 Hz, the number of bands of the previous frame is 39 and the number of bands of the current frame is 47. As shown, small changes in pitch frequency result in large changes in frequency, especially when the number of harmonics is large.

In Fig. 10, the frequency response of the seventh, eighteenth, and thirty-third harmonic bands according to the prior art is shown. These bands are meteor in both current and previous frames. In the correspondence of the above-mentioned harmonic bands, the frequency difference (eg, the seventh, eighteenth, and thirty-third bands of FIG. 10) of the raced harmonic causes a difference in amplitude and width frequency response. This results in the interaction of several different frequency band responses and also results in a distorted form of excitation spectrum as shown in FIG. In addition, if the pitch jumps, undesirable artifacts occur in the decoded speech.

An example of band / frequency correspondence in harmonic synthesis according to the present invention is shown in FIG. Harmonic synthesis is performed based on direct and inverse maps that provide a correspondence between the frequency bands of the current and previous frames. As shown, the number of the corresponding band, but may be different, the frequency band is hardly changed at the beginning and end of the frequency range (see the Δ f in band 33 of Figure 12).

Frequency response in the seventh, 18th, and 33rd harmonic bands according to the present invention is shown in FIG. 13, and as shown, the harmonic bands have the same amplitude and width. A small hill near the main peak corresponds to the harmonic fading of the previous frame. The frequency response of the excitation signal is shown in Figure 14, which shows a constant structure. It should be noted here that the different bands do not interact and do not overlap with the excitation signal configured. This makes it possible to evaluate the amplitude more accurately and reliably without causing dramatic results due to changes in pitch frequency.

Thus, in the current and previous frames, pairs of harmonics with the most similar frequency are selected and raced. The harmonics of the previous non-laced frame smoothly decrease to zero amplitude and the harmonics of the current non-laced frame smoothly increase to a predetermined amplitude.

또는

. 만약 현재 프레임의 주파수대역 m _c (또는 이전 프레임의 m _p )가 유성으로 결정되면, m_c ∈ I _P ^V 또는 m_p ∈ I _P ^V 이다. 만약 주파수대역 m _c 또는 m _p 가 무성으로 결정되면, m_c

I _c ^V 또는 m _p

I _P ^V 이다.

를 현재 프레임에 대한 피치 주파수라고 하면 (ω _o ^C =2f _o ^C ); 그리고, N _c 를 주파수대역의 개수라고 하면 (N _c =

), 여기에서,

는 샘플링 주파수의 값이다. 그러면, {M _mc }, m _c = 0,..., M _c -1 은 각 주파수대역에 대한 크기 그룹이고 I _c ^v 는 유성 대역으로 결정되는 주파수대역 그룹이다. 이전 프레임에서와 마찬가지로,

; N _p ; {M _mp }, m _p = 0,..., M _p -1 ; 그리고 I _p ^v 는 피치 주파수, 주파수대역의 개수, 크기 그룹, 그리고 유성 주파수대역 그룹이다.The following are the symbols used in the detailed description of the present invention. In the frequency bands m _c and m _p in the current and previous frames,

or

. If the frequency band m _c of the current frame (or m _p of the previous frame) is determined to be meteoric, then m _c ∈ I _P ^V or m _p ∈ I _P ^V. If the frequency band m _c or m _p is determined to be unvoiced, m _c

I _c ^V or m _p

I _P ^V.

Is the pitch frequency for the current frame (ω _o ^C = 2 f _o ^C ); And if N _c is the number of frequency bands ( N _c =

), From here,

Is the value of the sampling frequency. Then, { M _mc }, m _c = 0, ..., M _c - 1 is a size group for each frequency band and I _c ^v is a frequency band group determined as a meteor band. As in the previous frame,

; N _p ; { M _mp }, m _p = 0, ..., M _p - 1 ; And I _p ^v is the pitch frequency, the number of frequency bands, the magnitude group, and the meteor frequency band group.

, 즉

이고, 합성 신호

의 유성 부분은 적절한 대역 쌍 m = 0,..., N _p -1을 합함으로써 계산되며, 그 방정식은 아래에 보이는 바와 같다.Voiced sound synthesis is performed by the control oscillator bank device 176. The function of the control oscillator bank device 176 can be expressed by the following formula. If the pitch frequency increases,

, In other words

, Composite signal

The planetary part of is calculated by summing the appropriate band pair m = 0, ..., N _p -1 , the equation is shown below.

이다. 이때,

은 이전 및 현재 프레임의 주파수대역들 사이의 대응을 제공하는 직접사상이다.Where m is the number of band pairs. The m -th band pair < m _p , m _c > consists of the m _p band and the m _c band, where m _p = m and

to be. At this time,

Is a direct mapping that provides a correspondence between the frequency bands of the previous and current frames.

, 즉

이고, 합성 신호 S ^v (n)의 유성 부분은 적절한 대역 쌍 m = 0,..., N _c -1 을 합함으로써 계산될 수 있고 아래의 방정식으로 표현될 수 있다.If the pitch frequency decreases,

, In other words

And the planetary portion of the composite signal S ^v (n) can be calculated by summing the appropriate band pairs m = 0, ..., N _c -1 and can be represented by the following equation.

이다. 함수

은 현재 및 이전 프레임의 주파수대역들 사이의 대응을 제공하는 역사상이다. The m -th band pair < m _p , m _c > consists of the m _p band and the m _c band, where m _c = m and

to be. function

Is the history of providing a correspondence between the frequency bands of the current and previous frame.

, 즉

이고, 합성 신호 S ^v (n)의 유성 부분은 사상 없이 계산될 수 있으며, 방정식은 아래에 보이는 바와 같다.If the pitch frequency is constant,

, In other words

And the planetary part of the composite signal S ^v (n) can be calculated without mapping, and the equation is shown below.

그리고

가 m-번째 대역 쌍 ＜m _p , m _c ＞에서 만족된다면, 지속적인 조파가 생성된다. 진폭 보간은 아래의 공식에 따라 수행된다:The racing controller 172 controls the operation of the phase interpolator 173, the angular frequency interpolator 174, and the amplitude interpolator 175. Depending on the meteor state of the race band, three methods are available for interpolation. Condition

And

If is satisfied in the m -th band pair < m _p , m _c >, a continuous harmonic is produced. Amplitude interpolation is performed according to the following formula:

Here, M and M _{mp mc} is the size value for the previous and the current frame associated with the m _p and m _c band; n = 0, ..., L-1 is the number of samples; L is the length of the nonoverlapping portion of the speech frame; And R is the length of the hunting interval (0 < R < L ).

Interpolation of phase and angular frequency is performed according to the following formula:

From here,

Here, ø _mc (0) represents the corresponding phase of the m _c -th harmonic at the start of the current frame, which is the same as the phase of the harmonics at the end of the non-overlapping portion in the previous frame. That is, θ _mc (0) = 0 _mp (L) .

그리고

이면, 페이딩 조파가 생성되어 아래의 방정식 (14),(15)에 의해 보간이 수행된다.For the m -th band pair < m _p , m _c >

And

In this case, fading harmonics are generated and interpolation is performed by the following equations (14) and (15).

그리고

이면, 상승 조파(rising harmonic)가 생성되어 아래의 방정식 (16),(17)에 의해 조파 진폭 보간이 수행된다.For the m -th band pair < m _p , m _c >

And

In this case, a rising harmonic is generated, and harmonic amplitude interpolation is performed by the following equations (16) and (17).

Here, ø _mc (0) is the initial phase value ø ₀ Equivalent to, denotes the phase of the m _c -th harmonic at the beginning of the current frame.

Fig. 15 is a block diagram showing unvoiced sound synthesis according to the present invention, which includes a synchronous noise generator unit 181 on the decoder side, synchronized to a synchronous noise generator 96 on the encoder side. Therefore, the noise used for synthesis by the decoder is the same as the noise used for analysis by the encoder. The white noise signal waveform on the time axis is obtained from the white noise generator and is windowed by the hamming window device 182. The result is processed by the FFT device 183. Spectrum of the noise signal is multiplexed by the size M _m of the band determined as silent, the amplitude of the planetary band is set to zero.

Spectral transformation is performed by a Noise Spectrum Transformation unit 184. The transformed spectrum is subjected to an inverse fast Fourier transform by the IFFT device 185 using the phase value of the original noise signal. Then, in the Add and Overlap unit 186, the obtained noise signal is superimposed with the noise signal of the previous frame stored by the buffer 187 to produce an unvoiced portion. In the adder 190, a synthesized digital voice is made by adding voiced and unvoiced portions.

The above described embodiments are merely illustrative and are not intended to limit the present invention. The teachings described herein can be readily applied to other types of devices. The details of the invention are exemplary and are not intended to limit the scope of the claims. Many variations, modifications, and improvements are possible to those skilled in the art.

Included in the above.

Claims (46)

Divide an audio signal into a plurality of frames, determine a pitch frequency, determine voice information, indicate whether each of the plurality of frequency bands in each frame should be synthesized as a voiced or unvoiced band, and process the voice frame to frequency band A synthesis analysis method for determining spectral envelope information in a speech coding system based on synthesizing a synthesized digital speech signal from a data structure created by determining spectral envelope information indicative of a spectral magnitude of a) forming a group of spectral magnitude models by assigning fixed etalon values; b) synthesizing a model speech signal for the spectral magnitude model group under the voice frequency determination group and for the pitch frequency determined for the previous and current frames in the same manner as the received speech signal is synthesized at the decoder; c) calculating a spectrum of the model speech signal; d) estimating the spectrum of the actual speech signal by the spectrum of the model speech signal; And e) encoding a coefficient obtained from said estimated spectrum. The method of claim 1, In said step a), said spectral magnitude model group is formed separately for the voiced and unvoiced portions of a model speech signal spectrum. The method of claim 2, In step a), the spectral magnitude model group for the meteor portion of the spectrum is formed by assigning a fixed etalon value of 1 in the meteor band and otherwise assigning 0. The method of claim 2, And in step d), the planetary portion of the spectrum is estimated by the position of tuning the planetary excitation spectrum clip to the frequency band position using the Least Square Method. The method of claim 2, In step b), the unvoiced portion of the model speech spectrum is synthesized by producing a white noise signal in a unit amplitude range to provide synchronization characteristics of the synthesis method. The method of claim 2, In said step d), the unvoiced portion of the model speech signal spectrum is estimated by unvoiced excitation spectrum for each frequency band using least squares method. a) reducing the number of spectral sizes; b) evaluating a plurality of spectral size encoding methods; c) performing a spectral size encoding by selecting from among a plurality of spectral size encoding methods according to the evaluation. The method of claim 7, wherein In said step a), reducing said number of spectral magnitudes is based on a wavelet transform technique. The method of claim 8, In step b), the plurality of spectral magnitude encoding methods are wavelet transform technology and inter-frame prediction. Divide the speech signal into a plurality of frames, determine the pitch frequency, determine the meteor information, indicate whether each of the plurality of frequency bands in each frame should be synthesized as a voiced or unvoiced band, and process the speech frame to process the frequency band A method of synthesizing a synthesized digital speech signal from a data structure created by determining spectral envelope information indicating a spectral magnitude of a) configuring a frequency correspondence between the bands of the current and previous frame; b) synthesizing voiced sound components and racing the pairs of harmonics using pairs of harmonics having frequencies most similar to each other in the current and previous frames using the configured band frequency response, wherein all unpaired in the previous frame The harmonics smoothly decrease to zero amplitude and all unpaired harmonics in the current frame increase smoothly to a predetermined amplitude; c) synthesizing a speech component for the unvoiced frequency band; And d) synthesizing a speech signal by combining the synthesized speech components for the voiced and unvoiced frequency bands. The method of claim 10, In said step a), the band frequency correspondence is configured by forming direct and inverse maps of the frequency band system derived by the pitch frequencies of the previous and current frames. In the coding and decoding system of a speech signal, The voice signal coder A processor for dividing the input digital voice signal into a plurality of frames and analyzing the input digital voice signal in a time and frequency domain; An orthogonal transforming unit for transforming each frame to provide spectral data on a frequency axis; A pitch determination device for determining a pitch frequency for each frame; A voiced / voiceless determination device for generating a voiced / voiceless decision group using the determined pitch frequency; A spectral size determination device for measuring spectral magnitudes using an analysis method by synthesis; And And a parameter encoding device for encoding the determined pitch frequency, the measured spectral magnitude, and the voiced / unvoiced decision for each of the plurality of frames to combine the encoded data into a plurality of bits, The voice signal decoder A parameter decoding device for decoding the plurality of bits and providing pitch frequency, spectral magnitude, and voice / voice determination for each of the plurality of frames; A band frequency corresponding image forming apparatus for forming a band frequency corresponding image between bands of the current and previous frames; And a signal synthesizing apparatus for synthesizing a speech signal from the decoded pitch frequency, spectral magnitude, and voiced / unvoiced determination using the band frequency corresponding pseudo image. The method of claim 12, The coder further includes a frame classifier that classifies and assigns the frame classification to each frame in the time domain by the characteristics and ranges of signal values that change along the frame and by the signal oscillation characteristics of the first and second portions of the frame. And the voiced / voiceless determination device generates the voiced / voiceless decision group based on the assigned frame classification. The method of claim 13, The voiced / unvoiced determination device is characterized in that to use an adaptive threshold (adaptive threshold) according to the assigned frame classification. The method of claim 13, wherein the pitch determination device A pitch candidate group determining apparatus for determining a pitch candidate group based on analysis of a normalized autocorrelation function using direct or inverse order according to the assigned frame classification; An optimum candidate selection device for selecting an optimal candidate by measuring a pitch candidate group in a frequency domain; And And an optimum candidate improvement device for improving an optimal candidate value in a frequency domain. The method of claim 15, The optimum candidate selection device measures the pitch candidate group by a window function response scaled to obtain an appropriate sharpness of the approximation function in each band and to provide true pitch candidate selection. System. The method of claim 16, And the window function response is _scaled for a pitch frequency lower than a predetermined frequency F _scale . The method of claim 17, The window function response is scaled by a proportional sharpening procedure. The method of claim 18, The proportional sharpening procedure is performed by linear interpolation. The method of claim 19, The window function response scaled for different pitch frequencies is used as a look-up table. The method of claim 12, The parameter encoding device A scalar quantification device for quantifying a pitch frequency value; A spectral size wavelet reduction device for reducing the dimension of the spectral magnitude vector; A spectral size hybrid encoding device for encoding the reduced spectral magnitude vector by wavelet technology; And And a multiplexing device for combining the encoded data into a plurality of bits. The method of claim 21, The spectral size hybrid encoding device A wavelet encoding device for encoding the reduced spectral magnitude vector; An inter-frame prediction encoding apparatus for encoding the reduced spectral magnitude vector; And And a comparator for comparing the efficiency of the wavelet encoding apparatus with the efficiency of the inter-frame prediction encoding apparatus, and outputting data and decision bits corresponding to the superior apparatus to the multiplexing apparatus. The method of claim 12, The signal sum growth value is A voice synthesizer for synthesizing voiced sound components and racing the pairs of harmonics using pairs of frequency bands having the most similar frequencies in the current and previous frames, wherein all unpaired pairs in the previous frame are used. The harmonics decrease flat to zero amplitude and all unpaired harmonics in the current frame increase flat to become a predetermined amplitude; An unsynthesizing device for synthesizing a speech component for an unvoiced frequency band; And And an adder for synthesizing a model speech signal by summing synthesized speech components for the voiced and unvoiced frequency bands. The method of claim 12, The spectral size determination device A band frequency corresponding image forming apparatus for forming a frequency correspondence between bands of the current and previous frame; A voice synthesizer for synthesizing a model voice signal for a spectral magnitude model group based on the configured band frequency correspondence, pitch frequency, and voice decision group for previous and current frames; A first window device for processing a model meteor signal; An orthogonal conversion device for converting the model meteor signal windowed by the first window device into a frequency domain; A planetary size evaluation device for evaluating the planetary magnitude of the transformed model signal by the least square method; A synchronous noise generator for generating a model white noise signal having a unit amplitude range; A second window device for processing a model white noise signal; An orthogonal converter converting the model white noise signal windowed by the second window device into a frequency domain; And And an unvoiced magnitude estimating device for estimating unvoiced magnitude of the transformed model white noise signal by a least square method. The method of claim 24, Wherein the planetary growth value is assigned a fixed etalon value of 1 in the meteor band and 0 in other cases to form a model signal for a group of spectral size models. The method of claim 12, The voiced / unvoiced discriminator generates a voiced / unvoiced decision group using a window function response scaled to obtain an appropriate sharpness of each band approximation function and to provide true voiced / unvoiced crystal generation. system. The method of claim 26, And the window function response is _scaled for a pitch frequency lower than a predetermined frequency F _scale . The method of claim 27, And the window function response is scaled by a proportional sharpening procedure. The method of claim 28, And said proportional sharpening procedure is performed by linear interpolation. The method of claim 29, The window function response scaled for different pitch frequencies is used as a lookup table. The method of claim 30, And the voiced / unvoiced device tunes the position of the scaled response to the position of a frequency band peak. A method of coding and decoding a speech signal, the method comprising: The voice signal coding method is (a) dividing an input digital voice signal into a plurality of frames to be analyzed in a time and frequency domain; (b) orthogonally transforming each frame to provide spectral data on the frequency axis; (c) determining a pitch frequency for each frame; (d) generating a voiced / unvoiced crystal group using the determined pitch frequency; (e) measuring the spectral magnitude using an analytical method by synthesis; (f) encoding the determined pitch frequency, measured spectral magnitude, and the generated voiced / unvoiced decision for each of the plurality of frames; (g) combining the encoded pitch frequency, the measured spectral magnitude, and the voiced / unvoiced decision into a plurality of bits, The voice signal decoding method (aa) decoding the plurality of bits to provide pitch frequency, spectral magnitude, and voiced / unvoiced determination for each of the plurality of frames; (bb) configuring a band frequency correspondence between the bands of the current and previous frame; (cc) synthesizing a speech signal from the decoded pitch frequency, spectral magnitude, and voiced / unvoiced determination using the band frequency response. The method of claim 32, The coding method Classifying and assigning the frame classification to each frame of the time domain by characteristics and ranges of signal values varying along the frame and by signal oscillation characteristics of the first and second portions of the frame; And And in step (d), using the adaptive threshold value based on the assigned frame classification. The method of claim 33, Step (c) is (i) determining a pitch candidate group based on an analysis of normalized autocorrelation functions using direct or reverse order according to the assigned frame classification; (ii) measuring a pitch candidate group in the frequency domain to select an optimal candidate; And (iii) improving the optimal candidate value in the frequency domain. The method of claim 34, In step (ii), The pitch candidate group is measured by a window function response scaled to obtain an appropriate sharpness of each band approximation function and to provide true pitch candidate selection. The method of claim 35, wherein And the window function response is _scaled for a pitch frequency lower than a predetermined frequency F _scale . The method of claim 36, The window function response is scaled by a proportional sharpening procedure. The method of claim 37, wherein The proportional sharpening procedure is performed by linear interpolation. The method of claim 38, And said window function response scaled for different pitch frequencies is used as a lookup table. The method of claim 32, The step (cc) is Synthesizing voiced sound components and racing the pairs of harmonics using pairs of frequency pairs having the most similar frequencies to each other in the current and previous frames, wherein all non-paired harmonics in the previous frame The flattening decreases to zero amplitude and all unpaired harmonics in the current frame increase flattening to a predetermined amplitude; Synthesizing a speech component for the unvoiced frequency band; And Synthesizing a model speech signal by combining the synthesized speech components in the voiced and unvoiced frequency bands. The method of claim 32, In step (d), a voiced / unvoiced decision group is generated using a window function response scaled to obtain an appropriate sharpness of each band approximation function and to provide true voiced / unvoiced decision generation. . 42. The method of claim 41 wherein And the window function response is _scaled for a pitch frequency lower than a predetermined frequency F _scale . The method of claim 42, wherein The window function response is scaled by a proportional sharpening procedure. The method of claim 43, The proportional sharpening procedure is performed by linear interpolation. The method of claim 44, And said window function response scaled for different pitch frequencies is used as a lookup table. 42. The method of claim 41 wherein And said step (d) further comprises tuning the position of said scaled response to the position of a frequency band peak.

Images