JPH0981191A - Voice coding/decoding device and voice decoding device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To improve the pitch predicting precision by shortening the renewal period of pitch length compared with the renewal period of pitch predicting factor. SOLUTION: Frequency dividers 102, 107 are newly added to give the renewal period of pitch length shorter than that of pitch predicting factor. Though the instructions on analysis processing of pitch predicting factor are given to a pitch prediction analysing filter and a pitch prediction synthesizing filter with the same period as the past one, the instructions on the analysis processing of pitch length are given with a frequency faster than that in the analysis processing of the pitch predicting factor by frequency dividers 102, 107. By using these means, the renewal period of pitch length is shortened compared with the renewal period of the pitch predicting factor. Accordingly, the pitch predicting precision is improved in the voice coding/decoding device which conducts pitch prediction to plural sample length units.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a device for compressing and transmitting or storing voice waveform information.

[0002]

2. Description of the Related Art FIG. 4 shows a speech coding / decoding apparatus including pitch prediction processing of the prior art.

The encoder has a speech input terminal 400, a linear prediction analysis of the inputted speech, a linear prediction coefficient, a linear prediction analysis filter 403 for outputting a linear prediction residual signal, and a linear prediction residual. A signal is input, pitch prediction analysis is performed, a pitch length and a pitch prediction coefficient are encoded, a pitch prediction analysis filter 404 that outputs a pitch prediction residual signal, and a prediction residual that inputs and quantizes a pitch prediction residual signal It is composed of a difference quantizer 405, a linear prediction analysis filter, a pitch prediction analysis filter, and a controller 401 for controlling the operation of the prediction residual quantizer.

The decoder has an operation procedure opposite to that of the encoder, and a prediction residual inverse quantizer that generates a pitch prediction residual signal based on the coding information transmitted from the encoder. 408, a pitch prediction synthesis filter 409 that generates a signal having a pitch structure of speech, and a linear prediction synthesis filter 41.
0, an output terminal 411 of the synthesized speech signal, a linear prediction synthesis filter, a pitch prediction synthesis filter, and a controller 406 for controlling the operations of the prediction residual dequantizer.

In the above speech coding / decoding apparatus, the controller gives an instruction to each processing block to execute processing for each of a plurality of sample units. Both the pitch prediction analysis filter and the pitch prediction synthesis filter operate to calculate and process the pitch length and the pitch prediction coefficient at the synchronized timing according to the instruction.

Next, a specific example of the conventional pitch prediction processing will be described.

As a method for efficiently compressing and transmitting or accumulating a speech waveform, code-excited linear prediction coding (for example, "Code-Excited Linear Prediction (CELP): High") is used.
Quality Speech at Very Low Bit Rates ", MR Schro
eder and BS Atal, Proc. IEEE Int. Conf. on Acoust
ics, Speech and Signal Processing, pp. 937-940, 19
85) A block diagram of a known CELP encoder is shown in FIG. This encoder has a voice input terminal 50.
1, an auditory weighting filter 502 for auditorily masking an error signal by encoding, an adaptive codebook 503 for long-term prediction (pitch prediction), an excitation codebook 505 storing a plurality of excitation waveforms, Adaptive codebook 50
3 and the amplifiers 504 and 506 for amplifying the signals generated from the excitation codebook 505 with gains of g _a and g _s , respectively.
An adder 507, a synthesis filter 508 in which a speech linear prediction synthesis filter and an auditory weighting filter are connected in cascade, a subtractor 509, and an error energy minimization determiner 5 that determines the energy of an error signal generated by the subtractor 509.
It consists of 10.

In the CELP coding process, for example, a process such as a linear prediction analysis is performed for every 160 samples (which is called a frame), and is adapted for every 40 samples (which is called a subframe) obtained by dividing a frame into four. In general, the search processing of the codebook 503 and the excitation codebook 505 is performed.

The CELP codebook (adaptive codebook and excitation codebook) is selected by the error energy E as shown in equation (1).
_The purpose is to select the index _k that minimizes k from the codebook. The case of the adaptive codebook will be briefly described below.

[0010]

[Equation 1]

Here, X is a perceptual weighting filter 502.
The column vector of the input signal processed by the above equation (2). H is a lower triangular matrix having the impulse response of the synthesizing filter 508 as an element, and is represented by equation (3). P _k is a column vector generated at the k-th index of adaptive codebook 503 and is represented by equation (4). The above prediction using the adaptive codebook 503 may be called pitch prediction. g is a scalar gain (pitch prediction coefficient) given to the amplifier 504.

[0012]

[Equation 2]

[0013]

(Equation 3)

[0014]

(Equation 4)

Here, N represents the vector length (subframe length) of the excitation signal, and ^T represents the transposition operation.

To select the optimum pitch length for minimizing E _k , the index k (pitch length) that maximizes S _k shown in equation (5) is selected from the adaptive codebook 503. Will be done.

[0017]

(Equation 5)

The process of generating a signal from the adaptive codebook is shown in FIG.
Shown in The adaptive codebook 601 is specifically composed of a memory, and extracts a vector of only the subframe length (N) from the pitch length (k) position. Here, the extracted signal is stored in the pitch signal storage 602.

The pitch length (k) is the subframe length (N)
Various methods have been proposed to improve the pitch prediction accuracy by processing for shorter times and oversampling the signal of the adaptive codebook (for example, "Analysis and Improv").
ement of the Vector Quantization in SELP ", WBKle
ijn, Signal Processing IV, pp. 1043-1046, 1988, Mo P
itch Predictors with High Temporal Resolution ", P.
Kroon, ICASSP, pp. 661-664, 1990).

The parameters obtained by the adaptive codebook selection are
The optimum pitch length and (k _{o p t),} pitch prediction coefficient (g
_a ) (Pitch prediction coefficient selection is not described here because there are various known methods).

For the purpose of suppressing the increase in the amount of transmitted information and improving the accuracy of pitch prediction, it has also been proposed to shorten the pitch prediction coefficient update cycle with respect to the pitch length update cycle (for example, , Taniguchi et al., Japanese Patent Laid-Open No. 3-33898 "Pitch Prediction Method".

Further, the signal reproduced by the encoding / decoding has a smoother spectrum and pitch structure than the original signal. A device that emphasizes these in the decoding process is generally called a post filter. The process of emphasizing the pitch structure is basically the same as the pitch prediction technique, and generally, the information such as the pitch length and the pitch prediction coefficient coded by the coding process is used to perform the filtering process in subframe units ( For example, "Pitch Synchronous Innovat
ion CELP (PSI-CELP) -PDC half rate audio CODEC- ",
Oya, Suda, Miki, IEICE RCS93-78, pp. 63-70, 199
3).

In order to further increase the compression rate, a method of increasing the processing unit such as the frame length and the subframe length is widely used. For example, in PSI-CELP, which is the Japanese half-rate digital cellular standard system, the subframe length is in units of 80 samples, and the adaptive codebook search process sets the pitch length and pitch prediction coefficient for a signal of 80 samples.

[0024]

However, when the processing unit becomes long, there is a problem that the pitch length changes in the analysis range, especially at high pitch frequencies of women. For example, the sampling frequency is 8 kHz and the pitch frequency is 40
At 0 Hz, the processing unit of 80 samples includes 4 pitches, and it is quite possible that the pitch length changes between these 4 pitches. A pitch length error due to a change in pitch length within the subframe has a problem of causing a great deterioration in pitch prediction performance. Further, the method of updating the pitch prediction coefficient as shown in the related art has a problem that the prediction performance is insufficient.

[0025]

According to the present invention, a speech input terminal, a speech signal linear prediction analyzer, a pitch prediction analyzer for predicting a speech pitch signal, and a prediction for quantizing a prediction residual signal. A residual quantizer, a controller for operating the linear prediction analyzer, the pitch prediction analyzer, and the prediction residual quantizer in units of a plurality of sample lengths, a speech output terminal, and a linear prediction synthesizer, Pitch prediction synthesizer, prediction residual dequantizer, controller for operating the linear prediction synthesizer, pitch prediction synthesizer, and prediction residual dequantizer in multiple sample length units, and pitch length It comprises means for shortening the update cycle to be shorter than the pitch prediction coefficient update cycle.

[0026]

1 is a block diagram showing the configuration of a voice encoding / decoding device according to the present invention. The description of the parts common to FIG. 4 described in the related art will be omitted. The difference is that frequency dividers 102 and 107, which are means for shortening the pitch length update cycle from the pitch prediction coefficient update cycle, are newly added.

For the pitch prediction analysis filter and the pitch prediction synthesis filter, the pitch prediction coefficient analysis processing instruction is given in the same cycle as the conventional one, but the pitch length analysis processing instruction is given by the frequency divider. Instructions are given in a cycle earlier than analysis processing instructions. By doing so, the pitch length update cycle is shorter than the pitch prediction coefficient update cycle for processing.

Next, the process of generating the pitch prediction signal of the present invention will be described with reference to FIG. 2, corresponding to FIG. 6 described in the prior art. For convenience of description, the pitch length update unit will be referred to as a “sub-subframe”. FIG. 2 shows an example of the subframe length N and the subsubframe length N / 2. In this case, divide the sub-frame equally into the first half / second half,
The pitch length can be changed.

Reference numeral 201 denotes an adaptive codebook, and reference numeral 202 denotes a pitch signal storage for storing a signal generated from the adaptive codebook. Further, k1 and k2 indicate the pitch length of the first half and the pitch length of the second half, respectively.

FIG. 2 shows an example in which the process of extracting the signal of the processing unit length (N) from the adaptive codebook 201 is executed twice.

The procedure for determining the parameters for pitch prediction is the same as the procedure of the equations (1) to (5) described in the prior art. The difference is that the pitch prediction signal generation process represented by the equation (4) is changed as shown by the following equation (6).

The signal generated by the means of the present invention is represented by equation (6).

[0033]

(Equation 6)

By doing so, it is possible to cope with the change in the pitch length in the subframe. There are various possible processes for searching for the optimum pitch length (optimal k1, k2 in the above example) from the adaptive codebook. As a method with a large amount of calculation but the highest prediction accuracy, the pitch length (k
There is a process of generating all combinations of 1, k2) based on the equation (6). To determine the pitch length, the indexes k1 and k2 that maximize S _{k 1} and _{k 2} shown in the equation (7) as in the equation (5) are selected from the adaptive codebook.

[0035]

(Equation 7)

In the case of the equation (7), the amount of information required for pitch length transmission is K for each type of pitch length (k1, k2).
Then, it is 2 * log ₂ (K) bit. However, in the above method, the amount of pitch length transmission information becomes very large.

Next, an embodiment according to claim 2 will be described.

Generally, the pitch frequency of speech changes gradually, and the change amount of pitch length within a subframe is often limited to a very small range. By utilizing this property, the pitch update range for each sub-subframe can be limited. For example, when the subframe is divided into two, the following examples can be considered.

[0039]

(Equation 8)

In this case, in the sub-subframe in the latter half of the subframe, the pitch length can be changed by ± 1 sample length at most from the pitch length in the first half. In the case of the number (8), the amount of information required to transmit the pitch length is log _{2 where} K is the type of pitch length (k1).
(3K) bits, and the amount of information required for transmission can be reduced.

As another example of pitch length search processing, a method of determining the pitch length of the sub-subframe after selecting an average pitch length in subframe units can be considered.
When the sub-frame is divided into two, assuming that the average pitch length is km, the range of values that can be taken by k1 and k2 is, for example, an example such as Expression (9).

[0042]

[Equation 9]

In the example of the expression (9), the information of the average pitch length km is used to limit the movement of the pitch length within the subframe to 7 patterns. In the case of equation (9), the amount of information required for pitch length transmission is log ₂ (7K) bits, where K is the type of pitch length (km). A schematic representation of this pitch length movement is shown in FIG.

In FIG. 3, the change pattern of the pitch length in two sub-subframes is shown by seven arrows.

As in the above example, even if the change in pitch length within a subframe is limited, most cases of the actual change in pitch length of voice can be expressed, so that the increase in the amount of transmitted information can be suppressed to a small level. The pitch prediction accuracy can be improved.

Various limitations can be considered for the limitation of the update width of the pitch length such as applying the oversampling technique described in the prior art.

Next, a method of reducing the search processing amount of the pitch length changing within the subframe according to claim 3 will be described in the case where the change of the pitch length within the subframe is limited.

In the search, the calculation of the equation (7) is performed. In this calculation, for example, the change of the pitch length is calculated by the equation (9).
It limits to 7 patterns like.

First, consider the numerator term of equation (7).
In this section, calculating a portion of the X ^T H above, (X
^{This is} an inner product operation of ^T H) and P _{k 1} , _{k 2} . This is a process commonly called "Backward filtering" (eg, "Fast CESLP coding based on algebraic codes",
JP.Adoul, etc., Proc. IEEE Int. Conf. On Acoustic
s, Speech and Signal Processing, pp. 1957-1960, 19
87). When the length of the subframe is N, the amount of calculation of the numerator term is about 7N. However, when the sub-frame is divided into two, this inner product calculation can also be calculated by dividing into two, and by adding the inner product values of the first half / latter half according to the combination of movements of the pitch length. Therefore, the calculation amount of the inner product value is about 3N (= 6 * N / 2), which can be reduced to about half.

Next, consider the denominator term of the equation (7). This term represents the energy obtained by filtering P _{k 1} and _{k 2. If} the filter order is P (generally about 10), the denominator calculation amount is about 7 NP. However, when the subframe is divided into two, the amount of calculation of this energy can be reduced as follows.

First, the values are set only for the signals of the first half sub-subframes, and all of the latter half sub-subframes are set to 0.
The signal obtained by filtering the signal of is represented by the number (10).

[0052]

(Equation 10)

Here, the symbol @ represents a connection of vectors, the first half N / 2 length vector is P0 _{k 1} , and the second half N / 2 length vector is R _{k 1} . Similarly, a signal obtained by filtering only the signal of the latter sub-subframe is represented by the equation (11).

[0054]

[Equation 11]

Here, P1 _{k 2} is a vector of length N / 2.

The energy obtained by filtering P _{k 1} and _{k 2} is calculated by dividing it into the first half / second half of the subframe. The first half is P0 _{k 1} (k1 = km + 1, km, km
Only the calculation of -1) should be considered, and the processing amount is about 3NP / 2.
It is. In the second half, it is necessary to calculate the energy of the vector obtained by adding and combining R _{k 1} and P 1 _{k 2} . This energy is calculated by calculating the energy of each vector and calculating the inner product of two vectors, as shown in equation (12).

[0057]

(Equation 12)

Since the vector R _{k 1} corresponds to a zero input response (filtering with an input signal of 0), the calculation of energy is terminated with a length of M (eg, 10) shorter than N / 2 sample length. be able to. From the above, the amount of processing required to calculate the energy in the latter half is about 3M.
P + 3NP / 2 + 7N / 2 (7 in this formula is
It is the number of combinations of k1 and k2).

When the first half and the second half are combined, the processing amount is about 3P.
(N + M) + 7N / 2, which can be reduced compared to the processing amount (7NP) in which the subframe is not divided.

Next, an embodiment according to claim 4 will be described.

The present method can also be applied to a pitch emphasis filter for emphasizing the pitch structure of coded speech. Generally, the pitch enhancement filter is the same as the pitch prediction technique, and the information of the pitch length and the pitch prediction coefficient is updated for each subframe length using the information transmitted from the encoder.

When the pitch prediction filter of the present invention is included in the encoder side, the pitch length information changes within the subframe. Therefore, even in the pitch enhancement filter, the pitch length update period is set from the pitch prediction coefficient. It can be processed quickly.

Even when the pitch prediction filter on the encoder side calculates the pitch length and the pitch prediction coefficient in subframe units as described in the prior art,
If the optimum pitch length analysis / search process is performed again in the decoding process, the pitch length can be changed within the subframe, and the pitch enhancement filter process in which the pitch length update period is faster than the pitch prediction coefficient can be performed.

The pitch predictor is not limited to CELP as a speech coding method, but can be applied to all speech coding / decoding methods including pitch prediction.

[0065]

EFFECTS OF THE INVENTION In a speech coding / decoding apparatus that performs pitch prediction in units of a plurality of sample lengths, the pitch prediction accuracy can be improved by shortening the pitch length update cycle with respect to the pitch prediction coefficient update cycle. . Further, by providing a means for limiting the update amount of the pitch length within a certain range, the amount of information necessary for encoding the pitch length can be suppressed to be small. Also, the amount of processing required for pitch length search can be reduced. Moreover, by applying this method to a pitch enhancement filter, the quality of encoded and decoded speech can be improved.

[Brief description of drawings]

FIG. 1 is a diagram illustrating a speech coding / decoding apparatus including a pitch predictor of the present invention.

FIG. 2 is a diagram illustrating a process of generating a prediction signal from the pitch predictor of the present invention.

FIG. 3 is a diagram schematically illustrating an example of the movement of the pitch length when the change of the pitch length is limited in the pitch predictor of the present invention.

FIG. 4 is a diagram illustrating a speech encoding / decoding device including a conventional pitch predictor.

FIG. 5 is a block diagram of a conventional speech encoding process.

FIG. 6 is a diagram illustrating a process of generating a prediction signal from a conventional pitch predictor.

[Explanation of symbols]

201, 503, 601 Adaptive codebook 202, 602 Pitch signal store 100, 501, 400 Input terminal 502 Auditory weighting filter 504, 506 Amplifier 505 Excitation codebook 507 Adder 508 Synthesis filter 509 Subtractor 510 Error energy minimization determiner 101, 106, 401, 406 Controller 102, 107 Frequency divider 103, 403 Linear prediction analysis filter 104, 404 Pitch prediction analysis filter 105, 405 Prediction residual quantizer 108, 408 Prediction residual dequantizer 109, 409 pitch prediction synthesis filter 110,410 linear prediction synthesis filter

Claims (4)

[Claims] 1. A speech input terminal, a speech signal linear prediction analyzer, a pitch prediction analyzer for predicting a speech pitch signal, a prediction residual quantizer for quantizing a prediction residual signal,
A controller for operating the linear prediction analyzer, the pitch prediction analyzer, and the prediction residual quantizer in units of a plurality of sample lengths, a speech output terminal, a linear prediction synthesizer, a pitch prediction synthesizer, and a prediction In a speech coding / decoding apparatus including a residual dequantizer, a controller for operating the linear prediction synthesizer, the pitch prediction synthesizer, and the prediction residual dequantizer in units of a plurality of sample lengths, the pitch A speech coding / decoding apparatus comprising means for shortening a long update cycle shorter than a pitch prediction coefficient update cycle. 2. The speech coding / decoding apparatus according to claim 1, wherein the pitch length update amount is limited to a certain range. 3. The apparatus according to claim 1, wherein a pitch length search operation is performed on a signal obtained by dividing a plurality of sample lengths,
A speech coding / decoding device, characterized in that a combined operation of divided operation results is performed as a whole of a plurality of sample lengths. 4. A pitch decoding filter for emphasizing a pitch structure of decoded speech, wherein a pitch length updating cycle is shorter than an updating cycle of a pitch prediction coefficient. .