JPH11272299A - Method for embedding watermark bit during voice encoding - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide one method for secretly embedding electronic watermarks by use of some of voice codes provided by ITU-T-recommended G. 729 CS-ACELP which is a method for encoding digital voices. SOLUTION: When a key kp for labeling candidates for pulse positions involved in information having the candidates for the plurality of adjacent pulse positions in voice codes is introduced and watermark bits are embedded by use of a synthesizing method which employs a multi-pulse sound source code book used in coding voices, the fourth pulse position m3 corresponding to the watermark bit to be embedded is selected from the candidates for the pulse positions labeled by the key kp , whereby watermark information at 200 bit/sec. or less can be secretly embedded and transmitted as the pulse positions are controlled by a random bit series generated from a text, without providing acoustic discomfort at all even if the transmitted codes with the watermark information embedded therein is directly reproduced as voices, i.e., without letting outsiders know it.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a method for secretly embedding a digital watermark using a part of a voice code of digital voice.

[0002]

2. Description of the Related Art Conventionally, as the most basic technique for digitizing voice, linear pulse coding (Pulse Code) for quantizing the amplitude of a waveform based on a sampling theorem is known.
eModulation (hereinafter referred to as PCM) is known. A plurality of digital waveform values obtained by the PCM method are grouped into a frame, and one of the techniques for generating a speech code for each frame is code excitation linear prediction (CodeExited Lincar P).
rediction Audio Codes:
CELP) is known. Furthermore, this CE
Regarding LPs, ITU-Recommendation G.LP. 729 8 kbit / s CS-
ACELP (Conjugate Structure)
Algebraic CELP) has been announced,
This recommendation is one of the coding methods based on CELP, and relates to a technology capable of reproducing high-quality sound despite a significant reduction in the code amount.

However, in the transmission of sound using digital codes as in this recommendation, since the complete sound can be easily duplicated at the transmission destination, it is difficult to protect the copyright of the author, creator, or performer. The problem has been pointed out. As a measure for copyright protection, an attempt has been made to embed copyright information (digital watermark) for identifying an illegal copy of digital media using a digital code by using the ambiguity of human perception. (Example: Koshio Matsui: Digital Watermark, Journal of the Institute of Image Electronics Engineers of Japan, Vol. 26, N
o. 3, pp. 266-274, 1997).

For embedding a special signal such as a digital watermark in digital audio, a method of embedding a digital watermark using an audible masking phenomenon has been proposed by Boney et al. (Boney et al .: Digital w
atermarks foraudio signal
s, Proc. of the International
al Conference on Multimed
ia Computing and Systems,
pp. 473-480, 1996) has proposed a method of embedding and transmitting document data by imitating electronic noise. (Koshio Matsui et al .: Embedding Text Information in Speech Code in Adaptive Differential PCM Coding, Journal of Information Processing Society of Japan, Vol. 38, No. 10, p.
p. 2053-2061, 1997)

On the other hand, according to Iwakiri et al., International standards (recommendations) G. For 726, a clever method of embedding a digital watermark has also been proposed. (Munetoshi Iwakiri et al .: Adaptive Differential PC
Embedding text information in speech code in M coding,
IPSJ Transactions, Vol. 38, no. 10. p
p. 2053-2061, 1997)

[0006]

However, according to the method proposed by Boney et al. And Matsui et al., The embedding position may be specified by a third party. It is considered that the watermark information may be lost due to the large code compression performed. Recommendation G. As described above, Iwakiri et al. Have proposed a method of embedding a digital watermark skillfully as described above. 729, there is no proposal for such a skillful digital watermark embedding method. Therefore, an object of the present invention is to provide Recommendation G. Recommendation G.726 by a method different from that shown by Iwakiri et al. 729 is to provide a method for embedding a digital watermark in a compressed speech code and a simple method for hiding the presence of the digital watermark.

[0007]

The basic idea of the present invention for solving the above problems is to focus on the structure of a multi-pulse sound source used when encoding digital audio data, and to determine the bit structure in the synthesis process. This is to embed the serialized data. At this time, by arbitrarily selecting the speech code to be embedded and changing the embedding rule,
The presence of the implant can be hidden.

[0008] The present invention according to claim 1 is a method for embedding a watermark bit at least when digitally encoding voice using a fixed codebook and transmitting the same, wherein candidates for a plurality of adjacent pulse positions in the fixed codebook are provided. "1" or "0"
And a first key for selecting the pulse position candidate based on whether it is “1” or “0” is determined, and a bit position for embedding a watermark in the transmission speech code is determined by the first key. It is characterized by using the selected pulse position, focuses on the structure of a fixed codebook (multi-pulse sound source) used when encoding digital audio data, and embeds bit sequence data in the synthesis process. It is a thing.

According to a second aspect of the present invention, in the first aspect of the present invention, in the fixed codebook, each of the possible values of the predetermined number of the pulse position candidates is set to “1”. Or "0" is assigned, and a second key for selecting whether to perform the embedding of the watermark depending on whether the total value is "1" or "0" is determined, and obtained by feedback from the output speech code. The embedding of the watermark is performed by associating the obtained total value with the second key, and the audio code for embedding the watermark is unspecified, so that the third party can be selected. Makes it difficult to specify a speech code including watermark information, and it is possible to reduce the possibility that a key is analyzed by a third party.

According to a third aspect of the present invention, in the second aspect of the present invention, a third key which is reversely assigned to “1” and “0” in the first key is determined, Detecting whether the total value of a predetermined number of pulse position candidates is an even value or an odd value, and determining the first key so that each of the even value and the odd value of the total value does not have the same key. And one of the third keys is associated with each other, and the first key or the third key is placed in a bit position for embedding a watermark in the transmission voice code.
By using the pulse position selected by the key of (i), and changing the rule of embedding the watermark, the possibility that the key is analyzed even if the same key is used for a long period of time can be reduced.

[0011]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS An embodiment of the present invention will be described below with reference to the drawings and mathematical expressions. In the present embodiment, first, "I. Overview of Recommendation G.729" will be described, and then "II. Embedding method" (a text bit sequence is converted to a speech code using a multipulse sound source according to claim 1). ), Followed by "III. Method of improving confidentiality" (the method of concealing the presence of the embedment of claims 2 and 3 to make key analysis difficult), and "IV. Search algorithm" ( An algorithm of an actual operation for searching for a pulse position for embedding) will be described in the order of “V. Experimental results”.

I. Recommendation G. Outline of ITU 729 The following ITU is used as a reference for understanding the invention of the present application.
-T Recommendation G. 729 will be briefly described. Note that the purpose of the present invention is, as described above, the recommendation G. Recommendation G.729 provides a method for embedding a digital watermark in a speech code compressed using a fixed codebook as in 729, and a simple method for hiding the presence of the digital watermark. 729 is not intended to be described in detail, and therefore the following ITU-T Recommendation G. 729, IT
U.T. Recommendation G. Even though it is a part necessary for the essential configuration and operation of the G.729, a part that is not relevant to the present invention is deleted and described. Therefore, Recommendation G. 72
9 are insufficient for the explanation of Recommendation G.9. Please refer to the text of G.729.

ITU-T Recommendation G. 729, a conjugate structure-algebraic code excitation linear prediction coding scheme (hereinafter, CS-A
The audio codec is performed using the CELP method. One frame is 10 milliseconds long and has 80 sample points therein. The subframe is 5 ms long and has 40 sample points therein. The bit rate is 8
k bits / sec. That is, the audio is sampled, an 80-bit audio code is generated for each frame, and transmitted at a bit rate of 8 kbit / sec.

G. Recommendation using CS-SCELP method 7
In Recommendation G.29, 32 kbit / sec. 726, which realizes the same voice quality as that of the third generation personal communication, "Future Public Land Mobile Communication System (hereinafter, FPLMT)".
S)) and standardized with strict conditions such as securing error resilience, resulting in a high-quality codec with a small delay. Therefore, Recommendation G. 7
29 can be applied not only to FPLMTS but also to voice-compatible multiplexing devices for asynchronous transfer mode (ATM) and frame relay, and has already been implemented in some products.

Recommendation G. The 729 encoder selects (1) the coefficients of a linear prediction synthesis filter (corresponding to the shape of a human mouth) obtained by analyzing and quantizing the input speech, and (2) selecting from the adaptive codebook and the noise codebook which are the driving sound sources. And (3) a gain parameter (corresponding to the volume of voice) for adjusting the volume. Then, the decoder synthesizes speech from those parameters. Inter-frame prediction is applied to the linear prediction synthesis filter and the gain coefficient in order to increase the quantization efficiency. This is a method of predicting the current parameter from the parameter obtained in the previous frame and encoding the difference.

Recommendation G. 729 features,
(1) Algebraic in the noise codebook
(2) speeding up the operation and eliminating the need to accumulate all the noise codebook patterns as in the conventional CELP system;
(3) The memory can be saved, and the amount of calculation for pattern selection can be reduced. (4) The amount of calculation of the A-B-S method (analysis-by-synthesis: analysis by synthesis) for selecting a random codebook is large. If the bit rate is sufficiently allocated, use a method that can synthesize sounds as close as possible to the input speech. (5) Add a parity to the pitch period obtained from the adaptive codebook to detect bit errors during transmission Recommendation G. That is, the transmission error resistance is superior to that of H.726.

FIG. 729 is a block diagram illustrating the configuration of a basic encoder of FIG. The encoder in FIG. 1 calculates a coefficient of a linear prediction filter at a rate of once per frame of 10 milliseconds by a preprocessing unit 2 that performs high-pass filtering and scaling of an input signal 1 and a preprocessing signal. A linear prediction analysis unit 3 that converts the coefficient into a line spectrum pair (LSP), quantizes and outputs a linear prediction coefficient 7, a linear prediction synthesis filter 4 to which the linear prediction coefficient and the like are input, a preprocessing An adder 5 to which signals from the unit 2 and the linear prediction filter 4 are input and to which an error signal is input, an audible weighting filter 6 for filtering the output of the adder 5 using the linear prediction coefficient 7, and an audible weighting filter 6
, And the outputs of the fixed codebook search unit 8 and the pitch analysis unit 9 are input to the transmission bit stream 1 based on the linear prediction coefficient 7.
1; a gain quantization unit 12 to which the outputs of the fixed codebook search unit 8 and the pitch analysis unit 9 are input to output a gain; an adaptive codebook 13, which is a driving excitation;
Fixed codebook 14, which is both a driving excitation and a noise codebook, gain Gp15 of adaptive codebook 13, and gain G of fixed codebook 14
c16, and its operation is described in Recommendation G. 729 is not described in detail here.

Incidentally, Recommendation G. 729 is characterized by a fixed codebook (noise codebook) used for encoding.
The present invention also utilizes the features of the fixed codebook. Recommendation G. FIG. 2 shows a table showing the fixed codebook of H.729. FIG. 2 shows a multi-pulse excitation type fixed codebook, in which four pulse positions “m0” to “4” are selected from the candidates shown in the chart of 40 samples in a subframe of 5 ms.
"M3" and polarity information "s0" to "s3" are determined.

Recommendation G. At 729, the process of searching for pulse position and polarity information from the chart of FIG. 2 is as follows. For example, a 10th-order linear prediction filter corresponding to the linear prediction synthesis filter 4 in FIG. 1) (= G.729: Equation 2).

(Equation 1)

An audible weighting filter corresponding to the audible weighting filter 6 in FIG. 2) (= G.729: Equation 27). Here, γ1 and γ2 are weighting coefficients that are functions of the spectrum shape of the input signal for determining the characteristics of the perceptual weighting filter W (z).

(Equation 2)

Here, the synthesis filter W of the linear prediction filter and the audible weighting filter is used. A 40th-order code vector c derived from the chart of FIG.
(N) is expressed as the following equation (Equation 3) (= G.729: Equation 45).

(Equation 3) However, it is assumed that δ in the mathematical expression (Equation 3) is the following mathematical expression (Equation 4).

(Equation 4)

Here, first, the following equation (Equation 5) (= G.
729: The 40th-order code vector c (n) is processed using the filter P (z) in the equation 46).

(Equation 5) In this equation (Equation 5), T is a pitch delay. Β is the amount of gain g _p ^(m−1) of the adaptive codebook (corresponding to adaptive codebook 13 in FIG. 1) in the previous frame. 47).

(Equation 6) The impulse response h (n) of (z) is expressed by the following equation (Equation 7).
(= G.729: Equation 49).

(Equation 7)

Next, the target signal x '(n) is obtained by the following equation (Equation 8) (= G.729: Equation 50).

(Equation 8) X (n) in the above equation (Equation 8) is calculated from the weighted sound sw (n) to W (z). (N) is a convolution integral value of the impulse response h (n).

Next, x '(n) obtained by the equation (Equation 8)
Using the following equation (Equation 9) (= G.729: Equation 52)
To obtain d (n).

(Equation 9)

The following equation (Equation 10) (= G.72)
9: The value according to equation 54) is converted to a codebook C having an algebraic structure.
And Here, _mi is each pulse position.

(Equation 10) Further, the following equation (Equation 11) (Reference equation: G.729: Equation 5)
The value according to 5) is defined as energy E.

[Equation 11] Here, the matrix φ ′ (i, j) is expressed by the following equation (Equation 12).
(= G.729: Equation 56) (Equation 13) (= G.729:
Equation 57).

(Equation 12)

(Equation 13) The matrix φ (i, j) is expressed by the following equation (Equation 14) (=
G. FIG. 729: It is obtained by Expression 51).

[Equation 14]

The search for the codebook in the chart of FIG. 7
Abs (synthesis analysis: Analysis b) so as to maximize the value of C ² / E as shown in equation 53 of 29.
ySynthesis). Therefore,
Since the processing amount is generally enormous, a threshold value is introduced as described below to reduce the search processing amount.

[0027] Using the coefficient K ₃ = 0.4, the following equation (Equation 1)
5) Calculate the threshold thr ₃ by (= G.729: Equation 60).

(Equation 15)

[0028] In this case, the fourth of the search of the pulse _{m 3} is, t
This is performed only for a combination with pulse positions m ₀ , m ₁ , and m ₂ exceeding hr ₃ . In addition, an increase in processing delay is suppressed by using the maximum search processing amount TI _max for each frame. On the other hand, the transmitting side extracts each parameter from the transmitted code and synthesizes the output speech, so that the processing amount is smaller and faster than the encoding.

II. Embedding method Hereinafter, the present invention is described in Recommendation G. 729, the principle of a method of embedding watermark information in a speech code by applying the CS-ACELP coding method.

The fourth pulse information i in the chart of FIG.
Pulse position _{m 3} of _3, other _i 0 through i ₂ pulse positions m
₀ ~m Unlike _second candidate seen to have adjacent candidate. Here, as shown in FIG. 3 is a diagram showing a fourth substitution pulse position m ₃ pulse information, even voice code the optimum pulse position m _3, which is selected by replacing the candidate m _'3 adjacent thereto It is considered that the influence on the reproduced sound is small. By utilizing this, a special signal sequence such as watermark information is embedded in the multi-pulse sound source information portion of the speech code.

Firstly, introducing the key _{k p} performing labeling candidate pulse position _{m 3.} For example, as shown in FIG. 4 is a diagram showing the classification of pulse position _{m 3} of the fourth pulse information by the key _{k p,} if the key _k p = "00001111", the most significant bit key _{k p} since "0" is, assigned a "0" to a candidate of a pulse position _{m 3} {3}, assigning a "1" to the candidate {4} adjacent thereto. On the other hand, the key k
least significant bit of _p assigns a "1" to "1" is because the pulse position _{m 3} candidate {38}, assign "0" to a candidate {39} adjacent thereto. In this manner, the entire candidate pulse position m ₃ label of "0" and "1".

[0032] Here, when embedding a watermark bit "0" to the audio code, the key k _p, selects the pulse position m ₃ from candidates, labeled "0" in FIG.
On the other hand, when embedding the watermark bit “1”, the pulse position m is selected from the candidates labeled “1” in FIG.
Select ₃ . By repeating this, the binary watermark information can be embedded.

Further, the signer to know the key k _p, that the label is pulse position m ₃ included in the voice code checks "1" or "0", can be easily extracted watermark information.

III. By the method described in the confidentiality improving methods of "II. Embedding method", when subjected to embedding in all the sub-frame, selecting the position of the pulse position m ₃ of 800 bits every fourth pulse information per second By doing so, the watermark information is embedded, so that 200-bit watermark information can be embedded. However, when embedding watermark in all code using the same key k _p, more likely to third party unauthorized to analyze the key k _p. Therefore, the confidentiality was improved by the following method.

[0035] First, the total value of the candidate pulse position _m 0 ~m ₃ as shown in the following equation (Equation 16) and _{c p.}

(Equation 16) The value of the sum c _p will either 58 types shown in FIG. 5 the sum c _p is a table of possible values.

The pulse position m ₀ included in the speech code
If the random values of the candidates of ~m ₃ is, from the table of FIG. 2, the frequency of occurrence of the sum c _p is considered as the sum of the numbers extracted at random from almost successive natural numbers, therefore, close to a normal distribution It is considered to show characteristics.

In order to obtain an example of a normal distribution if the values of the above-mentioned candidates for the pulse positions m _{0 to} m ₃ are random, the inventor must use the English female voice Ews (see the table of FIG. 9 described later). Pulse position m included in the speech code obtained from
We investigated the occurrence frequency of the sum c _p with the ₀ ~m ₃ candidate, to obtain a survey results shown in FIG.

[0038] FIG. 6 is a diagram showing the frequency of occurrence of the four sum c _p of the pulse position m ₀ ~m _3, 6, occurrence frequency of the sum c _p is shows a normal distribution with schematic Therefore, the candidates for the pulse positions m _{0 to} m ₃ are almost randomly selected. Therefore, distributed arrangement of watermark bits is performed by the feedback processing structure shown in FIG.

FIG. 7A is a diagram showing feedback control in the transmitting device, and FIG. 7B is a diagram showing feedforward control in the receiving device using the feedback processing result of the transmitting device. FIG. 7 (a)
In, the input speech 21 is input to the G729 encoding device 22 and the output code 23 (m ₀ , 8 kbit / s)
m ₁ , m ₂ , and m ₃ ). A part of the output code 23 is fed back to the control device 24, and the control device 24 outputs the embedded code selection signal 25 to the G729 encoding device 22. In the G729 encoding device 22, the embedded code selection signal 2
5, the input speech 21 is encoded and output as an output code 23.

In FIG. 7B, the output code 23 is G
729 decoding device 26 and the control device 27
Is also input to the G729 decoding device 26 from the control device 27.
, An embedded code search signal 28 is output. G729
The decoding device 26 searches for an embedded code, extracts the embedded code, and outputs an output speech 29.

[0041] Here, the key k assigned to "0" and "1" to the total value of the 58 types that can take the sum _{c p} shown in FIG. 5
Introduce _con . This key k _con is a 58 bit 2
It is a decimal number.

[0042] Figure 8, the key corresponding to the sum c _p of the output speech code obtained by feedback using the control system of FIG. 7 k
It is a figure which shows the extraction process of the bit value _Cpb of _con . In Figure 8, first, we obtain the sum c _p from the output speech code obtained by feedback. Next, extracts the bit values _{c pb} key _{k con} corresponding to the total value _{c p.} When the extracted bit value _cpb is “1”, embedding a watermark bit in the audio code is performed. On the other hand, when the bit value _cpb is “0”, the watermark bit is not embedded in the audio code. By repeating this, the watermark bits can be distributed and arranged over the entire speech code. In this way, the key k _con
It is difficult for a third party who does not know the audio code to specify the audio code including the watermark information.

However, for example, the key k as described above
be introduced _con special to key _{k p} of FIG. 4 key _k p ( "0
When subjected to embedding for a long period of time using 0000000 "or" 1111111 "), the statistical characteristics of the watermark data is embedded, it is considered to be reflected in the speech code. Therefore, long-term use of the same key _{k p} is the watermark Hope in hiding the existence of If Hodokose the Inclusive, consider spreading the deviation of frequency of appearance of the sum c _p.

[0044] Here, the number shown in the table of FIG. 5, since it is composed of 2 number of successive sets all, sum c _p is an even number at a probability of 1/2. Therefore, the total value c _p is I to even

[Equation 17] Can spread.

IV. Search Algorithm FIG. 9 is an operation flowchart of an encoding procedure for a search algorithm for each pulse position using the key described above.
Will be described in detail with reference to FIG. 10 which is an operation flowchart of the decoding procedure.

In the following description of the search algorithm,
The notation is as follows. thr ₃ : search processing threshold value, S _max : maximum value of C / E,
TI _max : maximum search processing amount, time: search processing amount,
L0, ..., L4: _{loop, i 0, ..., i 4} : pulse position _{candidate, m 0, ..., m 3} : optimum pulse position, CE (x
_{0, x 1, x 2,} x 3): pulse position _{x 0, x 1,}
a function for calculating C / E using x ₂ and x ₃ , tbit: bit value to be embedded, mode: embedding execution flag (when embedding: 1, when not embedded: 0, initial value: 0), tap
(X, y): function to extract a bit value of y most significant bit of _{x, check (k con, c} p): Key _{k con}
Function of extracting a bit value _{c pb} corresponding to the sum value _{c p} of, get (T): function to extract one bit from the embedded data file T, put (tbit, T) : The extracted bit value tbit data file T Function to output to.

The processing shown in the operation flowchart of the encoding procedure of FIG. 9 is as follows. In step S1, to calculate the search processing threshold thr _3. In step S2, C
The maximum value S _max of / E is set to “0”. In step S3, the search processing amount time is set to “0”. Step S4
Then, the loop processing L0 is started using i ₀ = 0, 5,..., 35 as pulse position candidates. In step S5,
The loop processing L1 is started using i ₁ = 1, 6,..., 36 as pulse position candidates. In step S6, loop processing L2 is started using i ₂ = 2, 7,..., 37 as pulse position candidates.

[0048] If a hr ₃ (step S7: YES), the process proceeds to step S8, move on. In step S8, i ₃ =
, 38, the loop processing L3 is started.
In step S9, the state of the embedding execution flag “mod” indicating that there is no embedding or that it is the initial value
e is “0” or the mode of the embedding execution flag indicating that the flag is embedded is “1”.
And tap (k _p , a function that extracts the bit value of the (i ₃ −3) / 5th bit from the higher order of the key k _p
It is determined whether or not (i ₃ −3) / 5) = tbit (bit value to be embedded). mode is “0” or mode is “1” and tap (k _p , (i
₃ -3) / 5) = if a tbit (step S9:
YES), the process proceeds to step S10, and the mode is “0” or the mode is “1” and
_{tap (k p, (i 3} -3) / 5) if = not tbit: (step S9 NO), the process proceeds to step S14.

In step S10, the value of C / E, S
Is calculated using the pulse position candidates i ₀ , i ₁ , i ₂ , i _3.
CE (i ₀ , i ₁ , i ₂ ,
i ₃ ). In step S11, it is determined whether S, which is the value of C / E, is greater than S _max, which is the maximum value of C / E. If S is greater than S _max (step S
11: YES), the process proceeds to step S12, where S is
If it is not larger than _max (step S11: NO), the process proceeds to step S14. In step S12, C /
Let _{Smax be} the maximum value of E be S. Step S1
In _No. 3, the optimum pulse position m _{0 is set} to i ₀ , the optimum pulse position m _{1 is set} to i ₁ , the optimum pulse position m _{2 is set} to i ₂ , and the optimum pulse position m _{3 is set} to i ₃ . In step S14,
The loop processing L3 ends. In step S15, loop processing L4 is started using i ₄ = 4, 9,..., 39 as pulse position candidates.

In step S16, the embedding execution flag indicating that no embedding is performed, or the mode of the embedding execution flag indicating the initial value is "0", or the embedding execution indicating that the flag is embedded is performed. flag is a state mode is "1", and, from the top of the key _{k p} (i
_4-4) / 5 is a function of extracting a bit value of bit _{tap (k p, (i 4} -4) it is determined whether / 5) ≠ tbit (bit value embedded). mode is “0”, or mode is “1”, and
If tap (k _p , (i ₄ −4) / 5) ≠ tbit (step S16: YES), the process proceeds to step S17, where mode is “0” or mode is “1”. , And tap (k _p , (i ₄ -4) / 5)}
If it is not tbit (step S16: NO), the process proceeds to step S21. At step S17, CE (i ₀ , i ₁ , i), which is a function for obtaining C / E using the pulse position candidates i ₀ , i ₁ , i ₂ , and i _3, is used as the value of C / E.
₂ , i ₃ ). In step S18, it is determined whether or not S, which is the value of C / E, is greater than _Smax, which is the maximum value of C / E. If S is larger than _Smax (step S18: YES), the process proceeds to step S19, and if S is not larger than _Smax (step S18: NO).
Proceeds to step S21.

In step S19, S is set to _Smax , which is the maximum value of C / E. In step S20, the optimum pulse position _{m 0} and _{i 0,} the optimum pulse position _{m 1} a _{i 1,} and the optimum pulse position _{m 2} and _{i 2,} the optimum pulse position m
₃ and _{i 3.} In step S21, the loop processing L4
To end. In step S22, the search processing amount time
Is set to time + 1. In step S23, the search processing amount time to determine maximum search processing amount _{TI max} is greater than. If time is greater than TI _max (step S23: YES), the process proceeds to step S27, and if time is not greater than TI _max (step S23: NO), the process proceeds to step S24. In step S24, the loop processing L2 ends. Step S
At 25, the loop processing L1 ends. Step S26
Then, the loop processing L0 ends.

In step S27, the total value _{c p,} and the sum _{m 0 + m 1 + m 2} + m 3 of optimum pulse positions. In step S28, the embedding execution flag mode is set to the key k
a function of extracting a bit value _{c pb} corresponding to the sum value _{c p} of _{con check (k con, c p} ) and. In step S29, it is determined whether the embedding execution flag mode is “1”. If the mode is “1” (step S29: YES), the process proceeds to step S30, and if the mode is not “1” (step S29: N)
In O), the process proceeds to step S33.

In step S30, the bit value t to be embedded is
Let bit be get (T), which is a function that extracts one bit at a time from the data file T to be embedded. Step S3
In 1, a determination of whether or not the total value c _p is an even number,
If c _p is an even number: (step S31 YES), the process proceeds to step S32, if _{c p} is not an even number: (step S31 NO), the process proceeds to step S33. In step S33, the optimum pulse positions m ₀ , m ₁ ,
Output m ₂ and m ₃ .

The processing shown in the operation flowchart of the decoding procedure of FIG. 10 is as follows. In step S51,
The embedding execution flag mode is set to “0”. Step S
At 52, the optimum pulse position m ₀ ,
Extract m ₁ , m ₂ , and m ₃ . In step S53, the total value _{c p,} is the sum of the optimum pulse position _m 0 _{+ m} 1 _{+ m}
And ₂ + _{m 3.} In step S54, it is determined whether the embedding execution flag mode is “0”. mode
Is "0" (step S54: YES),
Proceeding to step S60, if the mode is not "0" (step S54: NO), proceeding to step S55.

In step S55, it is determined whether the value obtained by subtracting ₃ from m3 is 0 or a multiple of 5. m ₃ to 3
Is smaller than 0 or a multiple of 5 (step S5
5: YES), the process proceeds to step S56, _{m 3} from 3
Is not a multiple of 0 or 5 (step S5).
5: NO), the process proceeds to step S57. Step S5
In 6, the bit value tbit embedding a function of extracting a bit value of the upper key _{k p (m 3 -3) /} 5 bit _{tap (k p, (m 3} -3) / 5) and . At step S57, the value obtained by subtracting 4 from m ₃ is equal to or a multiple of 0 or 5. If the value obtained by subtracting from m ₃ 4 is a multiple of 0 or 5 (step S57: YES), the
Proceeds to step S58, the whether the value obtained by subtracting 4 from _{m 3} is 0 5
If not (step S55: NO), the process proceeds to step S59.

In step S58, the bit value t to be embedded is
The bit is defined as tap (k _p , (m ₃₎ , which is a function for extracting the bit value of the (m ₃ −4) / 5th bit from the higher order of the key k _p.
-4) / 5). In step S59, the extracted bit value tbit is output to the data file T. Step S
In 60, the embedded execution flag mode, the total value _{c p} is a function of extracting a bit value _{c pb} corresponding to _{check (k} con, _{c p)} of the key _{k con} with. Step S6
In 1, a determination of whether or not the total value c _p is an even number,
If c _p is an even number: (step S61 YES), the process proceeds to step S62, if _{c p} is not an even number: (step S61 NO), the process returns to step S52. In the step S62, the key _{k p} as the key _{k p,} step S
52, and returns to step S52 to step S61 or S
Repeat up to 62.

V. Experimental results V-1. Outline of Experimental System In the embodiment of the present invention in which a digital watermark is embedded as described above, in order for a third party to know the presence of the embedding due to abnormal reproduction sound quality, it is necessary that the embedding does not significantly degrade the sound quality. is important. Therefore, the inventor has made the recommendation G. A simulator according to the G.729 algorithm was created and an experiment was performed.

FIG. 11 is a chart of Japanese and English experimental voices using a male voice and a female voice. The experimental voices shown in the chart of FIG. 11 are 8 kHz of male and female utterances of Japanese and English extracted from FM radio and English conversation tape.
z, 16 bits.

Also, the following keys k _p and k _con are embedded in the embedding.
Was used. The values of these keys are given by the following equation (Equation 18):
As shown in (Equation 19), it is expressed by hexadecimal numbers from 0 to F.

(Equation 18)

[Equation 19]

[0060] Here, the reason that the key k _p as equation (Equation 18), when a special value as in Equation (Equation 18), as described above, the total value c _p obtained from the voice code
This is because it is considered that the statistical characteristic appears largest in the statistical value of.

The reason why the key k _con is represented by the equation (Equation 19) is that when the equation (Equation 19) is used, the watermark is embedded in all the codes, but in terms of deterioration of the reproduced sound quality, Because it is the maximum. That is, in the formula (number 18), handles the worst case of key _{k p,} in the formula (number 19), are dealing with the worst case of key _{k con.} Normally, such a situation must be avoided. However, in order to confirm the effectiveness of the present invention even in the worst case, an experiment was performed as a worst case.

In this experiment, RFC (Request For Co.) of Internet standard was used as watermark information.
The text data contained in the English language texts contained in the texts are used. This is characterized in that each character is represented by an 8-bit ASCII code, and statistically includes many bits “0”.

V-ii. Evaluation method of sound quality Opinion evaluation based on absolute judgment of the evaluator was used as a subjective evaluation method of audio data in which the watermark of this experiment was embedded. This means that a plurality of evaluators absolutely evaluate the sound quality in five levels, and obtain an average opinion value (MOS: M
ean Opinion Score).

In this experiment, the opinion evaluation criteria were set to very good: 5, good: 4, normal: 3, bad: 2, and very bad: 1. Further, in order to avoid the influence of the subject's prejudice, for each sound, two types, one without embedding and one with embedding, were prepared in a state where they could not be distinguished from each other in appearance.

Further, a system which can arbitrarily refer to each voice was prepared and evaluated by freely listening and comparing. With this, if there is a difference in auditory sound quality due to embedding,
It is considered that there is a large difference between the average opinion value MOS of the embedded voice and that of the non-embedded voice.

V-iii. Experimental Results and Discussion In this experiment, since the key k _con of the above equation (Equation 19) was used, the maximum amount of embedding of the watermark in the speech code was 200 bits per second. FIG. 13 is a table showing the average opinion value MOS when the sound quality of the reproduced sound in which the watermark is embedded and the reproduced sound without the watermark are evaluated by eight normal hearing persons in their 20s. According to the table of FIG. 13, the average opinion value MOS in the case where the watermark is embedded and in the case where the watermark is not embedded are about 3.7, respectively. The result is equivalent.

This means that in the above experiment of the present invention, there was almost no difference in audible sound quality due to embedding of the watermark data, that is, it indicates that the subject could not specify the embedded voice. Conceivable. Therefore, as described above, even in an experiment in which the worst condition was assumed using Expression (Equation 18) and Expression (Equation 19), the subject could not identify the voice with embedding. In the embodiment, it is considered that the third party rarely specifies the audio code having the watermark from the reproduced sound due to the audible difference in the reproduced sound quality.

Next, a part of the reproduced waveform was cut out, and the effect of the embedding process on the shape of the waveform was observed. In FIG. 12, FIG. 12 (a) shows a reproduced audio waveform without watermark embedding.
And FIG. 12 of the reproduced audio waveform with the watermark embedded.
(B) and FIG. 12 (c) showing their difference waveforms. These waveforms in FIG. 12 are portions corresponding to the pronunciation "think" in the English male voice data Em in FIG. 11, and are speech sections of 0.2 s. Therefore, the waveform (b) has the watermark information of 40 bits embedded therein.

Since the difference waveform (c) in FIG. 12 is not a straight line, it can be seen that there is a difference between the reproduced voice waveforms in FIGS. 12 (a) and 12 (b). This difference is considered to be due to a change in phase caused by a change in the pulse position of the codebook implemented in the present invention. However, human hearing is generally not good at feeling a phase shift. Therefore, unless an unnatural distortion occurs in the waveform shapes of FIG. 12A and FIG. 12B, even if the phases of both of them slightly change, the sound does not feel unnatural. . Therefore, even if these reproduced sounds are compared, it is considered that there is almost no significant difference in hearing. This can be inferred from the fact that there is no difference between the above-described chart of FIG. 13 with and without watermark embedding.

In a normal case, the speech code released using embedding of the watermark of the present invention in speech should be only the speech code with embedding, and the speech code for comparison as shown in the present experiment was used. Sounds without embedding are not released. Therefore, if the voice code is intercepted by an illegal means, it cannot be compared with a waveform without embedding such as the differential waveform (c) in FIG. Therefore, it is usually impossible for a third party to obtain a differential waveform as shown in FIG. Therefore, it is considered very difficult for a third party to specify the embedded speech code of the present invention from the shape of the reproduced waveform.

Next, considered with the key k _p as in Equation (Equation 18), when large quantities subjected to embedding watermark, statistical bit characteristics of the data to be embedded is reflected in the speech code . Therefore, the total value c _p when all voice codes of the female voice Ews in FIG. 11 are embedded by the same method as in FIG.
Was examined. As a result, key _{k p} a formula (number 1
8) The total value c of the four pulse positions when fixed using
FIG. 14 showing the appearance frequency of _p was obtained.

When FIG. 14 is compared with FIG. 6, it can be clearly seen that the embedding is clearly affected since every other bar graph has irregularities. Thus, using the special key k _p, such as the equation (equation 18) for a long time, statistical characteristics of the embedded bit sequence is reflected in the reproduced sound. That is, the possibility that the presence of the watermark embedded in the third party is noticed increases. In general,
Since the presence of the watermark might be more difficult to notice desired, using the technique of the present invention described in "III. Confidentiality improved method of" above, only the key k _p is varied, so as to diffuse the statistical bias I made it.

[0073] Figure 15 is from the 15 is a diagram showing the frequency of occurrence of the sum c _p of four pulse positions when the key k _p is varied by variation processing key k _p, shown in FIG. 14 It can be seen that such statistical bias has been diffused and improved. Thus, variation processing of the key k _p by the method of the present invention described in "III. Confidentiality improved method of" above is considered to be effective as a method for eliminating the statistical bias caused by implantation.

In the embodiment of the present invention described above, Recommendation G. 729 8 kbit / s, CS-ACEL
A method of secretly embedding watermark information in a speech code of P has been described. In addition, experiments using the method of the present invention have confirmed that even when watermark information or the like according to the present invention is embedded in a speech code, it is possible to reproduce speech without giving a listener of the speech code a sense of incongruity. Was.

By the way, digital voices are generally vulnerable to code errors, and if voice codes with errors are reproduced as they are, the sound quality is greatly degraded. Therefore, even in a speech code using the method of the present invention, if a code error occurs, the influence of the code error is considered to be large.
However, for example, by applying an error correction technique to a speech code using the method of the present invention as a countermeasure,
It is considered that random code errors can be dealt with.

The method of the present invention can also be used when secret information that the sender wants to notify only to a specific party is secretly embedded in a speech code and transmitted. For example, Ste
el other Simultaneous transmitis
Sion of speech and data u
sing code-breaking techni
ques, The Bell System Tech
medical Journal, Vol. 60, no.
9, pp. 2081-2105 (1981), Won
g Other transmitting data on t
he phaseof speech signal
s, The Bell SystemTechnica
l Journal, Vol. 61, no. 10, p
p. 2947-2970 (1982). In the case of this example, there is an advantage that the information can be transmitted to a third party who tries to obtain information by unauthorized means without even knowing the existence of the secret message.

The watermark embedding method of the present invention is very unlikely to be known to a third party as described above. It can be applied to embedding and secret transmission of document data.

[0078]

The watermark bit embedding method of the present invention focuses on the structure of a multi-pulse sound source used when encoding digital audio data, and embeds bit sequence data in the synthesis process. Recommendation G. 726
Of Recommendation G. by a method different from the method indicated by Iwakiri et al. Regarding 729, a method of embedding a digital watermark in a compressed speech code can be provided.

Further, by arbitrarily selecting a voice code to be embedded, it becomes difficult for a third party to specify a voice code including watermark information, and the possibility that the key is analyzed by the third party is reduced. In addition, by changing the rules of embedding, it is possible to reduce the possibility that the key will be analyzed even if the same key is used for a long period of time. Can provide a way to hide.

[Brief description of the drawings]

FIG. 729 is a block diagram illustrating the configuration of a basic encoder of FIG.

FIG. 729 is a chart showing a fixed codebook of G.729.

3 is a diagram showing a fourth substitution pulse position m ₃ of pulse information.

4 is a diagram showing the classification of pulse position m ₃ of the fourth pulse information by the key k _p.

5 is a table showing the four sum c _p possible values of the pulse position.

6 is a diagram showing the frequency of occurrence of the sum c _p of four pulse positions.

FIG. 7A is a diagram illustrating control of a transmission device;
(B) is a figure which shows control of a receiver.

FIG. 8 is a diagram illustrating a process of extracting a key bit value _cpb corresponding to a total value of four pulse positions of a feedback output speech code.

FIG. 9 is an operation flowchart of an encoding procedure.

FIG. 10 is an operation flowchart of a decoding procedure.

FIG. 11 is a chart of Japanese and English experimental voices by male and female voices.

12A is a diagram showing a speech waveform without watermark embedding, FIG. 12B is a diagram showing a speech waveform with watermark embedding, and FIG. 12C is a diagram showing speech waveforms with watermarks embedded therein. It is a figure which shows the waveform of the difference of ().

FIG. 13 is a chart showing an average opinion value of reproduced sound quality.

[14] The key k _p is a diagram showing the frequency of occurrence of the sum c _p of four pulse positions for fixed.

[15] The key k _p is a diagram showing the frequency of occurrence of the sum c _p of four pulse positions in the case of variations.

[Explanation of symbols]

1, 21 ... input voice, 2 ... preprocessing unit, 3 ...
Linear prediction analysis unit, 4 ... linear prediction synthesis filter, 5,
17 ... adder, 6 ... auditory weighting filter, 7
... Linear prediction coefficient, 8 ... Fixed codebook search, 9 ...
・ Pitch analysis, 10 ・ ・ ・ Parameter coding, 11 ・ ・
Transmission bit stream, 12: gain quantization, 13
... Adaptive codebook, 14 ... Fixed codebook, 15 ...
Adaptive codebook gain (pitch gain), 16... Fixed codebook gain, 22. 729 encoding device, 23 ... output code, 24, 27 ... control device, 25 ... embedded code selection signal, 26 ... decoding device, 28 ... embedded code search signal, 29 ... output voice, _{m 0 c p} · · · four total value of the pulse _position, the bit value of the key corresponding to the sum of the four pulse positions of the output speech code obtained by _{c pb} · · · _feedback, s 0 _~s 3 ··· polarity information,

Claims (3)

[Claims] 1. A method of embedding a watermark bit at least when digitally encoding and transmitting voice using a fixed codebook, wherein “1” or “1” is set as a candidate for a plurality of adjacent pulse positions in the fixed codebook. 0 ", and a first key for selecting the pulse position candidate based on" 1 "or" 0 "is determined. The first key is set at the bit position where the watermark is embedded in the transmission voice code. A watermark bit embedding method at the time of speech coding, characterized by using a pulse position selected by a key. 2. A method according to claim 1, wherein the fixed codebook assigns “1” or “0” to each of possible values of a predetermined total number of the pulse position candidates, and embeds the watermark. A second key for selecting whether to perform or not to perform according to whether the total value is “1” or “0” is determined, and the total value obtained by feedback from an output speech code is made to correspond to the second key, 2. The method according to claim 1, wherein the watermark is embedded. 3. A third key having a reverse assignment to the assignment of “1” and “0” in the first key is determined, and a total value of a predetermined number of the pulse position candidates is an even value. Or an odd value is detected, and one of the first key and the third key is associated with each of the even value and the odd value of the total value so as not to be the same key. The watermark bit at the time of voice encoding according to claim 2, wherein a pulse position selected by the first key or the third key is used as a bit position for embedding a watermark in the voice code. Embedding method.