JPH07248799A - Method and device for coding/decoding voice - Google Patents

Abstract

PURPOSE:To realize a method and device for coding/decoding voice capable of performing a precise voice coding process only by a small number of code books. CONSTITUTION:First of all, the voice data sampled in a prescribed time are discrete cosine-transformed from the data on a time base to the data on a frequency axis. Then, the data on the frequency axis are divided into an imaginary number component and a real number component. Then, an optimal frequency pattern most approximating to the data divided ;o into the imaginary number component and the real number component on the frequency axis from among frequency patterns beforehand stored in the code book 5 is selected, and the optimal frequency pattern is fitted to the data on the frequency axis, and a subtraction process is performed. The subtraction process by a waveform pattern is repeated until the maximum level of the data on the frequency axis after subtraction reaches a prescribed value.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a speech coding / decoding method and apparatus.

[0002]

2. Description of the Related Art Conventionally, a speech coding method of this kind has been as follows.

That is, a codebook corresponding to the waveform of voice data on the time axis is prepared in advance, and the codebook closest to the actual data on the time axis is selected from the codebooks. The number indicating the generated codebook is transmitted as compression encoded data of audio data.

[0004]

However, such a conventional method has the following problems. (1) Since the above codebooks correspond to waveforms on the time axis, various waveforms appear in voice data in order to perform compression with good sound quality. Therefore, the number of codebooks is enormous. Therefore, it is necessary to increase the capacity of the memory for storing the codebook. (2) Because a huge number of codebooks are required,
Therefore, in order to find the codebook corresponding to the waveform of the actual voice data, it is necessary to perform the process of applying the waveform of the voice data to a huge number of all codebooks, which takes a processing time accordingly.

The present invention solves the above problems and realizes a speech coding / decoding method and apparatus capable of performing highly accurate speech coding processing by using only a small number of codebooks. is there.

[0006]

In order to solve the above-mentioned problems, the present invention firstly requires that when audio data is encoded, the audio data is first subjected to discrete cosine transform in a predetermined time unit, and then the discrete cosine transform is performed. The codebook that detects the maximum level of the obtained data and the position information of this maximum level, and is close to the data around the maximum level in the data obtained by the discrete cosine transform from the codebook configured by the frequency pattern. Is selected, the codebook is subtracted from the data obtained by the discrete cosine transform, and the maximum level, the position information, the information indicating the codebook used for the subtraction, and the voice data within the predetermined time are selected. The encoded data of the audio data is generated based on the number of times the processing is performed.

Secondly, when the voice data is encoded by the above method, the number of times the codebook is subtracted from the data obtained by the discrete cosine transform is arbitrarily set to perform the encoding process, and the encoded data of the voice data is obtained. Is generated.

Thirdly, when the voice data is encoded by the first method, the minimum level for the subtraction is arbitrarily set in the level of the data obtained by the discrete cosine transform, and the obtained by the discrete cosine transform. The codebook is subtracted from the generated data to generate encoded data of voice data.

Fourthly, when the audio data is encoded by the first method, the same process as the process in the presence of voice is performed without performing a special process for the non-voice state even in the non-voice state. It is for generating encoded data of unvoiced data.

Fifth, when the speech data is encoded by the first method, the subtraction process is performed by matching the maximum level and the position of the codebook with the maximum level of the data obtained by the discrete cosine transform. It is for generating encoded data of audio data.

Sixth, when encoding the voice data by the first method, the data length of the encoded data to be generated is made variable according to the number of times the codebook is subtracted from the data obtained by the discrete cosine transform. To do.

[0012]

The present invention is based on the above-described method and configuration.
The discrete cosine transform is performed on the audio data sampled within a predetermined time from the data on the time axis to the data on the frequency axis. Next, the data on the frequency axis is divided into an odd component and an even component. Next, for the data on the frequency axis divided into the odd-numbered component and the even-numbered component, the optimum frequency pattern that is most approximated is selected from the frequency patterns stored in advance in a plurality of memories, and the data on the frequency axis is selected. Subtraction processing is performed by applying the optimum frequency pattern. The subtraction process using the waveform pattern is repeated until the maximum level of the data on the frequency axis after the subtraction reaches a predetermined value.

When the above processing is completed, (1) the maximum level of the data on the frequency axis before the processing, (2) the number of the frequency pattern used for the subtraction processing, and (3) the frequency position of the maximum level of the subtracted data. (4) Compressed and encoded data is generated and transmitted based on the information indicating the number of times the subtraction process is performed.

As a result, when the audio data is compression-encoded, the audio data is discrete cosine transformed into the data on the frequency axis, and then the data on the frequency axis is divided into its odd and even components. This increases the correlation coefficient in each of the odd-numbered component region and the even-numbered component region, so that the number of frequency patterns applied to the data on the frequency axis can be significantly reduced, and the frequency patterns stored in advance can be reduced. The number can be greatly reduced. Further, the capacity of the memory for storing the frequency pattern can be significantly reduced, and since the number of frequency patterns is small, it is possible to simplify the processing when compressing and encoding the audio data, and improve the processing speed. be able to.

Further, after achieving the above effects, the audio data compression-encoded data is obtained based on the maximum value of the data on the frequency axis before processing, the number of the subtracted waveform pattern, the position of the subtracted data, and the number of subtractions. Since the audio data is generated, it is possible to improve the compression efficiency of the audio data and ensure the reproducibility of the audio data.

[0016]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS An embodiment of the present invention will be described below with reference to the drawings.

First, the encoding of voice will be described. FIG. 1 is a block diagram of a speech coder for realizing a speech coding process according to the present invention. In FIG. 1, reference numeral 1 is an A / D converter for sampling and outputting audio data input as analog data at predetermined time intervals. Reference numeral 2 is a buffer for inputting the audio data sampled by the A / D converter 1, accumulating and outputting the audio data until a predetermined number is accumulated. In the present embodiment, the number of audio data stored in the buffer 2 is 12
Eight is set. Reference numeral 3 is a discrete cosine transforming means for performing a discrete cosine transform on the voice data output from the buffer 2. That is, as shown in FIG. 9, the audio data is converted from the data on the time axis (a) to the data on the frequency axis (b). Note that FIG. 9 will be described later. Reference numeral 4 is a maximum level detecting means for detecting the maximum level value of the voice data converted into the data on the frequency axis within a predetermined time and the frequency position where the maximum level exists. Reference numeral 5 is a codebook in which a pattern having a high frequency of occurrence is selected and stored from the patterns of the data on the frequency axis converted by the discrete cosine transform means 3. A frequency pattern as shown in FIG. 8 is used for the codebook 5, and FIG. 8 shows a typical frequency pattern. Although there are differences in frequency patterns depending on the size of the level, similar pattern shapes appear, so simply selecting a pattern with a high frequency of occurrence corresponds to the data on the frequency axis that actually appears sufficiently. be able to. A feature of the present invention is that a frequency pattern is used as a codebook, and the number of frequency patterns stored in the codebook 5 can be small. Also in this embodiment, the number of frequency patterns is 16 or less, which is sufficient. Reference numeral 6 is an optimum codebook selecting means for selecting from the codebook 5 a frequency pattern that is approximated at the maximum level frequency position output from the maximum level detecting means. Reference numeral 7 is a calculation means for comparing the actual data on the frequency axis with the frequency pattern selected from the codebook 5 and performing subtraction processing. Reference numeral 8 is an encoding means for generating and outputting compressed code data of audio data. The encoding means 8 outputs from the maximum level detection means 4 (1) the maximum level (hereinafter, simply referred to as maximum level) of the voice data converted into the data on the frequency axis within a predetermined time, and (2) the maximum. The frequency position where the level exists (hereinafter referred to as position information), and (3) the number of the frequency pattern output from the optimum codebook selecting means 6 (hereinafter referred to as the code number).
The compressed code data of the audio data is generated based on.

The speech coding process of the speech coding apparatus of the present invention having the above configuration will be described below with reference to FIGS. 2 to 10. FIG. 2 is a high-level flowchart showing an embodiment of the audio data encoding processing of the present invention. FIG. 3 shows steps in FIG. 2 (hereinafter referred to as ST)
6 is a flowchart showing a maximum level detection process in No. 4. FIG. 4 is a flowchart showing the optimum codebook selection process in ST6 of FIG. FIG. 5 is a flowchart showing the calculation process (subtraction) in ST7 of FIG. FIG. 6 is a flowchart showing the encoding process A in ST8 of FIG. FIG. 7 is a flowchart showing the encoding process B in ST11 of FIG. FIG. 8 is a diagram showing a representative pattern of frequency patterns stored in the codebook 7 of FIG. FIG. 9 is a diagram showing a state of data at a predetermined time when the audio data is subjected to the discrete cosine transform to be transformed from the data on the time axis to the data on the frequency axis. In FIG. 9, (a) shows the data on the time axis, (b) shows the state of discrete cosine transform of the data of (a) into the data on the frequency axis, and (c) shows the data on the receiving side. The reproduced voice data is shown by the data on the time axis, and (d) shows the data on the frequency axis before the inverse discrete cosine transform to obtain the data (c) on the time axis. FIG. 10 is a data structure diagram showing the structure of encoded compression-encoded data.

First, the outline of the speech coding process of this embodiment will be described with reference to FIG. ST before inputting voice data
The number of times the calculation process in 7 is repeated is set in the counter (ST1). As will be described later, the higher the set number of times, the higher the reproducibility of the compression-encoded audio data, and conversely, the smaller the set number of times, the lower the reproducibility of the compression-encoded audio data but the compression efficiency. Get higher

When voice data is input, the A / D converter 1
In step 1, the audio data is sampled in units of 128 every predetermined time. The A / D converter 1 outputs the sampled audio data in units of one sampling data. The buffer 2 stores the sampled voice data until 128 pieces are accumulated, and outputs it to the discrete cosine transform means 3. At this time, the audio data is in the state shown in FIG. The discrete cosine transform means 3 performs a discrete cosine transform on the input voice data into the data on the time axis and the data on the frequency axis (shown in FIG. 9B) (ST2), and separates it into odd and even components. Output above (ST3).

As described above, when the data on the frequency axis is separated into the odd component and the even component, the correlation coefficient in each region of the odd component and the even component increases, as shown in (b) of FIG. Data is generated.

Next, the maximum level of the data among the data on the frequency axis sampled 128 times within the predetermined time is detected (ST4) and the frequency position of the data having the maximum level is detected. ing. At this time, the maximum level means the size of the data, and the maximum data level in absolute value is detected.

If the maximum level is equal to or lower than the predetermined level (including 0) (ST5), it is considered that there is no sound, and the calculation processing based on the frequency pattern is not performed, and ST11 is set.
To the encoding process B. Here, according to the present embodiment, even if the process branches to the voiceless state, there is no special process for the voiceless state, which is the same as the process when there is voice ST1.
The process proceeds to the encoding process B of 1.

In ST5, if the maximum level is equal to or higher than the predetermined level, the process proceeds to the frequency pattern calculation process. The maximum level detecting means 4 outputs the maximum level (1) and the position information (2) to the optimum codebook selecting means 6.
The optimum codebook selecting means 6 sequentially extracts frequency patterns from the codebook 5. Assuming that there are five frequency patterns in the code book 5, the first frequency pattern is sequentially applied to the data on the actual frequency axis, and five data that are approximate to the data on the actual frequency axis are sequentially applied. Select from the frequency patterns. At this time, the frequency patterns stored in the codebook 5 are configured so that the maximum level is 1, as shown in FIG. Therefore, when applying the data on the actual frequency axis, the maximum level of the frequency pattern in the codebook 5 is multiplied by n so as to match the maximum level of the actual frequency pattern. In this way, the optimum codebook selecting means 6 selects the optimum frequency pattern from the codebook 5 (ST6) and outputs the code number (3) which is the number of the frequency pattern.

In the calculation process of ST7, the codebook 5 is optimized based on the code number (3) output from the optimum codebook selecting unit 6 on the data on the actual frequency axis output from the discrete cosine transforming unit 3. A process of extracting and subtracting a different frequency pattern is performed. The number of calculations (4) subtracted after the subtraction processing is output to the encoding means 8, and
The data on the frequency axis within the predetermined time after the subtraction is output to the maximum level detecting means 4.

In the encoding process A of ST8, the maximum level (1) and the position information (2) input from the maximum level detecting means 4 and the code number (3) input from the optimum codebook selecting means 6 are used to generate the data shown in FIG. As shown in (b), a frame forming the compression coded data is generated. The frame 1 is generated based on the first calculation processing in the calculation means 7, and the frame 2 is generated based on the second calculation processing. Therefore, the number of frames increases as the number of calculation processes increases.

When the encoding process A in ST8 by the encoding means is completed, the number of calculation processes set in ST1 is decremented (ST9), and it is determined whether the counter value has become 0 (ST10). If 0, ST
If it is not 0, the process goes back to the maximum level detecting means in ST3 to detect the maximum level in the data on the frequency axis after the subtraction processing, and ST5 → ST6 → ST7 → ST.
The process is repeated from 8 to ST9 to ST10 until the counter value becomes 0 in ST10.

Here, if a large counter value is set in advance in ST1, the number of calculations by the frequency pattern (ST7) increases, and it is possible to realize compression of voice data with high sound quality. On the other hand, if a large number of counter values are set in advance, the number of calculations will increase, and as a result, Figure 1
The number of frames of the compression-encoded data indicated by 0 (a) increases, the data length increases, and the compression rate of audio data decreases. As described above, according to the present invention, the compression rate of audio data can be arbitrarily changed.

Further, even if a large number of counter values are set in advance in ST1, the maximum level is set to a predetermined level in ST5 as in the case where the data on the frequency axis is almost the same as the frequency pattern in the codebook 5. If it falls below, even if the counter value does not reach 0 at that time, the calculation process is not repeated and is interrupted, and the encoding process B of ST11 is performed.
Go to. As described above, according to the present invention, when the data on the actual frequency axis and the frequency pattern are close to each other, the number of calculation processes is less than the preset counter value, and the actual frequency axis is When the data and the frequency pattern are not close to each other, the preset counter value is used, so that the compression rate can be increased while ensuring the reproducibility of the audio data.

The predetermined level in ST5 can be set arbitrarily. If the predetermined level is set high, the reproducibility of the audio data is deteriorated to that extent, while ST1
Since the calculation process is interrupted regardless of the counter value at, the compression rate increases accordingly. Further, if the predetermined level is set low, the reproducibility of the audio data becomes high, and on the other hand, the calculation process is continued until the counter value in ST1 reaches 0, so the compression rate becomes low accordingly.

At ST10, the counter value reaches 0,
Alternatively, when the maximum level becomes equal to or lower than the predetermined level in ST5, the process proceeds to the encoding process B in ST11. In the encoding process B of ST11, a process of combining frames 1 to n generated in the encoding process A of ST8 into one piece of compression encoded data according to the result of the calculation process at each number of times is performed. As shown in FIG. 10 (a), header information indicating that the audio data is within a predetermined time is placed at the head, and sequentially from frame 1 to frame n, compressed encoded data of the audio data within the predetermined time. Is being generated. When the compression encoded data is generated in this way, it is output to the outside.

In ST11, if the input voice data is in the non-voice state, the compression coded data is as shown in FIG.
It is generated by the configuration of only the header part in (a). When decoding the compressed coded data, if the compressed coded data is composed of only the header portion, it is determined that there is no voice. As described above, even when the voiceless state is determined, the process used only for the voiceless state is not specially performed, and the compression coded data is generated by the normal process when there is voice. Software can be reduced. Furthermore, when there is no audio data, the compression information decreases, so the compression rate of the audio data can be increased accordingly. The compression-coded data in the case of no voice may have a structure in which one frame is added to the header.

The above is the outline of the voice encoding process of the present embodiment. Hereinafter, the maximum level detection process of ST4 in FIG. 2 will be described in detail with reference to FIG.

First, the counter value is set and 0 is set to the maximum level (ST1). Here, the counter value indicates the number of samplings within a predetermined time. In the case of this embodiment, since the audio data is sampled 128 times within the predetermined time, the counter value becomes 128. Next, the voice data converted into the data on the frequency axis is input from the discrete cosine transform means 3. The following processing is repeated from the first sampled data to the 128th data on the frequency axis in order.

In ST2, the first sampled data is subtracted from the maximum level 0 set in ST1. If the first sampled data is set to 0, the subtraction result is 0, so the process proceeds to ST5 and ST1
Decrement the counter value at 128 from 127.

At ST6, it is determined whether or not the counter value at ST1 has become zero. In the above example, since the counter value is 127, the process returns to ST2. Again, ST2
It is assumed that the processing is repeated in the order of ST3 → ST5 → ST6, the processing proceeds with the maximum level being 0, and the data level is 50 when the counter value becomes 5. In ST2, level 50 is subtracted from maximum level 0. Since the subtraction result is -50, which is smaller than 0, the process proceeds to ST4. ST
In 4, the maximum level is input as 50 and the position information is input as 5. The process proceeds from ST5 → ST6 → ST2.

Next, it is assumed that the counter value becomes 6 and the data level is -30. In ST2, since the subtraction is performed by the absolute value, the level 30 is subtracted from the maximum value 50 updated last time. The subtraction result is 20, which is larger than 0, and thus the process proceeds to ST5. That is, the maximum level will not be updated. Processing is ST5 → ST6 →
Go to ST2.

It is assumed that the processing is repeated several times and the maximum level remains 50 when the counter value is 12.
It is assumed that the counter value is 13 and the data level is -80. In ST2, the absolute level 80 is subtracted from the maximum level 50. Since the subtraction result is -30, the value is smaller than 0, the process proceeds to ST4, the maximum level is updated from 50 to 80, and the position information is updated from 5 to 13.

In this way, the above processing is repeated up to the 128th sampled voice data, and the maximum level (1) and the position information (2) of the voice data converted into the data on the frequency axis within the predetermined time are obtained. Is detected and output.

The above is the maximum level detection processing. Less than,
The optimum codebook selection process of ST6 in FIG. 2 will be described in detail with reference to FIG.

First, the counter value is set, the maximum level (1) and the position information (2) are input from the maximum level detection processing 4, and an arbitrarily large value is input as the minimum subtraction error (S
T1). Here, the counter value indicates the number of frequency patterns stored in the codebook 5. In the case of the present embodiment, the number of frequency patterns is 5, as shown in (a) to (e) of FIG. For the minimum subtraction error, the actual data on the frequency axis and the frequency pattern in the codebook 5 are subtracted, and a value larger than the largest value that can occur in the sum of the subtraction results is input. is doing.

Next, the first stored frequency pattern is extracted from the codebook 5 (ST2). Here, it is assumed that (a) in FIG. 8 is taken out. The frequency pattern (a) is multiplied by n according to the maximum level (ST3). That is, the frequency pattern is stored with the maximum level height set to 1 in the codebook 5 as described above.
Therefore, assuming that the maximum level input from the maximum level detecting means 4 is 80, the frequency pattern (a) is multiplied by 80.

In ST4, the maximum level of the frequency-adjusted frequency pattern (a) is added to the actual maximum level of the data on the frequency axis and subtracted. Therefore, the actual maximum level of the data on the frequency axis disappears completely after the subtraction process. Then, the sum of the levels at each frequency position remaining after the subtraction is input to the subtraction error.

At ST5, the subtraction error at ST4 is subtracted from the minimum subtraction error at ST1. here,
The minimum subtraction error is 1000 and the subtraction error is 20.
When subtracted, it becomes 980, which is greater than 0 (ST6),
Go to ST7.

At ST7, the minimum subtraction error is 1000
To 20 and input 1 to the code number of the frequency pattern.

At ST8, the counter value at ST1 is decremented from 5 to 4. Next, the counter value is 0
Or not (ST9). Since the counter value is not 0 here, the process returns to ST2.

In ST2, the second frequency pattern is input from codebook 5. That is, here in FIG.
(B). Again, in accordance with the maximum level in the actual data on the frequency axis before the subtraction in ST4 of the previous time, the maximum level of the second frequency pattern (b) is set to n.
Double (ST3).

In ST4, the maximum level of the data on the actual frequency axis before the subtraction in the previous ST4 is matched with the maximum level of the second frequency pattern (b) to perform the subtraction, and the actual frequency remaining after the subtraction is performed. The sum of the levels at each frequency position of the data on the axis is input to the subtraction error.

At ST5, the subtraction error is subtracted from the minimum subtraction error. Here, the minimum subtraction error is updated to 20 the previous time, and the subtraction error is set to 10. Subtracting to 10 gives 0
Since it is larger (ST6), the minimum subtraction error is 20 to 1
It is updated to 0, and the code number is updated from 1 to 2 of the frequency pattern. The counter value is decremented,
Go to ST9 → ST2. The third frequency pattern is input from codebook 5 (ST2). That is, here, it is shown in FIG. Again, the maximum level of the third frequency pattern (C) is multiplied by n in accordance with the maximum level of the data on the first actual frequency axis (ST3).

In ST4, the maximum level of the data on the first actual frequency axis is matched with the maximum level of the third frequency pattern (C), and subtraction is performed. The sum of the levels at each frequency position is input to the subtraction error.

At ST5, the subtraction error is subtracted from the minimum subtraction error. Here, the minimum subtraction error is updated to 10 the previous time, and the subtraction error is set to 30. The subtraction results in 120, which is smaller than 0 (ST6). Therefore, the minimum subtraction error is not updated and remains 10, and the process proceeds to ST8 → ST9 → ST2.

The above processing is repeated until the counter value reaches 0 in ST9. Here, since the number of frequency patterns stored in the codebook 5 is 5, it is repeated 5 times. If the process proceeds to the fifth time, the minimum subtraction error in ST7 is 10, and the code number is 2, the second frequency pattern (b)
Is selected as the closest approximation of the frequency patterns in codebook 5. As described above, in the present invention, when the codebook is applied to the audio data to perform the compression encoding process, the audio data is subjected to the discrete cosine transform from the data on the time axis to the data on the frequency axis, and the frequency pattern is applied. There is. Since the frequency patterns that appear often have similar shapes, although there are differences in level, 1
It is not necessary to store as many as 000 patterns in advance, and even in the process of selecting the optimum codebook, only one is selected from a small number of frequency patterns, so that the process can be easily performed. The processing speed can be remarkably increased as compared with the conventional one.

The above is the optimum codebook selection processing.
Hereinafter, the calculation process (subtraction) in ST7 in FIG. 2 will be described in detail with reference to FIG.

First, the calculating means 7 inputs the maximum level (1) and the position information (2) from the maximum level detecting means 4, and the optimum codebook selecting means 6 inputs the code number (3).
Is input (ST1). Next, the frequency pattern is extracted from the codebook 5 based on the code number (3). Then, the frequency pattern is multiplied by n according to the maximum level in the actual frequency axis data within a predetermined time.

In ST4, the frequency pattern is subtracted from the actual frequency axis data in accordance with the maximum level frequency position in the actual frequency axis data in accordance with the maximum level frequency position in the actual frequency axis data (ST4). As a result of the subtraction, the level remaining in the actual frequency axis data is returned to the maximum level detecting means 4.

When the subtraction is performed here, it is performed in accordance with the maximum level frequency position in the actual frequency axis data and the maximum level frequency position in the frequency pattern. This reduces the counter value in ST1 of the figure. Even if you set and cancel the subtraction process with a small number of times,
Since the subtraction process is performed based on the maximum level of the data, it is possible to ensure a reasonable reproducibility and increase the compression rate.

The above is the calculation processing. Hereinafter, the encoding process A of ST8 in FIG. 2 will be described in detail with reference to FIG.

First, the relative level (1) is input from the maximum level detecting means 4 (ST1). Here, the relative level is the maximum level, but from ST4 to ST1 in FIG.
The maximum level changes while the processes up to 0 are repeated, and the relative level indicates the maximum level that changes in each process.

Next, the position information (2) corresponding to the relative level in ST1 is input from the maximum level detecting means 4 (ST2).

Then, the code number (3) is input from the optimum codebook selecting means 6. In ST4, each coded frame is generated from the relative level (1), the position information (2), and the code number (3) as shown in FIG.

The above is the encoding process A. Hereinafter, the encoding process B of ST10 in FIG. 2 will be described in detail with reference to FIG.

First, when the process of the number of calculation processes set in ST1 of FIG. 2 is finished, the compression coding process of the voice data within the predetermined time is finished, and the compression code of the voice data within the predetermined time is finished. The process will move to the process of generating the encoded data.

Here, first, the header information shown in FIG. 10A is generated. The header information includes a block identifier, a maximum level (1), and the number of calculations (4). The block identifier indicates each section of each audio data within a predetermined time, and by detecting the block identifier when reproducing the audio data, the audio data within a certain predetermined time is changed to the audio data within the next predetermined time. You can identify that you have moved to. Maximum level (1)
Is the maximum level of the voice data converted on the frequency axis within the predetermined time, and is the maximum level detected first in the maximum level detection process of ST4 in FIG. The number of calculations (4) is the count value set in ST1 of FIG.
It shows how many times the processing up to T9 was repeated.

The coding means 8 inputs the block identifier (ST1), the maximum level (1) from the maximum level detecting means 4 (ST2), and the calculation count (4) from the calculating means 7 ( ST3) shows the header information shown in (b) of FIG.

Then, each time the processing of FIG. 2 is repeated, the frame generated by the encoding processing A in ST8 of FIG. 2 is input (ST4), the frame is sequentially added to the header information, and the frame of FIG. A coded block as shown in (a) is generated (ST5), and the coded block is output to the outside. Here, the number of frames added to the header information increases according to the number of times the processing in FIG. 2 is repeated. Therefore, if the number of times the process in FIG. 2 is repeated is large, the data length of the coding block (compressed coded data) shown in FIG. 10A becomes long and the compression rate becomes low accordingly. On the other hand, if the number of times the process in FIG. 2 is repeated is small, the data length of the coded block (compressed coded data) shown in FIG. 10A becomes short and the compression rate becomes high accordingly. As described above, by arbitrarily setting the counter value in ST1 of FIG. 2, the compression rate of audio data can be changed.

The above is the encoding process B. As is apparent from the above description, according to the present invention, when compressing and encoding audio data, the data on the frequency axis is separated into the odd component and the even component, and then the audio data is first extracted from the data on the time axis. Discrete cosine transform on the frequency axis is performed, then it is separated into an odd component and an even component, and a codebook composed of frequency patterns is applied and subtracted. When this subtraction process is performed, the data on the frequency axis is separated into odd-numbered components and even-numbered components, and then the codebook is performed by the frequency pattern, so that the frequency pattern appears even if the level difference is different. Since the shapes of the patterns are similar to each other, the number of codebooks provided in advance can be small, and the selection process for selecting a codebook suitable for actual voice data can be easily executed with a small number of processes. In addition, the processing speed of the selection processing can be increased by the number of processings.

The compression coded data of the voice data is
The maximum level (1) obtained in the process of performing the subtraction process,
Generated from the position information (2), the code number (3), and the number of calculations (4), the data length of the compression-coded data increases in proportion to the number of times of the subtraction process when the number of times of the subtraction process is increased. I am trying. Accordingly, the data length of the compression-coded data, that is, the compression rate of the audio data can be arbitrarily changed by setting the number of times of the subtraction processing, and thus the compression rate of the audio data can be arbitrarily increased. Furthermore, even when the number of times of subtraction processing is set to be small and the compression rate is increased, since the subtraction processing is performed based on the maximum level of data, it is possible to guarantee the reproducibility to the extent that the content and characteristics of voice can be discriminated. it can.

The process of reproducing the audio data compressed and encoded as described above will be described below with reference to the drawings.

FIG. 11 is a block diagram of a speech decoding apparatus which realizes speech decoding processing according to the present invention. In FIG. 11, reference numeral 11 is a decoding means for extracting the maximum level (1), position information (2), code number (3), and number of calculations (4) from the input compression-coded data. Reference numeral 12 is a codebook that manages and stores the same number and the same number of frequency patterns as the frequency patterns used for the voice encoding process, using the same code number. Reference numeral 13 indicates the frequency of the voice data based on the maximum level (1), the position information (2), the code number (3), the number of calculations (4) input from the decoding means, and the frequency pattern input from the codebook 12. It is a calculation means that reproduces as data on the axis by calculation (addition). 1
Reference numeral 4 is an inverse discrete cosine transforming means for performing an inverse discrete cosine transform on the data on the frequency axis output from the calculating means 13 into the data on the time axis. 15 is an inverse discrete cosine transform means 1
4 is a buffer that outputs the audio data input from 4. Reference numeral 16 is a D / A that outputs data after D / A converting it in the unit of one sampling data input from the buffer 15.
It is an A converter.

The speech decoding process of the speech decoding apparatus of the present invention configured as above will be described below with reference to FIGS. 12 to 15. FIG. 12 is a high-level flowchart showing an embodiment of the audio data decoding processing of the present invention. FIG. 13 is a flowchart showing the decoding process A in ST1 of FIG. FIG. 14 shows ST3 of FIG.
5 is a flowchart showing a decoding process B in FIG.
FIG. 15 is a flowchart showing the calculation process (addition) in ST4 of FIG. FIG. 16 is a flowchart showing the inverse discrete cosine transform process in ST7 of FIG.

First, the outline of the speech decoding process of this embodiment will be described with reference to FIG. When the compression coded data is input, the decoding process A is performed (ST1). The decoding process A is a process of decomposing the coded block of FIG. 10A into header information and a frame, and detecting the maximum level (1) and the number of calculations (4) from the header information.

Next, the counter value is set based on the number of calculations (4) detected from the header information (ST).
2). That is, the addition process of ST4 is performed the same number of times as the number of subtraction processes performed in the encoding process of the audio data according to FIG.

Then, the decoding process B is performed. The decoding process B is performed on the frame n in the coded block of FIG.
Is a process for detecting the relative level (1), position information (2), and code number (3) set in each frame.

In ST4, calculation processing (addition) is performed in ST4. The calculation process extracts the corresponding frequency pattern from the codebook 12 based on the code number (3) input from the decoding means 11, and determines the maximum level of the frequency pattern based on the relative level (1) input from the decoding means 11. Multiply by n. In the code book 12, the maximum level of each frequency pattern is set to 1 and stored as shown in FIG. Next, the frequency pattern multiplied by n based on the relative level (1) is added to the position indicated by the position information (2). The position information (2) indicates the number of the position of the audio data within a predetermined time divided into 128, as in the case of the encoding process.

When the calculation process is completed, the counter value is decremented (ST5). Until the counter value reaches 0 (ST6), the process is repeated as ST3 → ST4 → ST5 → ST6. During this period, the relative level (1), the position information (2), and the code number (3) are sequentially detected from the subsequent frames,
A process of taking out the corresponding frequency pattern from the codebook 12, multiplying the frequency pattern by n, and adding it to the position indicated by the position information (2) is performed.

When the above processing is repeated and the counter value reaches 0, the data as shown in FIG. 9D is generated. The inverse discrete cosine transform of this data is performed, and
The data is converted into data on the time axis as shown in (ST7).
The data converted into the data on the time axis is output after being converted into analog data in the D / A conversion unit 16.

The above is the outline of the speech decoding processing according to the present embodiment. Hereinafter, with reference to FIG. 13, ST1 in FIG.
The decoding means A will be described in detail.

First, when the compressed coded data (coded block) is input to the decoding means 11 (ST1), the header information is extracted from the coded blocks constituting the compressed coded data, and the maximum level (1) and the number of calculations are calculated. (4)
Is output (ST2, ST3).

The above is the decoding means A. Below, FIG.
The decoding means B of ST3 in FIG. 12 will be explained in detail using.

First, the decoding means 11 analyzes the first frame in the compression coded data (coded block) (ST1). As a result of the analysis, the relative level (1), the position information (2), and the code number (3) are output (ST2, ST
3, ST4). Returning to the flowchart of FIG. 14, ST4 →
When the process proceeds from ST5 to ST6 to ST3 and returns to the decoding means B again, the above-described process is performed for the second frame in the compression coded data, and the counter value reaches 0 in ST6 of FIG. Repeat until.

The above is the decoding means B. Below, FIG.
The calculation process (addition) of ST4 in FIG. 12 will be described in detail using.

First, the calculation means 13 inputs the relative level (1), the position information (2) and the code number (3) from the decoding means 11 (ST1). Next, the frequency pattern corresponding to the code number (3) is extracted from the code book 12 (ST2). Then, the maximum level of the frequency pattern is multiplied by n based on the relative level (1) (S
T3). The frequency pattern multiplied by n is used as the position information (2).
Then, the maximum level position of the frequency pattern is aligned and added (ST4). When the addition of the frequency pattern is completed, the process returns to the flowchart of FIG. 12 and S of FIG.
ST5 → ST until the counter value reaches 0 at T6
The process of 6 → ST3 → ST4 is repeated, and the relative level (1), the position information (2), and the code number (3) obtained from the analysis result of the second frame in the compression coded data are again obtained in the calculation process of ST4. The frequency pattern of this time is added to the input result of the previous addition. The calculation process is repeated until the counter value reaches 0 in ST6 of FIG.

The above is the calculation process (addition). As is apparent from the above description, according to the present invention, when reproducing compressed audio data, first, the relative level (1), position information (2), code number (3), number of calculations ( 4) is detected, a codebook is selected and added according to the information obtained from the first frame of the compression encoded data. Then, the process is performed by the information obtained from the next frame of the compression-encoded data, and the decoding process is performed by repeating the process for the number of times based on the calculation count (4). By this decoding processing, the compression-encoded audio data can be reliably reproduced.

[0084]

As is apparent from the above description, according to the present invention, when compressing and coding voice data, first, the voice data is subjected to discrete cosine transform from data on the time axis to data on the frequency axis, Next, the data after the discrete cosine transformation is separated into an odd component and an even component, and the data on the frequency axis obtained as a result of the separation is subjected to a subtraction by applying a codebook composed of a frequency pattern. When performing this subtraction process, the data on the frequency axis is separated into the odd component and the even component, so the correlation coefficient increases in each region of the odd component and the even component. Patterns that are approximately similar to each other will appear. Moreover, if the codebook is composed of frequency patterns, the shapes of the patterns that appear are similar to each other even if the difference in level of the frequency patterns is different, so the number of codebooks to be provided in advance can be reduced. Even if it is drastically reduced, it can be adapted to the actual data on the frequency axis, and the selection process to select the codebook that matches the actual voice data can be easily executed with a small number of processing steps. It is possible to realize an effect that the processing speed of the selection processing can be increased by the number.

When the subtraction process is performed, the number of times of the subtraction process can be set arbitrarily. Therefore, if the number of times of the subtraction process is set in advance, it is possible to realize compression of voice data having high sound quality. On the other hand, if the number of times is set in advance, the number of calculations increases, the data length of the compression-coded data increases, and the compression rate of the audio data decreases. In this way, it is possible to arbitrarily change the compression ratio of audio data.

Further, even when the number of times of the above subtraction processing is set in advance, if the data on the frequency axis and the selected frequency pattern have almost the same shape,
When the maximum level of the actual data falls below the preset minimum level due to the subtraction, the subtraction process is interrupted at that time even if the number of times has not reached the above number. As described above, when the actual data on the frequency axis and the frequency pattern are close to each other, the number of times of the subtraction process is less than the preset number, and the actual data on the frequency axis and the frequency pattern are If is not approximate, the process is performed up to a preset number of times, so even if the number of times is set to a large number, if the data on the frequency axis and the selected frequency pattern have almost the same shape, the process is interrupted. As a result, the compression rate can be increased while ensuring the reproducibility of audio data.

Further, even when the silent state is determined,
Since the compression-encoded data is generated by the normal process when there is voice, the process used only for the non-voice state is not specially performed, and the software can be reduced accordingly. Furthermore, when there is no audio data, the compression information decreases, so the compression rate of the audio data can be increased accordingly.

Further, when performing the above subtraction processing, even if the number of subtraction processings is set to be small and the compression rate is increased, since the codebook is applied with reference to the maximum level of actual data, the subtraction processing is performed. It is possible to guarantee reproducibility to the extent that the content and characteristics of voice can be discriminated.

The compression coded data of the audio data obtained by the subtraction process is the maximum level (1) obtained by the process of the subtraction process, the position information (2), the code number (3), the number of calculations. Since the data length of the compression-coded data generated in (4) increases in proportion to the number of times of the subtraction processing when the number of times of performing the subtraction processing is increased, the compression rate of the audio data is Since the setting can be arbitrarily changed by setting the number of times, the data length of the compression encoded data can be arbitrarily changed and the compression rate of the audio data can be arbitrarily increased.

[Brief description of drawings]

FIG. 1 is a block diagram of a voice encoding process according to the present invention.

FIG. 2 is a high-level flow chart showing an embodiment of a voice data encoding process of the present invention.

FIG. 3 is a flowchart showing a maximum level detection process in ST4 of FIG.

FIG. 4 is a flowchart showing an optimum codebook selection process in ST6 of FIG.

5 is a flowchart showing a calculation process (subtraction) in ST7 of FIG.

FIG. 6 is a flowchart showing an encoding process A in ST8 of FIG.

FIG. 7 is a flowchart showing an encoding process B in ST11 of FIG.

8 is a diagram showing a representative pattern of frequency patterns stored in the codebook 5 of FIG.

FIG. 9 is a diagram showing a state of data at a predetermined time when voice data is subjected to discrete cosine transform and converted from data on a time axis to data on a frequency axis.

FIG. 10 is a data configuration diagram showing a configuration of encoded compression encoded data.

FIG. 11 is a block diagram for realizing a speech decoding process according to the present invention.

FIG. 12 is a high-level flowchart showing an embodiment of the audio data decoding process of the present invention.

13 is a flowchart showing a decoding process A in ST1 of FIG.

14 is a flowchart showing a decoding process B in ST3 of FIG.

FIG. 15 is a flowchart showing calculation processing (addition) in ST4 of FIG.

[Explanation of symbols]

 3 Discrete Cosine Transforming Means 4 Maximum Level Detecting Means 5 Codebook 6 Optimum Codebook Selecting Means 7 Calculating Means 8 Encoding Means

Claims (9)

[Claims] 1. A first process for performing a discrete cosine transform of audio data in a predetermined time unit, a second process for separating the data obtained by the discrete cosine transform into an odd component and an even component, and the separated process. A third process for detecting the maximum level of data and position information of this maximum level and a codebook composed of frequency patterns were used to approximate the data obtained by the discrete cosine transform to the data around the maximum level. The fourth process of selecting the codebook, the fifth process of subtracting the codebook from the data obtained by the discrete cosine transform, the maximum level, the position information, and the codebook used for the subtraction are shown. The encoded data of the audio data is generated based on the information and the number of times of the fifth processing performed on the audio data within the predetermined time. A speech encoding method comprising a sixth process. 2. A means for performing discrete cosine transform of audio data in a predetermined time unit, a means for separating the separated data into an odd component and an even component, a maximum level of the separated data and this maximum level. A means for detecting position information, a means for storing a codebook constituted by a frequency pattern, a means for selecting the codebook approximated to the data around the maximum level in the data obtained by the discrete cosine transform, Means for subtracting the codebook from the data around the maximum level, the maximum level, the position information, information indicating the codebook used for the subtraction, and the subtraction performed on the voice data within the predetermined time. A voice encoding device, comprising: means for generating encoded data of voice data based on the number of times. 3. A means for performing discrete cosine transform of audio data in a predetermined time unit, a means for separating the separated data into an odd component and an even component, a maximum level of the separated data and this maximum level. A means for detecting position information, a means for storing a codebook constituted by a frequency pattern, a means for selecting the codebook approximated to the data around the maximum level in the data obtained by the discrete cosine transform, Means for subtracting the codebook from data around the maximum level, means for arbitrarily setting the number of times of the subtraction, the maximum level, the position information, information indicating the codebook used for subtraction, the predetermined time Means for generating encoded data of audio data based on the number of times of the subtraction performed on the audio data in Audio coding device. 4. A means for performing discrete cosine transform of audio data in a predetermined time unit, a means for separating the separated data into an odd component and an even component, a maximum level of the separated data and a maximum level of the maximum level. Means for detecting position information, means for storing a codebook composed of frequency patterns, means for selecting the codebook that approximates the data around the maximum level in the data obtained by the discrete cosine transform, Means for subtracting the codebook from the data around the maximum level, means for arbitrarily setting the minimum level for the subtraction in the level of the data obtained by the discrete cosine transform, the maximum level, the position information, and subtraction Information indicating the codebook used, the number of times of subtraction performed on the audio data within the predetermined time. A speech coding apparatus comprising means for generating coded data of speech data based on a number. 5. A first process for performing a discrete cosine transform of audio data in a predetermined time unit, a second process for separating the data obtained by the discrete cosine transform into an odd component and an even component, and the separated process. A third process for detecting the maximum level of data and position information of this maximum level and a codebook composed of frequency patterns were used to approximate the data obtained by the discrete cosine transform to the data around the maximum level. A fourth process of selecting the codebook, a fifth process of subtracting the codebook from data around the maximum level, and a maximum level of
A fifth process for generating encoded data of audio data based on the position information, information indicating the codebook used for subtraction, and the number of times of the fifth process performed on the audio data within the predetermined time. And a voice coding method for generating coded data of voiceless data by the sixth process even in the voiceless state. 6. When performing the fifth process, the codebook obtained by the fourth process is multiplied by n in accordance with the maximum level of the data obtained by the discrete cosine transform, and the maximum of the data obtained by the discrete cosine transform is obtained. 2. The speech coding method according to claim 1, wherein the position of the maximum level of the codebook is aligned with the position of the level, and the data obtained by the discrete cosine transform is subtracted from the codebook. 7. The voice encoding method according to claim 1, wherein when performing the sixth process, the data length of the encoded data is variably generated according to the number of times of the fifth process. 8. A first process of inputting coded data of voice data generated by the voice coding method according to claim 1, a maximum level from the coded data, position information of this maximum level, and a frequency pattern. A second process configured to extract information indicating a codebook used in the encoding process, and selecting a codebook based on the information indicating the codebook, and selecting the codebook at the position indicated by the position information. And a third process for reproducing the voice data from the coded data by adding the maximum levels of the above. 9. A unit for inputting encoded data of voice data generated by the voice encoding device according to claim 2.
The number of operations performed on the voice data within a predetermined time is extracted from the encoded data, and the maximum level from the frame generated according to the number of operations, the position information of the maximum level, and the frequency pattern are used. Means for extracting information indicating the codebook used in the encoded processing,
A means for storing a codebook configured by a frequency pattern, a codebook is selected based on information indicating the codebook, and means for adding the maximum level of the codebook to the position indicated by the position information, A speech decoding apparatus, comprising: means for repeating the addition of the codebook by the number of times of the extracted operations, and reproducing speech data from the encoded data.