JP3661363B2 - Audio compression / decompression method and apparatus, and storage medium storing audio compression / decompression processing program - Google Patents

Description

[0001]
BACKGROUND OF THE INVENTION
The present invention relates to an audio compression / decompression method and apparatus for efficiently compressing / decompressing an audio signal by simple processing, and a storage medium storing an audio compression / decompression processing program.
[0002]
[Prior art]
Conventionally, various methods have been proposed as an encoding method for compressing and expanding an audio signal. One of them is JP-A-59-116973 (hereinafter referred to as the first prior art).
[0003]
This first prior art includes means for dividing input speech data every short time to obtain a short time speech signal sequence, spectrum envelope parameter extracting means for extracting a spectrum envelope parameter from the short time speech signal sequence, and the spectrum envelope. Means for calculating an impulse response sequence based on parameters, means for calculating an autocorrelation function sequence using the impulse response sequence, and a cross-correlation function using the impulse response sequence and the short time speech signal sequence Means for calculating a sequence, means for calculating and encoding a driving excitation signal sequence using the autocorrelation function sequence and the cross-correlation function sequence, means for outputting a combination of a spectral envelope code and a driving excitation signal, Furthermore, it has target signal calculation means for applying a predetermined correction to the short time audio signal.
[0004]
According to the first prior art, the position and gain of the driving sound source pulse can be determined efficiently when speech is encoded, and a certain amount of effect can be achieved in reducing the amount of calculation and the amount of memory used. Is obtained.
[0005]
However, in the first prior art, when speech synthesis is performed after encoding a speech signal such as a female voice, it is necessary to extract a large number of driving sound source pulses in order to obtain high-quality speech synthesis. For this reason, there is a problem that the compression rate is deteriorated.
[0006]
That is, a female voice is more complex than a male voice, and in order to obtain a highly accurate synthesized sound, it is necessary to extract a large number of driving sound source pulses, which ultimately results in a poor compression rate.
[0007]
On the other hand, as techniques for obtaining a high compression rate, there are JP-A-63-37399 (hereinafter referred to as the second prior art) and JP-A-3-4300 (hereinafter referred to as the third prior art).
[0008]
In the second conventional technique, pitch estimation is performed from a speech signal, a residual between an estimated value from a past pulse train and an actual signal is obtained, and a drive sound source pulse is calculated from the residual.
[0009]
In the third prior art, pitch estimation is performed, and a driving sound source (multi-pulse) for one pitch section is estimated. Then, by correcting the gain and phase of the multipulse, the other pitch sections are approximated by correcting the other pitch sections. Further, the second multi-pulse is estimated from the residual between the estimated value and the actual value. In addition to the multi-pulse signal, a noise code book may be used.
[0010]
[Problems to be solved by the invention]
The second and third prior arts described above calculate the period of repeating the same waveform, estimate the next period from the previous period, calculate the difference between the estimated part and the actual speech waveform, Since the driving sound source is calculated based on the difference, a high compression rate can be realized.
[0011]
However, since it is necessary to obtain the pitch or the difference, there is a problem that the calculation amount is large, and a memory having a large capacity is required to store the data.
[0012]
In addition, since the residual is calculated and the driving sound source pulse is calculated based on this residual, if a part of the data is lost, the lost data will have a large effect on the subsequent calculations, resulting in high accuracy. There is a big problem that it is impossible to perform proper speech synthesis.
[0013]
As described above, each of the conventional techniques has various problems. For example, the first conventional technique is a basic technique for obtaining a driving sound source pulse. However, in order to improve the quality of a synthesized sound, it is necessary to set up many driving sound source pulses, like a female voice. There is a problem that the compression rate is particularly bad for voice data. Moreover, although the 2nd prior art and the 3rd prior art can obtain a high compression rate, there is a problem that the calculation amount is large and the amount of memory used is also large, and further, because the difference information is used, it is vulnerable to data loss. There's a problem.
[0014]
Recently, portable information devices that handle audio data have been used in a wide range of fields. In this type of portable information device, since the calculation speed and memory capacity of the CPU are greatly restricted, it is a serious problem that the amount of calculation and the amount of memory used are large. In addition, the method using difference information has many problems in terms of improving the performance of products for information devices that need to consider data loss, and is not limited to portable devices, but also for real-time transmission on a computer network. This lack of data will also have a significant impact on the data being transmitted.
[0015]
As described above, each of the conventional speech coding methods has a common problem that the processing is complicated, and there is a problem that it is relatively difficult to increase the speed by hardware and parallel processing. In particular, the processing including the process for obtaining the pitch period has a large amount of calculation and has a great influence when an error occurs. Further, the conventional impulse response using the spectral envelope parameter and the method using the drive pulse have a problem that discontinuity occurs before and after the pulse, and this appears as noise.
[0016]
Therefore, the present invention has a simple processing content, enables easy hardware and parallel processing, enables efficient encoding, and enables audio data compression at a relatively high compression rate. An object of the present invention is to provide an audio compression / decompression method and apparatus and a storage medium storing an audio compression / decompression processing program.
[0017]
The speech compression / decompression method of the present invention cuts out a speech segment in a predetermined section from input speech, extracts a spectrum envelope parameter from the speech segment in the extracted predetermined section at a predetermined frequency, and estimates it by the extracted spectrum envelope parameter. A plurality of types of speech segments including a first speech segment created using a temporally predicted speech waveform and a temporally predicted speech waveform continuous with the temporally predicted speech waveform. Compare the similarity of the segment with the speech segment, select the speech segment with the highest similarity from the plurality of types of speech segment, and based on the data about the selected speech segment, It is characterized by including a process of encoding a speech piece to create encoded data.
[0018]
Further, after creating the encoded data, expand the encoded data, subtract the expanded data from the audio segment of the predetermined section that was cut out to obtain a residual, for the waveform of the residual, The plurality of types of speech segments are referred to, and a process of comparing the similarity between the plurality of types of speech segments and the residual waveform is performed to obtain encoded data.
[0019]
The plurality of types of speech pieces are created by using a second speech piece and a noise component that are created by using the speech waveform that has been subjected to the decompression process that is temporally behind the speech piece of the cut out predetermined section. A third voice piece, the first voice piece is updated in content after extraction of the spectral envelope parameter, and the second voice piece is a code created using the second voice piece. After the decompressed data is subjected to the decompression process, the contents are updated based on the decompressed data.
[0020]
In addition, the plurality of types of speech pieces have sections that are longer in time than the speech pieces of the cut out predetermined section, and when comparing the similarity with the cut out voice pieces of the predetermined section, In the range of the length of the voice segment, the similarity with the extracted voice segment of the predetermined section is compared, and the voice segment having the highest similarity is selected.
[0021]
The encoded data includes the number of the speech piece having the highest similarity part, position data indicating which part in the speech piece is included, and an amplitude adjustment parameter. Thus, the spectrum envelope parameter is added.
[0022]
The speech compression / decompression apparatus according to the present invention includes a speech segment extraction unit that extracts a speech segment of a predetermined section from input speech, and a spectral envelope parameter from the speech segment of the predetermined section that is extracted by the speech segment extraction unit at a predetermined frequency. A plurality of first speech segments created using a temporally predictive speech waveform estimated by the extracted spectral envelope parameter and a temporally predictive speech waveform continuous therewith, A similarity determination unit that refers to a plurality of types of speech pieces and compares the similarity between the plurality of types of speech pieces and the extracted speech piece of the predetermined section, and obtains a similarity, Based on the similarity by the similarity determination unit, an audio segment selection unit that selects an audio segment with the highest similarity, and an audio segment selected by the audio segment selection unit. Is characterized by having an encoding unit for encoding the speech segments of a predetermined interval data cut out the based on the, the.
[0023]
Also, a decompression unit that decompresses the data encoded by the encoding unit, and a speech segment update unit that updates the content of the corresponding speech segment using the data decompressed by the decompression unit or the spectrum envelope parameter And so on.
[0024]
And a residual generation unit that obtains a residual by subtracting the data expanded by the expansion unit from the extracted speech segment of the predetermined section, the similarity determination unit, the speech segment selection unit, and the encoding The expansion unit, the extension unit, and the residual generation unit form a loop in the order of processing, refer to the plurality of types of speech pieces for the residual waveform generated by the residual generation unit, and After performing the process of comparing the similarity with the speech piece, encoded data is created and output.
[0025]
Further, the plurality of types of speech pieces are a second speech piece created by using the speech waveform that has been subjected to the decompression processing that is temporally rearward of the extracted speech piece of the predetermined section, and a noise component. A third speech piece created, and the speech piece update unit updates the first speech piece after extraction of the spectral envelope parameter, and the coding is created using the second speech piece. After the data is subjected to the decompression process, the second audio piece is updated based on the data subjected to the decompression process.
[0026]
Further, the plurality of types of speech pieces have a section that is longer in time than the extracted speech section of the predetermined section, and the similarity determination unit includes a range of lengths of the plurality of types of speech pieces, The similarity is obtained by comparing the similarities, and the speech piece selection unit selects a speech piece having a portion with the highest similarity.
[0027]
Further, the encoded data is data represented by the number of the speech segment having the highest portion, the position data representing the highest portion, and the parameter for amplitude adjustment. It has a parameter.
[0028]
Further, the recording medium of the present invention cuts out a voice segment of a predetermined section from the input voice, extracts a spectrum envelope parameter from the voice segment of the cut out predetermined section at a predetermined frequency, and estimates it by the extracted spectrum envelope parameter. Referring to a plurality of types of speech segments including a first speech segment created using a temporally predicted speech waveform and a temporally predicted speech waveform that is continuous therewith, the plurality of types of speech segments and the speech of the predetermined section Compare the similarity with the piece, select the voice piece with the highest similarity, and encode the voice piece in the predetermined section based on the data about the selected voice piece to create the coded data It is a storage medium storing an audio compression / decompression processing program for causing a computer to execute processing.
[0029]
As described above, the present invention compares the similarity between each of a plurality of types of speech pieces and a speech piece of a predetermined section extracted from the input speech (for example, a speech piece having a length of about 4 msec), and obtains the highest similarity. A process of selecting a high speech segment and encoding the speech segment in the predetermined section cut out based on data about the selected speech segment is performed as a basic process. As a result, encoding can be performed with extremely simple processing, which can be advantageous when performing hardware and parallel processing.
In particular, when a predicted speech waveform estimated by a spectral envelope parameter is used, conventionally, only a temporally forward predicted speech waveform (impulse response) is generally used, but the present invention is estimated by a spectral envelope parameter. A speech segment is created using a temporally predicted speech waveform and a temporally backward predicted speech waveform that is continuous therewith.
In this way, in addition to the forward predicted speech waveform, the use of the backward predicted speech waveform in terms of time has the effect of reducing noise. That is, when a voice piece using only the impulse response (forward predicted voice waveform) is used, a voice piece whose waveform rises suddenly from a state where the voice level is almost 0 is generated. When processed, there is a problem in that a discontinuous point appears and that part appears as noise. On the other hand, when the backward predicted speech waveform behind the time is used, the discontinuity can be reduced as much as possible, and the quality of the compressed and expanded speech can be greatly improved.
[0030]
Also, after creating encoded data, the encoded data is decompressed, the residual data is subtracted from the extracted speech segment of the predetermined section, and a plurality of types are again applied to the residual waveform. By obtaining the encoded data by performing the process of obtaining the similarity once or more with reference to the voice piece, it is possible to obtain the encoded data with higher accuracy.
[0031]
Further, the plurality of types of speech pieces are created by using a second speech piece and a noise component created by using the speech waveform that has been subjected to the decompression process that is temporally behind the cut out speech piece of the predetermined section. By having the third speech piece, the input speech can be encoded efficiently and with high accuracy.
[0033]
In addition, the content of the first speech segment is updated after the extraction of the spectral envelope parameter, and the second speech segment is encoded data created using the second speech segment is subjected to the decompression process. Since the contents are updated based on the decompressed data, the extracted audio data in the predetermined section is different from the codebook having the fixed contents as in the past. The optimum audio piece is always stored for the piece, and high-quality encoding is possible.
[0034]
In addition, the encoded data includes a speech segment number having a similar partial speech segment, position data indicating which portion in the speech segment, data represented by parameters for amplitude adjustment, depending on circumstances. In addition, it can be expressed by data including spectral envelope parameters. Therefore, the encoded data is about several bytes of data, and significant data compression is possible. In general, since the voice hardly changes suddenly, when each processing target speech piece is considered to be about 4 msec, the change of the spectrum envelope parameter is gradual, and once in 10 processing target speech pieces. Extracting the spectrum envelope parameter at a certain frequency provides sufficient accuracy. Therefore, even if the spectrum envelope parameter is added, the data can be greatly compressed.
[0035]
DETAILED DESCRIPTION OF THE INVENTION
Embodiments of the present invention will be described below. Before describing a specific embodiment, first, the basic processing contents of the embodiment of the present invention will be described.
[0036]
FIG. 1 shows an input voice waveform. For example, a voice segment of about 4 msec is cut out from such an input voice waveform. The extracted speech segment (hereinafter referred to as the processing target speech segment) h1 is compared with the speech segment stored in the speech segment table, and the speech segment with the highest similarity is selected from the speech segment table and selected. Encoded data is created using the voice piece. The reason why the processing target speech piece is set to 4 msec is that the best result can be obtained by cutting out with a length of about 4 msec in the system used in this embodiment. That is, if the length of the processing target speech piece is shorter than 4 msec, the sound quality is improved, but it leads to a decrease in the compression rate, and if it is longer than 4 msec, the compression rate is advantageous. This is because the sound quality may be deteriorated.
[0037]
By the way, the voice segment table mentioned here has voice segments created from a plurality of elements as shown in FIG. 2 (in this example, four voice segments A1 to A4). The method of creating will be described later. Note that the latest speech segment is always stored in the speech segment table, and the speech segment table shown in FIG. 2 shows the contents of the speech segment table at a certain time.
[0038]
Now, assuming that the speech fragment table shown in FIG. 2 is the latest content, the processed speech fragment h1 of about 4 msec extracted in FIG. 1 is located at which part of which speech fragment in the speech fragment table. Determine whether they are most similar. In this case, it is determined that the processing target speech piece h1 is most similar in the portion from the position p1 of the speech piece A2 of the speech piece table. This most similar part is called a similar part.
[0039]
As a result, the encoded data of the processing target speech piece h1 can be represented by the scale for matching the speech piece number A2, the position p1, and the speech level of the speech piece table.
[0040]
That is, in this case, there are four voice segment numbers A1 to A4 in this case, so it can be represented by 2 bits, and the position p1 is 128 sampling if the length of each voice segment is 16 msec. Since it is a point (assuming the sampling frequency is 8 kHz), it can be represented by 7 bits. In order to adjust the sound level, for example, if it is adjusted in 128 steps, it can be represented by 7 bits. Therefore, when these are added up, it can be expressed as 16-bit, that is, 2-byte data.
[0041]
On the other hand, if the processing target speech piece h1 has a data amount of about 2 bytes at each sampling point, a data amount of 64 bytes exists if the number of sampling points is 32. Become. Therefore, the amount of data after encoding is 1/32 of the original data.
[0042]
If spectrum envelope parameters are used, about 4.5 bytes are required as the data. However, in general, since the voice hardly changes rapidly, when each processing target speech piece is considered to be about 4 msec, the change of the spectrum envelope parameter is moderate, and about once in 10 processing target speech pieces. Extracting the spectral envelope parameter at a frequency of sufficient accuracy can be obtained. Therefore, even if the spectral envelope parameter is added, the encoded data may be data that is significantly compressed relative to the original data. it can.
[0043]
As described above, in the present invention, the processing itself is simple, and the audio data can be efficiently compressed.
[0044]
Next, specific embodiments of the present invention will be described.
[0045]
FIG. 3 is a flowchart for explaining the processing procedure of the embodiment of the present invention. In FIG. 3, first, a processing target speech piece h1 of about 4 msec is cut out from the input speech (step s1). This process is the process described with reference to FIG. Then, it is determined whether or not a spectrum envelope parameter is to be extracted (step s2). If a spectrum envelope parameter is required, the spectrum envelope parameter is extracted (step s3). Note that, as described above, since the voice is unlikely to change abruptly, if each processing target speech piece to be cut out is considered to be about 4 msec, the change in the spectrum envelope parameter is moderate. Therefore, sufficient accuracy can be obtained by extracting the spectrum envelope parameters at a frequency of about once in 10 speech pieces to be processed.
[0046]
Then, in the next step s4, referring to the speech segment table at that time, the speech segment having the similar part with the highest similarity is selected. For example, if the content of the speech segment table at that time is the content shown in FIG. 2 for the processing target speech segment h1 at a certain time, the processing target speech segment h1 is the position of the speech segment A2 in the speech segment table. It is determined that the part from p1 is most similar, and the speech piece A2 is selected as a speech piece having a similar part.
[0047]
Next, an encoding process is performed based on the data (speech piece number, position, magnification for matching the speech level) of the selected speech piece A2 (step s5).
[0048]
Then, it is determined whether or not the compression process has been completed (step s6). If the compression process has been completed, the encoded data encoded in step s5 is output (step s7), and the input speech is input. It is determined whether or not all compression processes have been completed (step s8). If completed, the process ends. If not completed yet, the process returns to step s1.
[0049]
On the other hand, if the compression process is not finished in step s6, the decompression process (step s9) and the residual generation process (step s10) are performed, then the process returns to step s4, and the loop formed from step s4 to step s10 Process. Hereinafter, this loop process will be described.
[0050]
As described above, for example, it is determined that the portion from the position p1 of the speech segment A2 in the speech segment table is most similar to the processing target speech segment h1, and the speech segment A2 having the similar portion is selected. To do. Then, encoding processing is performed based on the data (speech number, position, magnification for matching the sound level) of the selected speech piece A2. The same process is repeated several times without terminating the compression process at this stage. That is, after being encoded in step s5, the encoded data is once decompressed (step s7), and then a residual generation process is performed (step 8).
[0051]
This residual generation process is a process of subtracting the encoded and expanded audio data from the original input audio (in this case, the processing target audio piece h1) and taking the difference. That is, as shown in FIG. 4, the decompressed audio data H1 is subtracted from the processing target audio fragment h1, and the residual d1 is obtained. Then, with respect to the obtained residual d1, a process of selecting a speech segment having a portion having a highest similarity (similar portion) with reference to the speech segment table at that time is performed. By performing such a process once or more, highly accurate compressed data can be obtained, but sufficient accuracy can be obtained even about twice.
[0052]
By the way, the decompression process performed in step s9 is performed according to a processing procedure as shown in the flowchart of FIG.
[0053]
That is, the encoded data is input (step s11), and it is determined whether or not the spectrum envelope parameter is updated (step s12). That is, when the spectrum envelope parameter is extracted, the value of the spectrum envelope parameter so far is updated to the value of the new spectrum envelope parameter (step s13).
[0054]
Next, referring to the speech segment table at that time, the speech segment having the portion with the highest similarity (similar portion) is selected based on the encoded data (step s14). Then, decompressed data is created based on the selected audio piece data (step s15). Then, it is determined whether or not the process is completed (step s16). If the processing is not finished, the contents of the speech segment table up to that point are updated with the new speech segment using the data expanded in step s15 (step s17).
[0055]
If there is more encoded data, the same processing is performed on the encoded data.
[0056]
Note that this decompression process is not only used as one of the processing procedures of FIG. 3, but is also used in the decompression process alone. For example, when encoded data is stored in a predetermined memory, it is also used when the encoded data is decompressed.
[0057]
When the decompression process ends in this way, residual generation is performed in the flowchart of FIG. 3 (step s10). In other words, as described above, as shown in FIG. 4, the decompressed audio data H1 is subtracted from the audio fragment h1, and the residual d1 is obtained. Then, with respect to the obtained residual d1, the speech segment having the highest similarity (similar portion) is selected with reference to the speech segment table at that time (the speech segment table newly updated after the decompression process). The process of doing. By performing such processing once or more, highly accurate compressed data can be obtained, but sufficient accuracy can be obtained even about twice as described above.
[0058]
By the way, the speech segment table used in the above processing includes at least speech segments created by the following elements.
[0059]
(1) Currently, audio data that has already been subjected to compression / expansion processing (voice data that has been subjected to compression / expansion processing later in time with respect to the processing target audio piece) is used for the cut out processing target audio piece. Here, the expression that the time that has already passed is referred to as “backward in time” and the future time is referred to as “forward in time” is used.
[0060]
For example, it is assumed that the input voice is shown in FIG. 6A, and the input voice up to a certain time t1 has already been compressed and expanded, and the compressed and expanded voice waveform is as shown in FIG. 6B. If the processing target speech piece h1 is currently at h1, the processing target speech piece h1 has a predetermined portion of the compressed and expanded speech waveform shown in FIG. 6B (immediately before the processing target speech piece h1). (Compressed and decompressed speech waveform) is used as a speech piece. This corresponds to, for example, A2 speech segment in the speech segment table shown in FIG. The time length of the audio piece is about 16 msec.
[0061]
(2) A temporal forward prediction speech waveform estimated from a spectral envelope parameter in the vicinity of the processing target speech segment and a temporal backward prediction speech waveform continuous therewith are used.
[0062]
As previously mentioned, the spectral envelope parameters do not need to be sent for each segmented speech fragment. This is because it is considered that the voice hardly changes abruptly. For example, if the spectrum envelope parameter is sent at a rate of once to several to a dozen speech pieces to be processed, Good. In this sense, the expression “nearby” spectral envelope parameters of the processing target speech segment is used here.
[0063]
Note that the temporal forward prediction speech waveform estimated from the spectral envelope parameters in the vicinity of the current speech piece to be processed and the temporal backward prediction speech waveform continuous therewith are shown in FIG. In addition to (speech waveform) x1, it indicates a backward predicted speech waveform x2 that is temporally backward.
[0064]
Thus, in addition to the impulse response (forward predicted speech waveform), using the backward predicted speech waveform that is temporally rearward has the effect of reducing noise. That is, when a voice piece using only the impulse response (forward predicted voice waveform) is used, a voice piece whose waveform rises suddenly from a state where the voice level is almost 0 is generated. When processed, there is a problem in that a discontinuous point appears and that part appears as noise. On the other hand, when the backward predicted speech waveform behind the time is used, the discontinuity can be reduced as much as possible, and the quality of the compressed and expanded speech can be greatly improved.
[0065]
(3) Use a noise waveform.
[0066]
This noise waveform may be given by random numbers, or may be sampled from actual input speech.
[0067]
As described above, the content of the speech segment table used in the present invention includes at least the speech segment described in (1) to (3). Each of these voice pieces is held as a voice piece having a length of about 16 msec in a state as shown in FIG. 2, for example, and the latest data is always stored.
[0068]
FIG. 8 is a block diagram showing the configuration of the audio compression / decompression apparatus of the present invention. In FIG. 8, the voice input from the voice input unit 1 is cut out by the voice cutout unit 2 as a processing target voice piece of about 4 msec, for example, as described above. The extracted processing target speech piece is compared with several speech pieces A1, A2,..., An in the speech piece table 4 by the similarity determination unit 3 to obtain a similarity. Then, the speech segment having the highest similarity (similar portion) is selected by the speech segment selection unit 5.
[0069]
The encoding unit 6 performs an encoding process based on data about the selected speech segment (speech segment number, position, magnification for matching the speech level) and the like. If the encoding process is terminated at this stage, the encoded data is output from the encoded data output unit 7. At this time, when the spectrum envelope parameter is used, an encoding process is performed by adding the spectrum envelope parameter extracted by the spectrum envelope parameter extraction unit 8.
[0070]
On the other hand, if the encoding process is not completed, the encoded data encoded by the encoding unit 7 is expanded by the expansion unit 9, and the residual generation process is performed by the residual generation unit 10. This expansion processing and residual generation processing are the processing of step s9 and step s10 in the flowchart in FIG.
[0071]
As described above, the residual generation process is a process of subtracting the encoded and expanded audio data from the original input audio (in this case, the processing target audio piece h1) and taking the difference. That is, as shown in FIG. 4, the decompressed voice data H1 is subtracted from the voice piece h1, and the residual d1 is obtained. Then, with respect to the obtained residual, a process of selecting a similar partial speech segment having the highest degree of similarity is performed with reference to the speech segment table at that time.
[0072]
The decompression process performed in the decompression unit 9 is performed according to the processing procedure shown in the flowchart of FIG. Then, the speech segment update unit 11 updates the contents of the speech segment table 4 using the decompressed speech data. In addition, when the spectral envelope parameter is extracted from the spectral envelope parameter extracting unit 8, the speech segment updating unit 11 performs temporal forward prediction speech waveform estimated by the spectral envelope parameter and temporal backward prediction continuous thereto. The voice waveform is also updated. In this way, the latest speech segment is always stored as the content of the speech segment table 4.
[0073]
Since the overall operation of the audio compression / decompression apparatus having such a configuration has been described with reference to the flowchart of FIG. 4, description of the operation is omitted here.
[0074]
The present invention is not limited to the above-described embodiment, and various modifications can be made without departing from the gist of the present invention. For example, the processing target speech piece to be cut out is 4 msec in the above-described embodiment. This is because the best result was obtained by using 4 msec in the system used in the above-described embodiment. . However, this numerical value may differ depending on the system to be used, and is not limited to this, and an optimal time can be set according to the system to which the present invention is applied. Further, the contents of the speech fragment table shown in FIG. 2 are merely examples, and the present invention is not limited to this.
[0075]
The processing program for performing the audio compression / decompression processing of the present invention described above can be stored in a storage medium such as a floppy disk, an optical disk, or a hard disk, and the present invention includes these storage mediums. Yes, and a method of obtaining data from the network may be used.
[0076]
As described above, according to the present invention, the similarity between each voice piece in the voice piece table and the voice piece in a predetermined section cut out from the input voice is compared, and the voice piece with the highest similarity is selected. The process of encoding the cut out predetermined section based on the data about the selected speech piece is performed as a basic process. As a result, encoding can be performed with a very simple process.
[0077]
Also, after creating encoded data, the encoded data is decompressed, and the residual waveform obtained by subtracting the decompressed data from the processing target speech fragment is referred to the speech fragment table again. Then, by obtaining the encoded data by performing the process of obtaining the similarity a plurality of times, it is possible to obtain even higher quality encoded data.
[0078]
The speech segment stored in the speech segment table is a speech segment created using a speech waveform that has already been compressed and decompressed in time behind the target speech segment, and is temporally estimated from the spectral envelope parameter. By having at least a speech segment created using the forward predicted speech waveform and the temporal backward predicted speech waveform, and a speech segment created from noise components, an efficient and high-quality code can be used when encoding the input speech. Can be realized. In particular, when a speech segment is created from a predicted speech waveform estimated from a spectrum envelope parameter, in the present invention, in addition to a temporal forward predicted speech waveform estimated from a spectrum envelope parameter, a temporally backward predicted speech waveform Therefore, noise can be reduced and the voice quality can be greatly improved.
[0079]
In addition, since the content of each speech piece is updated after the encoded data is decompressed or the spectral envelope parameters are extracted, a codebook having a fixed content as in the past is used. In contrast, the optimum speech segment is always stored for the speech segment to be processed, and high-quality encoding is possible.
[0080]
In addition, the encoded data includes a speech segment number having a similar partial speech segment, position data indicating which portion in the speech segment, data represented by parameters for amplitude adjustment, depending on circumstances. In addition, it can be expressed by data including a spectral envelope parameter, and a large amount of data compression is possible.
[0081]
As described above, according to the present invention, it is possible to perform high-quality audio compression / decompression with simple processing contents and high efficiency, and can be extremely advantageous when performing hardware or parallel processing.
[Brief description of the drawings]
FIG. 1 is a diagram showing an example in which input speech is cut out in a predetermined section in order to explain an embodiment of the present invention.
FIG. 2 is a diagram showing an example of an audio fragment table according to the embodiment of the present invention.
FIG. 3 is a flowchart for explaining a processing procedure according to the embodiment of the present invention.
FIG. 4 is a diagram illustrating processing for obtaining a residual component in the embodiment of the present invention.
FIG. 5 is a flowchart illustrating a decompression processing procedure according to the embodiment of the present invention.
FIG. 6 is a diagram for explaining an example of creating a speech segment in a speech segment table from an expanded speech waveform in the embodiment of the present invention.
FIG. 7 is a diagram illustrating an example in which a speech segment in a speech segment table according to the embodiment of the present invention is created from a temporally forward predicted speech waveform estimated from a spectral envelope parameter and a temporally backward predicted speech waveform.
FIG. 8 is a block diagram showing the configuration of the audio compression / decompression apparatus according to the embodiment of the present invention.
[Explanation of symbols]
1 Voice input part
2 Voice segmentation part
3 Similarity determination unit
4 Voice one side table
5 Voice segment selector
6 Coding section
7 Encoded data output section
8 Spectrum envelope parameter extraction unit
9 Extension part
10 Residual generator
11 Voice segment update part
h1 target speech piece
A1, A2, A3, A4 Voice fragment stored in the voice fragment table
p1 Position of similar partial speech in speech segment

Claims (12)

Cut out a segment of speech from the input speech
Extracting spectral envelope parameters from the segmented speech segment at a predetermined frequency,
With reference to a plurality of types of speech segments including a first speech segment created using a temporal forward prediction speech waveform estimated by the extracted spectral envelope parameter and a temporal backward prediction speech waveform continuous therewith, the plurality of speech segments Compare the similarity between the type of audio piece and the extracted audio piece of the predetermined section, select the audio piece with the highest similarity from the plurality of types of audio piece, and select the data about the selected audio piece An audio compression / decompression method characterized by including a process of generating encoded data by encoding an audio piece of the cut-out predetermined section. After creating the encoded data, the encoded data is expanded, and the expanded data is subtracted from the extracted audio segment of the predetermined section to obtain a residual. The encoded data is obtained by referring to types of speech segments and performing a process of comparing the similarity between the plurality of types of speech segments and the residual waveform. Audio compression encoding method.   The plurality of types of speech pieces are created by using a second speech piece and a noise component that are created by using the speech waveform that has been subjected to the decompression process that is temporally behind the speech piece of the cut out predetermined section. A third voice piece, the first voice piece is updated in content after extraction of the spectral envelope parameter, and the second voice piece is a code created using the second voice piece. 3. The audio compression / decompression method according to claim 2, wherein after the decompressed data is subjected to the decompression process, the content is updated based on the decompressed data.   The plurality of types of speech pieces have a section that is longer in time than the speech piece of the cut out predetermined section, and when comparing the similarity with the cut out voice piece of the predetermined section, 4. The speech piece having a portion with the highest similarity is selected by comparing the similarity with the speech piece of the cut-out predetermined section in the range of the length. The audio compression / decompression method described.   The encoded data is the number of a speech piece having the portion with the highest degree of similarity, position data indicating which part in the speech piece is, a parameter for amplitude adjustment, and, depending on the case, 5. The audio compression / decompression method according to claim 4, wherein the data is obtained by adding a spectrum envelope parameter. An audio segment extractor for extracting an audio segment of a predetermined section from the input speech;
A spectral envelope parameter extracting unit that extracts a spectral envelope parameter from a voice segment of a predetermined section cut out by the voice piece cutting unit at a predetermined frequency;
A plurality of types of speech pieces including a first speech segment created using a temporal forward prediction speech waveform estimated by the extracted spectral envelope parameter and a temporal backward prediction speech waveform continuous therewith,
A similarity determination unit that refers to the plurality of types of speech pieces and compares the similarities between the plurality of types of speech pieces and the extracted speech piece of the predetermined section;
An audio piece selection unit that selects an audio piece with the highest similarity based on the similarity by the similarity determination unit;
An encoding unit that encodes the speech segment of the predetermined section extracted based on the data about the speech segment selected by the speech segment selection unit;
An audio compression / decompression apparatus comprising: A decompression unit for decompressing the data encoded by the encoding unit;
A speech segment update unit that updates the content of the corresponding speech segment using the data expanded by the expansion unit or the spectrum envelope parameter;
7. The audio compression / decompression apparatus according to claim 6, further comprising: A residual generation unit that subtracts the data expanded by the expansion unit from the clipped audio segment of the predetermined section to obtain a residual;
The similarity determination unit, the speech segment selection unit, the encoding unit, the decompression unit, and the residual generation unit form a loop in the order of processing,
For the residual waveform generated by the residual generation unit, refer to the plurality of types of speech pieces, and perform a process of comparing the similarity with the plurality of types of speech pieces, and then encoding data. 8. The audio compression / decompression apparatus according to claim 7, wherein the audio compression / decompression apparatus is created and output. The plurality of types of speech pieces are created by using a second speech piece and a noise component created by using the speech waveform that has been subjected to the decompression process that is temporally rearward of the extracted speech piece of the predetermined section. A third voice piece
The speech segment update unit updates the first speech segment after the extraction of the spectral envelope parameter, and the encoded data created using the second speech segment is subjected to the decompression process. 9. The voice compression / decompression apparatus according to claim 6, wherein the second voice piece is updated based on the decompressed data. The plurality of types of speech pieces have a section that is longer in time than the cut-out speech section of the predetermined section,
The similarity determination unit calculates a similarity by comparing similarities in a range of lengths of the plurality of types of speech pieces;
10. The voice compression / decompression method according to claim 6, wherein the voice piece selection unit selects a voice piece having a portion having the highest degree of similarity.   The encoded data is data represented by the number of the voice segment having the highest part, the position data representing the highest part, and the parameter for amplitude adjustment. 11. The audio compression / decompression apparatus according to claim 10, further comprising: Cut out a segment of speech from the input speech
Extracting spectral envelope parameters from the segmented speech segment at a predetermined frequency,
With reference to a plurality of types of speech segments including a first speech segment created using a temporal forward prediction speech waveform estimated by the extracted spectral envelope parameter and a temporal backward prediction speech waveform continuous therewith, the plurality of speech segments The similarity between the type of speech piece and the speech piece in the predetermined section is selected, the speech piece having the highest similarity is selected, and the speech piece in the predetermined section is selected based on the data about the selected speech piece. A storage medium storing an audio compression / decompression processing program for causing a computer to execute a process of generating encoded data by encoding a computer.

Images