JP2009003078A - Speech encoding device, speech decoding device, speech encoding method, speech decoding method, and program - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To effectively encode and decode a speech signal which is digitalized with a sampling frequency suitable for recording conversation etc. for language learning. <P>SOLUTION: A vector quantization section 5 performs vector quantization on a spectrum handed over from a spectrum perfection section 31 by using a sorted vector quantization (VQ) table 41 or 43 in which a representative vector sorted in order of energy is stored, and an index attached to the representative vector used for approximation is handed over to an index difference calculation section 45 etc.. The index difference calculation section 45 hands over an index difference which is a value calculated by subtracting the index just before the time from the index handed over from the vector quantization section 5, to an entropy encoding section 35. The entropy encoding section 35 performs entropy encoding on the index difference. When decoding speech, the index difference is added to the index just before the time. <P>COPYRIGHT: (C)2009,JPO&INPIT

Description

  The present invention relates to a voice encoding device, a voice decoding device, a voice encoding method, a voice decoding method, and a program that are required when executing voice signal compression and decompression.

As encoding methods for audio signal compression used in cellular phones, digital audio players, etc., so far, μ-law, ADPCM (Adaptive Differential Pulse Code Modulation), MP3 (MPEG Audio Layer-3), VSELP ( Vector Sum Excited Linear Prediction), ITU-T Recommendation G. A CELP (Code-Excited Linear Prediction) type compression method represented by 729 has been put into practical use. Patent Document 1 discloses a technique using vector quantization as an audio signal compression technique.
Japanese Patent Laid-Open No. 10-63299

  In the case of digitizing speech for use in learning a foreign language, it is considered appropriate to sample the speech at a sampling frequency of about 16 kHz. This is because it is possible to maintain the characteristics of various languages at this level of sampling frequency, and from the viewpoint of ensuring the sound quality necessary for learning, the amount of data can be reduced even if the sampling frequency is increased further. This is because the effect is small for the increase.

  However, the compression noise that appears in the CELP compression method has a problem that it is not appropriate for language learning, although it does not interfere with communication between two speakers of the same native language through conversation. In addition, although μ-law and ADPCM enable voice reproduction with sound quality sufficient for language learning, since the encoding rate is high, these compression methods are used in portable devices where the storage capacity of the storage device is limited. When used, there is a problem that the recording capacity is reduced and the recording time is shortened. MP3 is intended for audio signal compression when higher quality audio reproduction is required than for language learning purposes, such as music appreciation purposes, and is effective at the above sampling frequency of about 16 kHz. There was a problem that could not be done.

  The present invention has been made in view of the above circumstances, and enables encoding and decoding suitable for low bit rate encoding of an audio signal digitized at a sampling frequency appropriate for recording such as language learning conversation. An object is to provide a speech encoding device, a speech decoding device, a speech encoding method, a speech decoding method, and a program.

In order to achieve the above object, a speech encoding apparatus according to the first aspect of the present invention provides:
A framing unit that divides the digital audio signal into framed digital audio signals that are digital audio signals for each frame that is a predetermined time interval;
A frequency converter that converts the frequency of the framed digital audio signal and generates a digital spectrum for each frame;
A vector quantization table composed of representative vectors identified by the attached index and having a large absolute value in ascending or descending order of the index;
A vector quantization unit that obtains the index corresponding to the digital spectrum by vector-quantizing the digital spectrum using the vector quantization table;
An index storage unit that stores the index obtained by the vector quantization unit in association with the frame corresponding to the index;
The index obtained by the vector quantization unit is acquired from the vector quantization unit, and is stored in the index storage unit in association with the frame that is temporally earlier than the frame corresponding to the index. An index difference calculation unit that acquires an index from the index storage unit and calculates a difference between the acquired indexes;
An encoding unit that generates a code by entropy encoding the difference calculated by the index difference calculation unit;
Is provided.

  Due to the continuity and stationarity of the audio signal, the difference value is biased, and therefore can be efficiently encoded by entropy encoding.

  The index difference calculation unit obtains, for example, the index obtained by the vector quantization unit from the vector quantization unit, and is associated with the frame temporally immediately preceding the frame corresponding to the index. An index stored in the index storage unit is acquired from the index storage unit, and a difference between the acquired indexes is calculated.

  The vector quantization table is composed of representative vectors expressed as one point in a multi-dimensional space using coordinate axes with element numbers corresponding to high and low frequencies, and there are a plurality of representative vectors having the same absolute value. When the plurality of representative vectors are arranged in ascending or descending order of the index, the element number assigned to the coordinate axis corresponding to the component having the maximum absolute value among the components corresponding to the coordinate axis of each representative vector is It is desirable that the index is attached so as to increase monotonously.

  According to the representative vector to which the index is attached in this way, since the index value is likely to be biased in the time zone in which the energy of the audio signal is substantially constant, the encoding efficiency is further improved.

  The vector quantization table includes a plurality of band-specific tables, and the band-specific table is associated with a table band, each of which is a specific band, and matches a typical speech spectrum pattern in the table band. The vector quantization unit vector-quantizes the digital spectrum for each quantization band that is the same as or more subdivided than the table band. When performing vector quantization for each quantization band, the band-specific table corresponding to the table band including the quantization band may be used.

  Since the audio signal has different characteristics for each band, it may be possible to efficiently perform vector quantization by referring to a different vector quantization table for each band.

  A code length monitoring unit that obtains the code length of the code generated by the encoding unit and determines whether the code length is equal to or less than a preset target code length, and the encoding unit includes the code If it is determined by the length monitoring unit that the code length is longer than the target code length, a portion corresponding to a relatively low energy deletion band of the digital spectrum divided into predetermined deletion bands It is desirable to perform entropy coding again after being excluded from the target of entropy coding.

  When there is a restriction on the amount of stored information and the amount of transmitted information, it is possible to minimize degradation of reproduced speech by encoding only the band considered to be important for speech decoding.

    For example, the frequency conversion unit performs a modified discrete cosine transform on the framed digital signal to generate the digital spectrum for each frame.

In order to achieve the above object, a speech decoding apparatus according to the second aspect of the present invention provides:
According to an index attached to a digital spectrum belonging to a current frame that is one of frames in a predetermined time interval and generated from a digital audio signal and to a digital spectrum belonging to a frame that is earlier in time than the current frame As a result of performing vector quantization using a vector quantization table composed of representative vectors that are distinguished from each other and that have a large absolute value in ascending or descending order of the index, the values obtained as values that characterize each digital spectrum are obtained. A decoding unit that decodes the difference from data entropy-encoded as an amount corresponding to the current frame in which the difference between the indexes calculated from the index is;
An index storage unit for storing the index in association with the frame corresponding to the index;
The index corresponding to the frame, the difference decoded as an amount corresponding to the frame by the decoding unit, the index stored in the index storage unit in association with a frame that is temporally earlier than the frame An index calculation unit obtained by adding to
The vector quantization table;
An inverse vector quantization unit that approximately restores the digital spectrum by using the representative vector retrieved from the vector quantization table by searching the representative vector corresponding to the index obtained by the index calculation unit;
A frequency inverse transform unit for performing frequency inverse transform on the digital spectrum restored by the inverse vector quantization unit to restore a digital audio signal belonging to the frame to which the digital spectrum belongs;
Is provided.

In order to achieve the above object, a speech encoding method according to a third aspect of the present invention includes:
A framing step of dividing the digital audio signal into framed digital audio signals that are digital audio signals for each frame that is a predetermined time interval;
A frequency conversion step of frequency-converting the framed digital audio signal to generate a digital spectrum for each frame;
The digital spectrum is vector-quantized using a vector quantization table that is configured by a representative vector that is identified by the attached index and that has a large absolute value in ascending or descending order of the index, thereby corresponding to the digital spectrum. A vector quantization step to find the index;
An index storing step for storing the index obtained by the vector quantum step in association with the frame corresponding to the index;
The index obtained by the vector quantization step is acquired from the vector quantization step and stored in the past index storing step in association with the previous frame in time than the frame corresponding to the index. An index difference calculating step for acquiring an index and calculating a difference between the acquired indexes;
An encoding step for generating a code by entropy encoding the difference calculated by the index difference calculating step;
Consists of

In order to achieve the above object, a speech decoding method according to the fourth aspect of the present invention provides:
According to an index attached to a digital spectrum belonging to a current frame that is one of frames in a predetermined time interval and generated from a digital audio signal and to a digital spectrum belonging to a frame that is earlier in time than the current frame From the index obtained as a value that characterizes each digital spectrum as a result of vector quantization using a vector quantization table that is identified and composed of representative vectors having a large absolute value in ascending or descending order of the index A decoding step of decoding the difference from the data entropy-encoded as an amount corresponding to the current frame in which the difference between the calculated indexes is;
An index storing step of storing the index in association with the frame corresponding to the index;
The index corresponding to the frame is stored in the past index storing step in which the difference decoded as an amount corresponding to the frame by the decoding step is correlated with a past frame in time than the frame. An index calculation step to be obtained by adding to
An inverse vector quantization step of approximately restoring the digital spectrum by using the representative vector searched from the vector quantization table for the representative vector corresponding to the index obtained by the index calculation step;
Frequency inverse transform step for restoring the digital audio signal belonging to the frame to which the digital spectrum belongs by performing frequency inverse transform on the digital spectrum restored in the inverse vector quantization step;
Is provided.

In order to achieve the above object, a program according to the fifth aspect of the present invention provides:
On the computer,
A framing step of dividing the digital audio signal into framed digital audio signals that are digital audio signals for each frame that is a predetermined time interval;
A frequency conversion step of frequency-converting the framed digital audio signal to generate a digital spectrum for each frame;
The digital spectrum is vector-quantized using a vector quantization table that is configured by a representative vector that is identified by the attached index and that has a large absolute value in ascending or descending order of the index, thereby corresponding to the digital spectrum. A vector quantization step to find the index;
An index storing step for storing the index obtained by the vector quantum step in association with the frame corresponding to the index;
The index obtained by the vector quantization step is acquired from the vector quantization step and stored in the past index storing step in association with the previous frame in time than the frame corresponding to the index. An index difference calculating step for acquiring an index and calculating a difference between the acquired indexes;
An encoding step for generating a code by entropy encoding the difference calculated by the index difference calculating step;
Is executed.

In order to achieve the above object, a program according to the sixth aspect of the present invention provides:
On the computer,
According to an index attached to a digital spectrum belonging to a current frame that is one of frames in a predetermined time interval and generated from a digital audio signal and to a digital spectrum belonging to a frame that is earlier in time than the current frame From the index obtained as a value that characterizes each digital spectrum as a result of vector quantization using a vector quantization table that is identified and composed of representative vectors having a large absolute value in ascending or descending order of the index A decoding step of decoding the difference from the data entropy-encoded as an amount corresponding to the current frame in which the difference between the calculated indexes is;
An index storing step of storing the index in association with the frame corresponding to the index;
The index corresponding to the frame is stored in the past index storing step in which the difference decoded as an amount corresponding to the frame by the decoding step is correlated with a past frame in time than the frame. An index calculation step to be obtained by adding to
An inverse vector quantization step of approximately restoring the digital spectrum by using the representative vector searched from the vector quantization table for the representative vector corresponding to the index obtained by the index calculation step;
Frequency inverse transform step for restoring the digital audio signal belonging to the frame to which the digital spectrum belongs by performing frequency inverse transform on the digital spectrum restored in the inverse vector quantization step;
Is executed.

  ADVANTAGE OF THE INVENTION According to this invention, the code | cord | chord handled at the time of compression and decompression | restoration of the audio | voice signal digitized with the sampling frequency appropriate for recording of language learning conversation etc. can be made into a thing with a low encoding rate.

  Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings.

(Embodiment 1)
First, a correspondence relationship between an audio signal as a time domain signal and an audio signal, ie, a spectrum, as a frequency domain signal will be described. In the following description, as a general rule, simply referring to an audio signal indicates an audio signal as a signal in the real time domain.

  An audio signal, which is a function of time, is basically equivalent to a spectrum time series representing the frequency characteristics of each time zone. There are various policies regarding what time zone the time axis is divided into, and how to obtain a spectrum representative of a certain time zone from an audio signal, and the present invention is applied to such various policies. can do. As an example, in the present embodiment, the concept of frame and block is introduced as a concept for dividing the time axis, and when obtaining the spectrum from the audio signal, MDCT (Modified Discrete Cosine Transform) in units of the block is used. ) And the above-described MDCT coefficient accumulation processing in units of frames are employed.

  FIG. 1 shows the relationship between frames, the relationship between blocks, and the relationship between frames and blocks. It is assumed that the horizontal direction of the paper is the time axis. As shown in the figure, the time axis is first divided into units having a predetermined time length called frames, and each frame is further divided into units having a shorter time length called blocks. However, the frames have overlapping portions corresponding to half the length of the blocks, and the blocks also have overlapping portions corresponding to the half length of the blocks. One frame includes L blocks from block 0 to block L-1. L is determined to be L = 4, for example, in consideration of the processing time required for vector quantization and the capacity of the VQ table, as will be described later. As shown in the figure, each frame has a time interval such as time t-2Δt, time t-Δt, time t, and time t + Δt, where Δt is the time corresponding to the start time or end time of two adjacent frames. Corresponds to each time for each Δt. In the present embodiment, since one spectrum is allocated to one frame, the spectrum is a spectrum at time t-2Δt, a spectrum at time t-Δt, a spectrum at time t, a spectrum at time t + Δt, and so on. In addition, a time series in which the time interval is Δt is formed.

  FIG. 2 schematically shows how to obtain the spectrum in the frame corresponding to time t. One block is a unit for performing one MDCT. Assuming that the number of audio signal samples included in one block is M, the result of MDCT is a total of M / 2 from the 0th order to the (M / 2-1) th order. The MDCT coefficient is obtained. In the case of an audio signal digitized at a sampling frequency of about 16 kHz, M is preferably 256, for example. The order of the MDCT coefficient corresponds to the frequency. That is, the higher-order MDCT coefficient corresponds to the high-frequency component of the audio signal. Therefore, it can be said that the graph in which the vertical axis is the MDCT coefficient and the horizontal axis is the order corresponds to the spectrum for each block, as shown in the figure.

The j (0 ≦ j ≦ M / 2-1) -th MDCT coefficient obtained as a result of MDCT performed in the block k (0 ≦ k ≦ L−1) included in the frame corresponding to the time t is _expressed as X _{t, j , K.} Then, j is an amount corresponding to the frequency. That is, the magnitude of j corresponds to the frequency level. Therefore, hereinafter, it may be expressed as “frequency j”.

In the frame corresponding to time t, L spectra for each block are obtained. That is, the block _{0, X t, 0,0, X} t, 1,0, ···, X t, Motomari spectral represented by _{M / 2-1,0,} for block 1, X _{t, 0 , 1} , X _{t, 1, 1} ,..., X _{t, M / 2-1 , 1} are obtained, and, for block L-1, X _{t, 0, L-1} , _{Xt, 1, L-1} ,..., _{Xt, M / 2-1, L-1} are obtained.

The L spectrums for each block are subjected to an integration process as indicated by a dotted arrow in FIG. 2 to finally complete one spectrum corresponding to time t. That is, the MDCT coefficients are represented by Xt _{, 0, 0} , _{Xt, 0} , ₁ , ..., _{Xt, 0, L-1} , _{Xt, 1} , ₀ , _{Xt, 1} , ₁ , ... , X _{t, 1, L-1} , X _{t, 2, 0} , ..., X _{t, M / 2-2, L-1} , X _{t, M / 2-1} , ₀ , X _{t, M / The} spectrum corresponding to the time t is completed by arranging like _{2-1, 1} ,..., _Xt , _{M / 2-1, L-1} .

Of the spectrum corresponding to the time t, the component corresponding to the j-th order MDCT coefficient is arranged in the time-series order of the blocks and expressed as a vector, and F _{t, j} . That is, _{Ft, j} = { _{Xt, j, 0} , _{Xt, j, 1} ,..., _{Xt, j, L-1} }. Further, the spectrum corresponding to time t is obtained by arranging vectors F _{t, j in the order} of F _{t, 0} , F _{t, 1} ,..., F _{t, M / 2-1} as shown in FIG. It can be said that there is.

Thus, the vector F _{t, j} corresponds to the frequency j which is a part of the spectrum corresponding to the time t. Hereinafter, the portion of the spectrum corresponding to the frequency j is referred to as a partial spectrum.

  The correspondence between the audio signal and the spectrum has been clarified. Next, configurations of the speech encoding device and the speech decoding device according to the present embodiment will be described. From the viewpoint of ensuring convenience for the user, in this embodiment, the speech encoding device and speech decoding device are integrated as a speech encoding / decoding device into a single device. .

  FIG. 3 shows a physical configuration of the speech encoding / decoding device 3 according to this embodiment. The voice encoding / decoding device 3 is, for example, a mobile phone.

  The speech encoding / decoding device 3 includes a CPU 121, a ROM (Read Only Memory) 123, a storage unit 125, a speech processing unit 141, a wireless communication unit 161, and an operation key input content processing unit 171. These are connected to each other via a system bus 181. The system bus 181 is a transmission path for transferring commands and data.

  The ROM 123 stores operation programs for speech encoding and decoding, representative vectors necessary for vector quantization, and the like.

  The storage unit 125 includes a RAM (Random Access Memory) 131 and a hard disk 133, and stores digital audio signals, MDCT coefficients, and the like. In particular, in this embodiment, the speech encoding / decoding device 111 requires information based on the speech signal of the immediately preceding time for processing at a certain time in both speech encoding and speech decoding. The storage unit 125 plays an important role as a delay processing buffer memory that stores such information at least temporarily.

  The audio encoding / decoding device 3 further includes a microphone 151, a speaker 153, an antenna 163, and operation keys 173.

  The microphone 151 collects the user's voice on the transmission side, that is, the encoding side, and delivers it to the voice processing unit 141. The speaker 153 issues the restored voice delivered from the voice processing unit 141 to the user on the receiving side, that is, the decoding side. The antenna 163 transmits the code delivered from the wireless communication unit 161 to the receiving side, that is, the decoding side speech encoding / decoding device 3 as a wireless signal, or the transmitting side, that is, the wireless signal transmitted from the encoding side device 3. Is received and delivered to the wireless communication unit 161. The operation key 173 is used when the user changes various initial setting values given in advance by his / her own judgment, or when the user on the transmission side, that is, the encoding side, selects the device 3 on the reception side, that is, the decoding side, which is the other party of the call. It is used to transmit the user's intention to the device 3 when specifying. For example, in the case of a mobile phone, this specification is performed using a telephone number assigned to each mobile phone.

  The voice processing unit 141, the wireless communication unit 161, and the operation key input content processing unit 171 are under the control of the CPU 121 via the system bus 181.

  FIG. 4 is a block diagram showing a functional configuration when the speech encoding / decoding device 3 according to the present embodiment functions as a speech encoding device. As shown in the figure, the speech encoding / decoding device 3 includes an A / D conversion unit 4, a DC (Direct Current) removal unit 23, a framing unit 25, a level adjustment unit 27, and a frequency conversion unit. 29, spectrum completion unit 31, vector quantization related processing unit 33, entropy encoding unit 35, code length monitoring unit 37, band data deletion unit 39, low band sorted VQ table 41, and high band sorted VQ A table 43 is provided.

  The CPU 121 in FIG. 3 functions as the A / D conversion unit 4 in FIG. 4 by operating in cooperation with the audio processing unit 141 and the storage unit 125 in accordance with the operation program written in the ROM 123. The CPU 121 also operates in cooperation with the storage unit 125 in accordance with the operation program written in the ROM 123, so that the DC removal unit 23, the framing unit 25, the level adjustment unit 27, the frequency conversion unit 29, the spectrum completion unit 31, the vector It functions as a quantization-related processing unit 33, an entropy encoding unit 35, a code length monitoring unit 37, and a band data deletion unit 39. The ROM 123 stores representative vectors necessary for vector quantization as a database, and functions as the low-frequency sorted VQ table 41 and the high-frequency sorted VQ table 43.

  The A / D converter 4 in FIG. 4 converts the input analog audio signal into a digital audio signal and outputs the digital audio signal to the DC removing unit 23. The sampling frequency is preferably about 16 kHz, but may be 11.025 kHz, 22.05 kHz, or the like.

  The DC removal unit 23 removes the direct current component of the digital audio signal input from the A / D conversion unit 4 and outputs it to the framing unit 25. The reason why the DC component of the audio signal is removed is that the DC component is almost irrelevant to the sound quality. The direct current component can be removed by, for example, a known high-pass filter.

  The framing unit 25 divides the signal input from the DC removal unit 23 into frames described with reference to FIGS. 1 and 2 and outputs the frame to the level adjustment unit 27. Basically, one frame is a processing unit for audio signal compression. However, in this embodiment, as will be described later, the processing in a certain frame requires the result of the processing in the immediately preceding frame in time, so in this sense, two frames are processing units for audio signal compression. It becomes.

The level adjustment unit 27 adjusts the level of the input audio signal for each frame, and outputs the level-adjusted signal to the frequency conversion unit 29. Level adjustment is to make the maximum value of the amplitude of a signal included in one frame fall within a specified number of bits (hereinafter referred to as suppression target bits). For example, if the maximum amplitude of a signal in one frame is n bits and the suppression target bit is N bits, the level adjustment is performed for all the signals in the frame by LSB (Least Significant Bit: least significant bit) for the number of shift_bits that satisfy the following equation: ) Can be realized by shifting to the side.
shift_bit = 0 (when n ≦ N), shift_bit = nN (when n> N)

  At the time of audio reproduction, since it is necessary to restore the signal whose amplitude is suppressed to the suppression target bit or less, it is necessary to output a signal representing shift_bit as a part of the audio compression signal. Therefore, the level adjustment unit 27 delivers the level-adjusted signal to the frequency conversion unit 29 and delivers the shift_bit to the entropy coding unit 35 for inclusion in the encoding target.

  The frequency conversion unit 29 performs frequency conversion on the signal input from the level adjustment unit 27 and outputs the signal to the spectrum completion unit 31. In the present embodiment, as described above, MDCT is used for frequency conversion. The frequency conversion unit 29 executes MDCT for each block described above, generates a spectrum for each block shown in FIG. 2, and delivers it to the spectrum completion unit 31.

The spectrum completion unit 31 in FIG. 4 first sorts the MDCT coefficients input from the frequency conversion unit 29 for each frequency. This is an operation for accumulating the spectrum for each block, as indicated by the dotted arrow in FIG. 2, to complete the spectrum corresponding to the frame. Subsequently, the spectrum completion unit 31 collectively vectorizes the coefficients in the same frequency band and outputs them to the vector quantization related processing unit 33. Here, the vector generated as a result of vectorization is the vector F _{t, j} = {X _{t, j, k} | k = 0, 1,..., L−1} already described with reference to FIG. It is.

  As described above, when signals in the same frequency band are collectively vectorized, for example, when many stationary signals are included, the accuracy of subsequent vector quantization is improved.

The vector quantization related processing unit 33 receives the vector F _{t, j} created by the spectrum completion unit 31, and performs the processing described later with reference to the low-frequency sorted VQ table 41 and the high-frequency sorted VQ table 43. An index difference is calculated, and the calculated index difference is delivered to the entropy encoding unit 35.

  As shown in FIG. 5, the vector quantization related processing unit 33 includes a vector quantization unit 5, a representative vector index storage unit 47, and an index difference calculation unit 45.

The vector quantization unit 5 refers to a VQ (Vector Quantization) table that stores representative vectors representing a plurality of speech patterns, and the vectors F _{t, j} created by the spectrum completion unit 31 and each of the vectors stored in the VQ table. The representative vectors are compared, the most similar representative vector is selected, and the index i attached to the representative vector is output to the index difference calculation unit 45 and the representative vector index storage unit 47.

Various criteria can be considered for selecting a representative vector similar to a vector to be vector-encoded. In the present embodiment, the representative vector is selected as follows. That is, i _MAX representative vectors stored in the VQ table are represented as {V _i | i = 1,..., I _MAX }, V _i = {v _{i, k} | k = 0,. 1}, the elements X _{t, j, k} of the vector F _{t, j} to be encoded are compared with the elements v _{i, k of} the i-th representative vector V _i stored in the VQ table. the vector F _{t, j} and V _i as the difference e _i is minimized between the vector V _i, is selected as the representative vector. The difference e _i is calculated by the following equation.
e _i = (X _{t, j, 0} -v _{i, 0} ) ^ 2 + (X _{t, j, 1} -v _{i, 1} ) ^ 2 + ... + (X _{t, j, k} -v _{i, k} ) ^ 2
However, the symbol “^” represents a power.

The number of representative vectors i _max and the number of blocks per frame, that is, the vector length L are determined in consideration of the processing time required for vector quantization, the capacity of the VQ table, and the like. For example, a free combination is conceivable, for example, the vector length L is 2 and the number of representative vectors is 256, or the vector length L is 4 and the number of representative vectors is 8192 (= 2 ¹³ ).

In the present embodiment, the representative vectors stored in the VQ table are indexed in ascending energy order. In other words, the representative vectors are sorted in order of energy, and if the energy of the representative vector V _i is E (V _i ),
E (V ₁ ) ≦ E (V ₂ ) ≦ ・ ・ ・ ≦ E (V _iMAX )
It is. Where energy E (V _i ) is
E (V _i ) = | V _i | ^{^ 2} = v _{i, 0} ^{^ 2} + v _{i, 1} ^{^ 2} + ... + v _{i, L-1} ^{^ 2}
Define as follows. In this embodiment, the VQ table storing the sorted representative vectors is used.

  In addition, since the sound often has different characteristics between the high frequency part and the low frequency part, in this embodiment, different VQ tables are used for the high frequency and the low frequency.

  Therefore, in the present embodiment, as the VQ table, the low-frequency sorted VQ table 41, which is a VQ table storing sorted representative vectors for use in low-frequency vector quantization, and the high-frequency vector. A high-frequency sorted VQ table 43, which is a VQ table storing sorted representative vectors for use in quantization, is used.

Vector F _{t, j} = {X _{t, j, k} | k = 0, 1,..., L−1} (j = 0, 1,..., M / 2) created by the spectrum completion unit 31 In (-1), the boundary between the high frequency band and the low frequency band may be, for example, a place where j indicating the frequency band is simply divided in half. That is, F _{t, 0} , F _{t, 1} ,..., F _{t, M / 4-1} are low frequencies, F _{t, M / 4} , F _{t, M / 4 + 1 ,. , M / 2-1} should be high. Therefore, in the vector quantization unit 5, the low-frequency vectors F _{t, 0} , F _{t, 1} ,..., F _{t, M / 4-1} are stored in the low-frequency sorted VQ table 41. Compared with the representative vector, the index i attached to the most similar representative vector is output. Similarly, the high-frequency vectors F _{t, M / 4} , F _{t, M / 4 + 1} ,..., F _{t, M / 2-1} are stored in the high frequency sorted VQ table 43. Compared with the representative vector, the index i attached to the most similar representative vector is output.

  Subsequent processing performed by the vector quantization related processing unit 33 includes some kind of delay processing such as using the result of vector quantization in the immediately preceding frame in terms of time. Therefore, in order to facilitate understanding, such processing will be described step by step with reference to FIG.

The representative vector index storage unit 47 receives the index i from the vector quantization unit 5 and stores it. The representative vector index storage unit 47 functions as a buffer memory for performing delay processing. When the vector quantization related processing unit 33 starts processing at time t, as shown in FIG. 6A, the representative vector index storage unit 47 obtains the result of processing at time t−Δt corresponding to the immediately preceding frame. As i (t-Δt, j) which is the index of the representative vector most similar to the vector F _{t-Δt, j} corresponding to the frequency j at time t-Δt. The process at time t is started when the vector Ft _{, j} is input to the vector quantization unit 5.

As shown in FIG. 6B, the vector quantization unit 5 obtains i (t, j) which is the index of the representative vector most similar to the input vector F _{t, j} .

  Next, as illustrated in FIG. 6C, the vector quantization unit 5 delivers the obtained i (t, j) to the index difference calculation unit 45 and the representative vector index storage unit 47. The representative vector index storage unit 47 receives and stores i (t, j) from the vector quantization unit 5, and passes i (t−Δt, j) stored so far to the index difference calculation unit 45.

Subsequently, as shown in FIG. 6D, an index difference calculation unit that receives i (t, j) from the vector quantization unit 5 and i (t−Δt, j) from the representative vector index storage unit 47. 45 is the index difference Δi (t, j),
Δi (t, j) = i (t, j) -i (t-Δt, j)
Ask for. Then, the index difference calculation unit 45 outputs the obtained index difference Δi (t, j) as shown in FIG. The output destination is the entropy encoding unit 35 as shown in FIGS.

  At the stage when the processing at time t is completed, as shown in FIG. 6E, the representative vector index storage unit 47 stores an index i (t, j) corresponding to the frequency j at time t. That is, except for the fact that the variable representing the time has changed from t-Δt to t, the state shown in FIG. Therefore, the processing similar to the processing from FIG. 6A to FIG. 6E may be repeated for each frame corresponding to time t + Δt and later after time.

  As shown in FIGS. 4 and 5, the entropy encoding unit 35 receives the shift_bit from the level adjustment unit 27 and the index difference Δi (t, j from the index difference calculation unit 45 in the vector quantization related processing unit 33. ), Entropy-encode these received quantities to generate a code, and output the generated code as an audio compression signal. The CPU 121 shown in FIG. 3 issues a transmission command to the wireless communication unit 161 based on the operation program stored in the ROM 123, and the wireless communication unit 161 accordingly receives the code via the antenna 163 by wireless communication. That is, it is performed by transmitting it to the speech encoding / decoding device 3 on the speech decoding side. The entropy encoding unit 35 also outputs the generated code to the code length monitoring unit 37. This is because it is necessary to determine whether or not the code length of the generated code satisfies a predetermined limit, as will be described later.

  Entropy coding is a coding method that uses the statistical properties of a signal to convert a code into a shorter code. Huffman coding, arithmetic coding, and range coder coding Is known. A feature of the entropy encoding method is that the compression rate is not constant even if the information compression accuracy is constant. That is, when entropy encoding is applied to a plurality of pieces of data having the same length but different contents, in general, due to the difference in the appearance frequency of data elements in the original data, The code length varies. In general, it is difficult to predict the compression rate before encoding, and it is not known until entropy encoding is actually performed whether or not a high compression rate can be obtained. On the other hand, the present invention aims at encoding at a low bit rate, and if the speech encoding / decoding device 3 is, for example, a mobile phone, the code length is limited due to communication infrastructure or the like. . In principle, the entropy encoding unit 35 entropy-encodes all the information received from the level adjustment unit 27 and the vector quantization related processing unit 33 in order to minimize degradation of speech quality. When the entropy encoding unit 35 actually performs entropy encoding, if the compression rate happens to be low enough not to satisfy the above-mentioned restriction on the code length, the information to be encoded is appropriately selected. It is necessary to thin out and re-encode.

  Therefore, in this embodiment, a code length monitoring unit 37 and a band data deleting unit 39 are provided. The code length monitoring unit 37 receives the code generated by the entropy encoding unit 35, measures the code length, and monitors whether the code length is within a predetermined target code length. If the code length monitoring unit 37 determines that the target code length has been exceeded as a result of such monitoring, the code length monitoring unit 37 notifies the band data deletion unit 39 to that effect. When the band data deleting unit 39 receives a notification that the code length is too long, the band data deleting unit 39 maintains the sound quality even if the MDCT coefficient at j is deleted from the frequency band excluded from the encoding target, specifically, the frequency j. J that is considered to have relatively little influence on the point is determined, and the determination result is notified to the entropy encoding unit 35. Upon receiving such notification, the entropy encoding unit 35 excludes the band to be deleted determined by the band data deletion unit 39 from the encoding target, and redoes the entropy encoding. The code generated again is monitored again by the code length monitoring unit 37. If the code length is still too long, more band to be excluded from the encoding target is determined by the band data deleting unit 39, This is fed back to the entropy encoding unit 35. Such loop processing is repeated until the code length of the code generated by the entropy encoding unit 35 becomes equal to or less than the target code length.

As described above, the band data deleting unit 39 selects a band that has little influence even if it is deleted in terms of preventing voice quality deterioration from among the bands corresponding to each frequency j. Various criteria can be considered for determining a band that has little influence even if it is deleted. In the present embodiment, a band having a small energy is deleted. In this way, the band to be deleted can be determined relatively easily. That is, as energy at frequency j, energy E (F _{t, j} ) is
E (F _{t, j} ) = | F _{t, j} | ^{^ 2} = X _{t, j, 0} ^{^ 2} + X _{t, j, 1} ^{^ 2} + ... + X _{t, j, L-1} ^{^ 2}
The band corresponding to the frequency j having a small energy E (F _{t, j} ) is preferentially deleted. Note that the deletion of the band corresponding to the frequency j is specifically performed by replacing all components of the vector _{Ft, j} with 0, for example.

  The procedure of the above operations performed by the information amount monitoring unit 37 and the band data deleting unit 39 is summarized in the flowchart shown in FIG. Physically, as described above, the CPU 121 functions as the information amount monitoring unit 37 and the band data deleting unit 39 by operating in cooperation with the storage unit 125 according to the operation program written in the ROM 123.

It is assumed that MDCT in the frame corresponding to time t is completed, and the vector F _{t, j} (0 ≦ j ≦ M / 2-1) is already stored in the storage unit 125.

The CPU 121 loads the vector F _{t, j} (0 ≦ j ≦ M / 2-1) from the storage unit 125 to an internal register (not shown) of the CPU, and calculates the energy E (F _{t, j} ) at the frequency j. Then, sorting is performed based on the calculated E (F _{t, j} ), and a priority is assigned to each frequency j so that a band with low energy is preferentially deleted (step S7). However, at first, the entire spectrum band is to be encoded (step S11). Subsequently, the CPU 121 entropy-encodes data to be encoded to generate a code (step S13), and obtains a code length (step S15). Further, the CPU 121 determines whether or not the obtained code length is equal to or smaller than a predetermined target code length (step S17). When it is determined that the calculated code length is equal to or smaller than the target code length (step S17; Yes), the process is terminated. When it is determined that the target code length is exceeded (step S17; No), the process proceeds to step S19. In step S19, the CPU 121 sets the remaining band except for the band with the highest deletion priority among the bands that have been encoded at the time of the previous entropy encoding (step S13) as a new encoding target. (Step S19), the process returns to step S13 to perform entropy encoding again.

  In this way, even if it is unavoidable to exclude some bands from the encoding target, the bands that are considered to have a large impact on the reproduced speech quality due to their relatively high energy are considered as encoding targets. Expected to remain. Therefore, it is possible to suppress the deterioration of the reproduced voice quality due to the band deletion to the minimum.

Even if the value of E (F _{t, j} ) over various frequencies j is considered after fixing time _t, the value of E (F _{t, j} ) over a sufficiently long time with frequency j fixed However, E (F _{t, j} ) itself has various values, large and small. However, the value of E (F _{t, j} ) -E (F _{t-Δt, j} ), which is the energy difference between adjacent times, is sufficient for a sufficiently long time, even if considered over various frequencies j. However, relatively small values appear frequently. This is due to the fact that the audio signal is continuous and that a steady state often appears in the audio signal.

As shown in FIG. 6 and the like, the vector quantization unit 5 causes the vector F _{t, j} to be the representative vector V _{i (t, j)} , the vector F _{t-Δt, j} is the representative vector V _{i (t-Δt, j)} , respectively. As described above, since the representative vectors are sorted in order of energy, the index assigned to the representative vector already has a meaning as an index of energy of the representative vector, and E (F _{As the value of t, j} ) -E (F _{t-Δt, j} ), considering that a relatively small value appears frequently, it is attached to two representative vectors adjacent in time series. The index difference Δi (t, j) (= i (t, j) −i (t−Δt, j)) output from the index difference calculation unit 45 is a relatively small value. It is concluded that appears frequently.

  In general, when the value to be encoded is biased, the compression efficiency of entropy encoding is improved. Therefore, according to the present embodiment, the index difference Δi (t, j), which is an amount that becomes a relatively small value with high frequency, is entropy-encoded, so that the encoding efficiency is high and the code length is short. A high-quality audio signal can be transmitted with a code.

  Therefore, according to the present embodiment, it is possible to reduce the encoding rate while adopting a sampling frequency suitable for recording such as language learning conversation. For example, the audio encoding / decoding device 111 according to the present embodiment can compress an audio signal having a sampling frequency of about 16 kHz to a rate of about 16 kbps.

  In the above description, the process is described in which the difference value of the index added to the representative vector between time t and time t−Δt is sent from the speech coding apparatus to the speech decoding apparatus. In order to enable audio reproduction by such processing, it is natural that, at least for the first frame to be encoded, the value of the index itself must be sent as an initial value from the former device to the latter device. There is. Therefore, in the present embodiment, for a frame corresponding to the time when the user of the speech coding apparatus starts speaking, such an initial value is sent to the speech decoding apparatus. Furthermore, if only the difference is kept sent, errors due to electrical errors at the time of reception may accumulate on the audio decoding side, and the audio may not be reproduced correctly. A value is to be sent.

  FIG. 8 is a block diagram showing a functional configuration when the speech encoding / decoding device 3 according to the present embodiment functions as a speech decoding device. As shown in the figure, the speech encoding / decoding device 3 includes an entropy decoding unit 8, a vector inverse quantization related processing unit 49, a time order rearrangement unit 51, a frequency inverse transformation unit 53, and a level reproduction unit. 55, a frame synthesizing unit 57, and a D / A converting unit 59, and further includes a low-frequency sorted VQ table 41 and a high-frequency sorted VQ table 43 that function as a speech encoding device.

  3 operates in cooperation with the storage unit 125 in accordance with the operation program written in the ROM 123, so that the entropy decoding unit 8, the vector inverse quantization related processing unit 49, the time order rearrangement unit 51 in FIG. It functions as a frequency inverse transform unit 53, a level reproduction unit 55, and a frame synthesis unit 57. The CPU 121 also functions as the D / A conversion unit 59 of FIG. 8 by operating in cooperation with the audio processing unit 141 and the storage unit 125 according to the operation program written in the ROM 123.

  When the speech encoding / decoding device 3 according to the present embodiment operates as a speech decoding device, information transmitted by means such as wireless communication as a result of another speech encoding / decoding device 3 operating as the speech encoding device. Are compressed by the antenna 163. The wireless communication unit 161 stores information collected by the antenna 163 in the storage unit 125 in accordance with an instruction issued by the CPU 121 based on an operation program stored in the ROM 123.

  The entropy decoding unit 8 decodes an audio compression signal that is a signal encoded by entropy encoding. Subsequently, the entropy decoding unit 8 outputs the index difference Δi (t, j) of the information obtained as a result of the decoding to the vector inverse quantization related processing unit 49, and among the information, shift_bit is converted to the level reproduction unit. To 55. The vector inverse quantization related processing unit 49 receives the index difference Δi (t, j) from the entropy decoding unit 8 and refers to the low frequency sorted VQ table 41 and the high frequency sorted VQ table 43 to be described later. Thus, an appropriate representative vector is selected, and the selected representative vector is delivered to the time order rearrangement unit 51 as an amount used for approximate restoration of the spectrum.

  As shown in FIG. 9, the vector inverse quantization related processing unit 49 includes an index calculation unit 9, a representative vector index storage unit 61, and a vector inverse quantization unit 63.

  The index calculation unit 9 exchanges information with the representative vector index storage unit 61 as will be described later, so that the speech encoding / decoding device 3 as the speech encoding device is obtained by vector quantization at the frequency j. The index attached to the representative vector is calculated and delivered to the vector inverse quantization unit 63.

  The vector inverse quantization unit 63 acquires the representative vector with the index delivered from the index calculation unit by searching the low-frequency sorted VQ table 41 or the high-frequency sorted VQ table 43. . When the frequency j to be processed corresponds to the previously defined low frequency, the low frequency sorted VQ table 41 corresponds to the frequency j, and the frequency j corresponds to the high frequency. Are searched in the high frequency sorted VQ table 43. The vector inverse quantization unit 63 outputs the retrieved representative vector to the time order rearrangement unit 51 as a result of approximating the portion of the band corresponding to the frequency j in the spectrum for each frame.

  The index calculation unit 9 and the representative vector index storage unit 61 store the index attached to the representative vector obtained by the vector quantization at the frequency j at the time t by the speech encoding / decoding device 3 as the speech encoding device. When the calculation is performed by exchanging information in (3), some kind of delay processing is performed, such as using the result of vector quantization in the immediately preceding frame in terms of time. Therefore, in order to facilitate understanding, such processing will be described step by step with reference to FIGS. 10 and 11.

The representative vector index storage unit 61 receives the index i from the index calculation unit 9 and stores it. The representative vector index storage unit 61 functions as a buffer memory for performing delay processing. When the vector inverse quantization related processing unit 49 starts processing at time t, as shown in FIG. 10A, the representative vector index storage unit 61 performs processing at time t−Δt corresponding to the immediately preceding frame. As a result, i (t-Δt, j) which is the index of the representative vector most similar to the vector F _{t-Δt, j} corresponding to the frequency j at time t-Δt is stored. The processing at time t is started when the index difference Δi (t, j) is input to the index calculation unit 9.

As shown in FIG. 10B, the representative vector index storage unit 61 delivers the stored index i (t−Δt, j) to the index calculation unit 9. Next, as shown in FIG. 10C, the index calculation unit 9 uses the index difference Δi (t, j) input from the entropy decoding unit 8 as the index i ( By adding to t−Δt, j), an index i (t, j) at time t is obtained. That is, the index calculation unit 9 performs an operation of i (t, j) = i (t−Δt, j) + Δi (t, j). Subsequently, as shown in FIG. 11A, the index calculation unit 9 delivers the obtained index i (t, j) to the vector inverse quantization unit 63 and the representative vector index storage unit 61. Thereafter, as shown in FIG. 11B, the vector inverse quantization unit 63 searches the VQ table for the representative vector to which the received index i (t, j) is attached, while the representative vector index. The storage unit 61 stores the received index i (t, j) itself. Finally, as shown in FIG. 11C, the vector inverse quantization unit 63 outputs the retrieved representative vector V _{i (t, j)} to the time order rearrangement unit 51. At this stage, the representative vector index storage unit 61 stores an index i (t, j) corresponding to the frequency j at time t. That is, except for the fact that the variable representing the time has changed from t-Δt to t, the state shown in FIG. Therefore, the processing similar to the processing from FIG. 10A to FIG. 11C may be repeated for each frame corresponding to time t + Δt and later after time.

  The time order rearrangement unit 51 in FIGS. 8 and 9 approximately reproduces the spectrum by collecting representative vectors corresponding to each frequency j from the vector inverse quantization unit 63 in the vector inverse quantization related processing unit 49. Then, the spectrum for each block is approximately reproduced by rearranging the components so as to follow the dotted arrow in FIG. 2 in the reverse direction. Subsequently, the time-order rearranging unit 51 delivers the spectrum for each block to the frequency inverse transform unit 53 in FIG. The frequency inverse transform unit 53 performs inverse MDCT on the spectrum for each block input from the time order rearrangement unit 51 and outputs the result to the level reproduction unit 55. The level reproduction unit 55 adjusts the level of the signal input from the frequency inverse conversion unit 53 by referring to the shift_bit input from the entropy decoding unit 8 to return to the original level. Output. The frame synthesizing unit 57 synthesizes frames that are processing units of encoding and decoding, and outputs the synthesized signal to the D / A conversion unit 59. The D / A converter 59 converts the digital signal input from the frame synthesizer 57 into an analog signal and outputs it as an audio reproduction signal.

(Embodiment 2)
Hereinafter, a speech encoding / decoding device according to Embodiment 2 of the present invention will be described. In this embodiment, the low frequency sorted VQ table 41 and the high frequency sorted VQ table 43 stored as a database in the ROM 123 of the speech encoding and decoding apparatus 3 according to the first embodiment are encoded efficiently. It is the same as that of Embodiment 1 except for further improvements.

In the first embodiment, for the index i (1 ≦ i ≦ i _MAX ) attached to the representative vectors stored in the low frequency sorted VQ table 41 and the high frequency sorted VQ table 43, the representative vector V _i Energy E (V _i )
E (V ₁ ) ≦ E (V ₂ ) ≦ ・ ・ ・ ≦ E (V _iMAX )
It was imposed that the relationship was established. However, when a plurality of representative vectors happen to have the same energy, there is no particular limitation on how to index each of the plurality of representative vectors.

  On the other hand, in the present embodiment, even when a plurality of representative vectors happen to have the same energy so as to further improve the encoding efficiency, the plurality of such vectors are determined according to a predetermined policy in consideration of the continuity of the audio signal. Are indexed.

That is, p representative vectors with equal energy
_{V i1 = {v i1,0, v} i1,1, ···, v i1, k [i1, MAX], ···, v i1, L-1},
_{V i2 = {v i2,0, v} i2,1, ···, v i2, k [i2, MAX], ···, v i2, L-1},
...
V _ip = {v _{ip, 0} , v _{ip, 1} , ..., v _{ip, k [ip, MAX]} , ..., v _{ip, L-1} }
(Where v _{i, k [i, MAX]} represent the component having the maximum absolute value among the components of V _i )
When attaching an index such that i1 <i2 <... <ip,
k [i1, MAX] ≦ k [i2, MAX] ≦ ・ ・ ・ ≦ k [ip, MAX]
Is established.

  Hereinafter, in order to facilitate understanding, what the predetermined policy is is described with an example.

For example, based on the constraint E (V ₁ ) ≦ E (V ₂ ) ≦ ・ ・ ・ ≦ E (V _iMAX ), the representative vector having the smallest energy to the 14th in ascending order of energy Suppose that the index to be attached is fixed. That is,
E (V ₁ ) <E (V ₂ ) <・ ・ ・ <E (V ₁₄ )
And, E (V ₁₄₎ is V _1, V _2, ···, smaller than the energy of any representative vectors than V _14, to. Furthermore, it is also assumed that the index to be assigned is determined based on the above-mentioned restrictions for the 19th representative vector to the representative vector having the maximum energy in ascending order of energy. That is,
E (V ₁₉ ) <E (V ₂₀ ) <・ ・ ・ <E (V _iMAX )
It is also assumed that E (V ₁₉ ) is larger than the energy of any representative vector other than V ₁₉ , V _{20 ,.} In addition, the four representative vectors remaining when V ₁ , V ₂ ,..., V ₁₄ , V ₁₉ , V _{20 ,.} And Then, although it is determined that the four representative vectors from 15 to 18 should be assigned to the four representative vectors, 15 is assigned as an index to which representative vector, 16 is assigned to which representative vector, In the first embodiment, which representative vector is assigned 18 is arbitrary. Since the constraint is due to an inequality sign, no matter how the four representative vectors are indexed, E (V ₁₅ ) = E (V ₁₆ ) = E (V ₁₇ ) = E (V ₁₈ This is because the above constraints are satisfied.

  On the other hand, in this embodiment, an index is attached | subjected according to the above-mentioned policy also with respect to several representative vectors which have such an equal energy.

  Hereinafter, as an example, it is assumed that the number L of blocks per frame is L = 5. Then, each representative vector has L, that is, five components.

Therefore, as shown in FIG. ^{_{12, | V 15 | 2 =}} | V 16 | 2 = | V 17 | 2 = | V 18 | 4 pieces of representative vectors V _15, V ₁₆ with equal energy as ^2, V ₁₇ and V ₁₈ can be expressed in a five-dimensional space provided with five coordinate axes from the 0th axis to the 4th axis. The zeroth axis corresponds to block 0, the first axis corresponds to block 1, the second axis corresponds to block 2, the third axis corresponds to block 3, and the fourth axis corresponds to block 4. When R is equal to the absolute value of the above four representative vectors, that is, R = | V ₁₅ | = | V ₁₆ | = | V ₁₇ | = | V ₁₈ |, the tip of the representative vector is 5 It exists on a spherical surface with a radius R centered on the origin in the dimensional space.

In the following, as an example, the component v ₁₅ , ₀ , v ₁₅ , ₁ , v ₁₅ , ₂ , v ₁₅ , ₃ , v ₁₅ , ₄ of the representative vector V ₁₅ has the largest absolute value v ₁₅ , ₀ , and the component v _{16, 0} representative vectors _{V 16, v 16,1, v 16,2} , v 16,3, absolute value of v _16,4 is v _16,1 up components, components v _{17, 0} representative vectors _{V 17, v 17,1, v 17,2} , v 17,3, absolute value of v _17,4 biggest component is v _17,2, representative vector V ₁₈ components _{v 18,0, v 18,1, v 18,2} , v 18,3, absolute value of v _18,4 biggest component is assumed to be v _18,3.

This indexing of each representative vector is consistent with the above-described predetermined policy introduced in the present embodiment. Also, from the viewpoint of the relationship with the spectrum represented by the MDCT coefficient, it can be said that the number assigned to the coordinate axis corresponds to the frequency level. Furthermore, according to FIG. 12, qualitatively, V ₁₅ makes a small angle with the 0th axis, V ₁₆ makes a small angle with the first axis, V ₁₇ makes a small angle with the second axis, V ₁₈ makes a small angle with the third axis.

Hereinafter, as an example, when each representative vector is displayed as a partial spectrum, each vector is assumed to have a shape shown in FIG. That is, the bar graphs drawn with solid lines in FIGS. 13A, 13C, 13D, and 13E are partial spectra corresponding to the representative vectors V ₁₆ , V ₁₅ , V ₁₇ , and V ₁₈ , respectively. . Components v _{16, 0} of as V _{16 above, v 16,1, v 16,2, v} 16,3, since the absolute value of v _16,4 biggest component is the v _16,1, FIG as shown in 13 (a), in the partial spectrum corresponding to V _16, the frequency corresponding to the block 1 reaches a peak frequency.

In FIG. 13 (b) ~ (e) , for reference, partial spectrum corresponding to the representative vector V ₁₆ is shown in dotted lines.

  Here, in order to prevent confusion in understanding, terms are organized. As described above, the partial spectrum is a portion corresponding to the frequency j in the entire spectrum. The peak frequency here refers to a frequency that causes a peak in the partial spectrum, among frequencies obtained by further subdividing the frequency j band.

  As is clear from FIG. 2, in the present embodiment, strictly, the arrangement of the MDCT coefficients in the partial spectrum is in block order, that is, in time order, not in frequency order. However, as shown by the dotted arrows in FIG. 2, a partial spectrum is constructed by considering the time order as the frequency order, and the spectrum corresponding to one frame is completed by arranging such partial spectra in the order of the frequency j. Therefore, when discussing the approximation by the representative vector, it is assumed that the arrangement of the MDCT coefficients in the partial spectrum is in order of frequency. This is simply because, in this embodiment, MDCT in units of blocks is adopted as frequency conversion after introducing a plurality of time sections of frames and blocks. For example, one block corresponds to one frame, that is, MDCT is applied to the audio signal of the entire frame, or the concept of time division different from that of the frame or block is introduced, or another frequency conversion is adopted. In this case, temporal elements may be excluded from the arrangement of the frequency conversion coefficients. The arrangement of the MDCT coefficients in the partial spectrum, which is a unit for performing vector quantization, may be in time order as in the present embodiment. Since this embodiment effectively uses the continuity of the audio signal on the time axis and the frequency axis, even if the time series is handled in part of the spectrum generation process, it is based on a consistent policy. This is because the effect of the present invention is produced as long as the spectrum generated in this way is handled.

As described above, the partial spectrum corresponding to the representative vector V ₁₆ is shown in FIG. For other representative vectors, the peak frequency of the V ₁₅ is the frequency corresponding to the block 0 is a peak frequency, the peak frequency of the V ₁₇ is the frequency corresponding to the block 2 becomes the peak frequency, the peak frequency of V ₁₈ to block 3 The corresponding frequency becomes the peak frequency. Since the peak frequency becomes higher in the ascending order of the index in this way, in order to facilitate understanding, in the following, as shown in FIGS. 13 (c), (d), and (e), representative vectors V ₁₅ , V ₁₇ , V ₁₈ happens to correspond to a partial spectrum corresponding to the representative vector V ₁₆ that is shifted in the frequency axis direction by the index difference. For example, it is assumed that the partial spectrum corresponding to V ₁₇ is obtained by shifting the partial spectrum corresponding to V ₁₆ by one block in the high frequency direction. At this time, the MDCT coefficient originally corresponding to the highest frequency of V ₁₆ , that is, the block 4 is made to correspond to the lowest frequency, that is, the block 0 in V ₁₇ . Considering only such a case, there is no problem in describing the features of the present embodiment because the difference between the peak frequency and the index matches. Further, in the present embodiment, there is a problem when a plurality of representative vectors have the same energy, and in view of this point, when each representative vector is obtained by parallel movement in the frequency axis direction as described above, Since it is obvious that the energy of all the representative vectors is equal, it is convenient for understanding.

Hereinafter, vector quantization performed at time t and time t + Δt for a certain frequency j will be described. Assume that the representative vector V ₁₆ is selected as the representative vector most similar to the vector F _{t, j} representing the partial spectrum at time t. That is, assume that F _{t, j} is approximated by V ₁₆ . Various cases can be considered as to which representative vector the vector F _{t + Δt, j} representing the partial spectrum at the next time is approximated. For example, it may be approximated by a representative vector having higher energy than the representative vectors V ₁₅ , V ₁₆ , V ₁₇ and V ₁₈ , or may be approximated by a representative vector having lower energy. As described above, when approximated by a representative vector having energy different from that of the representative vector used for approximation at the immediately preceding time, an effect specific to the present embodiment does not occur, and only the same effect as in the first embodiment occurs. .

On the other hand, the vector F _{t + Δt, j} representing the partial spectrum at the next time may be approximated by any of the representative vectors V ₁₅ , V ₁₆ , V ₁₇ , V ₁₈ . This is not a rare situation, but rather a frequent occurrence, due to the continuity of the audio signal and the continuity, ie the nature of the audio signal often having a steady-state time zone. is there. Then, among them, there is a high possibility that V ₁₆ is the same representative vector the representative vector that has been selected at time t is the time of the immediately preceding is selected again. That is, F _{t + Δt, j} is the same as or almost the same as F _{t, j,} and there is a high possibility that the same vector is selected as the representative vector for approximation. This is due to the continuity and stationarity described above. In such a case, the index difference becomes 0 due to 16-16.

The next most likely is that F _{t + Δt, j} is approximated by V ₁₅ or V ₁₇ . Because of the continuity and stationarity described above, the partial spectrum corresponding to F _{t + Δt, j} changes more significantly when the partial spectrum corresponding to F _{t, j} is slightly changed. This is because there are more cases. Referring to FIG. 13, partial spectrum corresponding to V ₁₅ or V _17, since the frequency Bundake peak corresponding to one block in comparison with the partial spectrum corresponding to V ₁₆ are obtained by moving, V ₁₆ can be said to have changed slightly. On the other hand, since the partial spectrum corresponding to V ₁₈ has a peak shifted by the frequency corresponding to two blocks compared to the partial spectrum corresponding to V ₁₆ , relatively, V ₁₆ changed greatly. It can be said that it is a thing. Therefore, the possibility that F _{t + Δt, j} is approximated by V ₁₅ or V ₁₇ is higher than the possibility that F _{t + Δt, j} is approximated by V ₁₈ . If F _{t + Δt, j} is approximated by V _15, the index of the difference, by 15-16, -1. When F _{t + Δt, j} is approximated by V ₁₇ , the index difference becomes 1 due to 17-16. When F _{t + Δt, j} is approximated by V ₁₈ , the index difference is 2 due to 18-16.

  As is clear from the above, when the energy of the representative vector used for approximation at time t + Δt is equal to the energy of the representative vector used for approximation at time t, the difference between the indexes of both representative vectors is the present embodiment. As long as the index is attached to the representative vector according to the above-mentioned constraints newly introduced in, the probability of becoming 0 is the highest, followed by the probability of becoming +1 and the probability of becoming -1, followed by +2. And the probability of becoming -2, and so on. That is, the index difference is biased such that the smaller the absolute value, the higher the appearance frequency. Due to the presence of such bias, the efficiency of encoding by entropy encoding is increased.

  Thus, according to the present embodiment, in the case where the energy of the spectrum is substantially constant over a plurality of consecutive frames, in addition to the high efficiency of encoding for the same reason as in the first embodiment. It can be expected that the encoding efficiency is further increased. In addition, because of the continuity and continuity of the audio signal, it frequently occurs when the energy of the spectrum is almost constant over a plurality of consecutive frames. Therefore, there are many cases where an effect peculiar to the present embodiment occurs.

  In addition, this invention is not limited to the said embodiment, A various deformation | transformation and application are possible. The above-described hardware configuration, block configuration, and flowchart are examples, and are not limited.

  For example, a mobile phone has been described as the speech encoding / decoding device 3 shown in FIG. 1, but the PHS (Personal Handyphone System), the PDA (Personal Digital Assistants), or a general personal computer has The invention can be easily applied. That is, the said embodiment is for description and does not restrict | limit the scope of the present invention.

  In the above-described embodiment, different VQ tables are used for high frequency and low frequency in vector quantization. However, the same V Q table may be used for all frequency bands of speech. Good. Further, the frequency band may be further divided and a different VQ table may be used for each frequency band.

  Also, although energy is often mentioned in ascending or descending order, even if only one of them is mentioned, it may be in ascending or descending order as long as a consistent policy is adopted as a whole. .

It is a figure which shows the frame division | segmentation of an input audio | voice signal, and the relationship between 1 frame and each block. It is a figure which shows typically the relationship between the MDCT for every block and the spectrum in a frame unit in Embodiment 1 of this invention. It is a figure which shows the physical structure of the audio | voice encoding and decoding apparatus which concerns on Embodiment 1 of this invention. It is a figure which shows the flow of the signal processing at the time of the audio | voice encoding and decoding apparatus which concerns on Embodiment 1 of this invention functions as an audio | voice encoding apparatus. It is a figure which shows the detail of the vector quantization related process part which concerns on Embodiment 1 of this invention. It is a figure which shows the flow which calculates | requires an index difference in the encoding side in Embodiment 1 of this invention. It is a figure which shows the flow of code length monitoring in Embodiment 1 of this invention. It is a figure which shows the flow of the signal processing when the audio | voice encoding / decoding apparatus which concerns on Embodiment 1 of this invention functions as an audio | voice decoding apparatus. It is a figure which shows the detail of the vector inverse quantization related process part which concerns on Embodiment 1 of this invention. It is a figure which shows the first half of the flow which calculates | requires a representative vector in the decoding side in Embodiment 1 of this invention. It is a figure which shows the second half of the flow which calculates | requires a representative vector in the decoding side in Embodiment 1 of this invention. It is the figure which represented typically the several representative vector of equal energy in Embodiment 2 of this invention in multidimensional space. It is the figure which represented typically the mode of the vector quantization before and behind a predetermined time interval in Embodiment 2 of this invention.

Explanation of symbols

  DESCRIPTION OF SYMBOLS 3 ... Speech encoding and decoding apparatus, 4 ... A / D conversion part, 5 ... Vector quantization part, 8 ... Entropy decoding part, 9 ... Index calculation part, 23 ... DC removal unit, 25 ... framing unit, 27 ... level adjustment unit, 29 ... frequency conversion unit, 31 ... spectrum completion unit, 33 ... vector quantization related processing unit, 35 ... Entropy encoding unit, 37... Code length monitoring unit, 39... Band data deletion unit, 41... Low band sorted VQ table, 43. .. Index difference calculation unit, 47... Representative vector index storage unit of vector quantization related processing unit, 49... Vector dequantization related processing unit, 51. Frequency inverse conversion unit, 55... Level reproduction unit, 7... Frame synthesis unit, 59... D / A conversion unit, 61... Representative vector index storage unit of vector inverse quantization related processing unit, 63... Vector dequantization unit, 121. CPU, 123 ... ROM, 125 ... storage unit, 131 ... RAM, 133 ... hard disk, 141 ... audio processing unit, 151 ... microphone, 153 ... speaker, 161 ...・ Wireless communication unit, 163... Antenna, 171... Operation key input content processing unit, 173... Operation key, 181.

Claims (11)

A framing unit that divides the digital audio signal into framed digital audio signals that are digital audio signals for each frame that is a predetermined time interval;
A frequency converter that converts the frequency of the framed digital audio signal and generates a digital spectrum for each frame;
A vector quantization table composed of representative vectors identified by the attached index and having a large absolute value in ascending or descending order of the index;
A vector quantization unit that obtains the index corresponding to the digital spectrum by vector-quantizing the digital spectrum using the vector quantization table;
An index storage unit that stores the index obtained by the vector quantization unit in association with the frame corresponding to the index;
The index obtained by the vector quantization unit is acquired from the vector quantization unit, and is stored in the index storage unit in association with the frame that is temporally earlier than the frame corresponding to the index. An index difference calculation unit that acquires an index from the index storage unit and calculates a difference between the acquired indexes;
An encoding unit that generates a code by entropy encoding the difference calculated by the index difference calculation unit;
A speech encoding device comprising: The index difference calculation unit
The index obtained by the vector quantization unit is acquired from the vector quantization unit, and the index stored in the index storage unit in association with the frame immediately preceding the frame corresponding to the index Is calculated from the index storage unit, and the difference between the two acquired indexes is calculated.
The speech coding apparatus according to claim 1. The vector quantization table is:
It consists of representative vectors that are expressed as one point in a multidimensional space using coordinate axes with element numbers corresponding to high and low frequencies, and for a plurality of representative vectors having the same absolute value, When arranged in ascending or descending order of the index, the element number assigned to the coordinate axis corresponding to the component having the maximum absolute value among the components corresponding to the coordinate axis of each representative vector is monotonously increased. The index is attached,
The speech encoding apparatus according to claim 1 or 2, characterized in that The vector quantization table is:
With multiple bandwidth tables,
The band-specific table is
Each consisting of a representative vector that is associated with a table band that is a particular band and matches a typical speech spectrum pattern in the table band;
It is characterized by
The vector quantization unit includes:
The digital spectrum is vector-quantized for each quantization band that is the same as or more subdivided with the table band, and the quantization band is included when vector quantization is performed for each quantization band. Using the band-specific table corresponding to the table band,
The speech coding apparatus according to any one of claims 1 to 3, wherein A code length monitoring unit for obtaining a code length of the code generated by the encoding unit and determining whether the code length is equal to or less than a preset target code length;
The encoding unit includes:
When the code length monitoring unit determines that the code length is longer than the target code length, it corresponds to a deletion band with relatively low energy in the digital spectrum divided into predetermined deletion bands. Entropy coding again after excluding the part from entropy coding,
The speech coding apparatus according to any one of claims 1 to 4, wherein the speech coding apparatus is characterized in that: The frequency converter is
Transforming the framed digital signal into a modified discrete cosine transform to generate the digital spectrum for each frame;
The speech encoding apparatus according to claim 1, wherein the speech encoding apparatus is a part of the speech encoding apparatus. According to an index attached to a digital spectrum belonging to a current frame that is one of frames in a predetermined time interval and generated from a digital audio signal and to a digital spectrum belonging to a frame that is earlier in time than the current frame As a result of performing vector quantization using a vector quantization table composed of representative vectors that are distinguished from each other and that have a large absolute value in ascending or descending order of the index, the values obtained as values that characterize each digital spectrum are obtained. A decoding unit that decodes the difference from data entropy-encoded as an amount corresponding to the current frame in which the difference between the indexes calculated from the index is;
An index storage unit for storing the index in association with the frame corresponding to the index;
The index corresponding to the frame, the difference decoded as an amount corresponding to the frame by the decoding unit, the index stored in the index storage unit in association with a frame that is temporally earlier than the frame An index calculation unit obtained by adding to
The vector quantization table;
An inverse vector quantization unit that approximately restores the digital spectrum by using the representative vector retrieved from the vector quantization table by searching the representative vector corresponding to the index obtained by the index calculation unit;
A frequency inverse transform unit for performing frequency inverse transform on the digital spectrum restored by the inverse vector quantization unit to restore a digital audio signal belonging to the frame to which the digital spectrum belongs;
A speech decoding apparatus comprising: A framing step of dividing the digital audio signal into framed digital audio signals that are digital audio signals for each frame that is a predetermined time interval;
A frequency conversion step of frequency-converting the framed digital audio signal to generate a digital spectrum for each frame;
The digital spectrum is vector-quantized using a vector quantization table that is configured by a representative vector that is identified by the attached index and that has a large absolute value in ascending or descending order of the index, thereby corresponding to the digital spectrum. A vector quantization step to find the index;
An index storing step for storing the index obtained by the vector quantum step in association with the frame corresponding to the index;
The index obtained by the vector quantization step is acquired from the vector quantization step and stored in the past index storing step in association with the previous frame in time than the frame corresponding to the index. An index difference calculating step for acquiring an index and calculating a difference between the acquired indexes;
An encoding step for generating a code by entropy encoding the difference calculated by the index difference calculating step;
A speech encoding method comprising: According to an index attached to a digital spectrum belonging to a current frame that is one of frames in a predetermined time interval and generated from a digital audio signal and to a digital spectrum belonging to a frame that is earlier in time than the current frame From the index obtained as a value that characterizes each digital spectrum as a result of vector quantization using a vector quantization table that is identified and composed of representative vectors having a large absolute value in ascending or descending order of the index A decoding step of decoding the difference from the data entropy-encoded as an amount corresponding to the current frame in which the difference between the calculated indexes is;
An index storing step of storing the index in association with the frame corresponding to the index;
The index corresponding to the frame is stored in the past index storing step in which the difference decoded as an amount corresponding to the frame by the decoding step is correlated with a past frame in time than the frame. An index calculation step to be obtained by adding to
An inverse vector quantization step of approximately restoring the digital spectrum by using the representative vector searched from the vector quantization table for the representative vector corresponding to the index obtained by the index calculation step;
Frequency inverse transform step for restoring the digital audio signal belonging to the frame to which the digital spectrum belongs by performing frequency inverse transform on the digital spectrum restored in the inverse vector quantization step;
A speech decoding method comprising: On the computer,
A framing step of dividing the digital audio signal into framed digital audio signals that are digital audio signals for each frame that is a predetermined time interval;
A frequency conversion step of frequency-converting the framed digital audio signal to generate a digital spectrum for each frame;
The digital spectrum is vector-quantized using a vector quantization table that is configured by a representative vector that is identified by the attached index and that has a large absolute value in ascending or descending order of the index, thereby corresponding to the digital spectrum. A vector quantization step to find the index;
An index storing step for storing the index obtained by the vector quantum step in association with the frame corresponding to the index;
The index obtained by the vector quantization step is acquired from the vector quantization step and stored in the past index storing step in association with the previous frame in time than the frame corresponding to the index. An index difference calculating step for acquiring an index and calculating a difference between the acquired indexes;
An encoding step for generating a code by entropy encoding the difference calculated by the index difference calculating step;
A program that executes On the computer,
According to an index attached to a digital spectrum belonging to a current frame that is one of frames in a predetermined time interval and generated from a digital audio signal and to a digital spectrum belonging to a frame that is earlier in time than the current frame From the index obtained as a value that characterizes each digital spectrum as a result of vector quantization using a vector quantization table that is identified and composed of representative vectors having a large absolute value in ascending or descending order of the index A decoding step of decoding the difference from the data entropy-encoded as an amount corresponding to the current frame in which the difference between the calculated indexes is;
An index storing step of storing the index in association with the frame corresponding to the index;
The index corresponding to the frame is stored in the past index storing step in which the difference decoded as an amount corresponding to the frame by the decoding step is correlated with a past frame in time than the frame. An index calculation step to be obtained by adding to
An inverse vector quantization step of approximately restoring the digital spectrum by using the representative vector searched from the vector quantization table for the representative vector corresponding to the index obtained by the index calculation step;
Frequency inverse transform step for restoring the digital audio signal belonging to the frame to which the digital spectrum belongs by performing frequency inverse transform on the digital spectrum restored in the inverse vector quantization step;
A program that executes

Images