JP4935280B2 - Speech coding apparatus, speech decoding apparatus, speech coding method, speech decoding method, and program - Google Patents

Description

  The present invention relates to a speech encoding device, speech decoding device, speech encoding method, speech decoding method, and program that are required when performing analysis / synthesis speech compression / decompression.

  In the field of mobile communications such as digital mobile phones, a low bit rate (about 8 kbps) voice compression coding method is required to cope with the increase in subscribers. For example, as an 8 kbps speech coding method, ITU-T Recommendation G. 729, there is a speech encoding method.

  The speech coding method according to the above-mentioned recommendation is basically a method of coding a speech signal after decomposing it into a prediction coefficient and a residual signal by predictive analysis. As prediction analysis, for example, linear prediction analysis and MLSA analysis (for example, see Non-Patent Document 1) are known.

  Since the bit rate is limited to, for example, about 8 kbps as described above, it is necessary to perform encoding so that the code length per unit time is constant.

  Conventionally, therefore, an encoding method based on vector quantization or the like has been adopted as an encoding method. In such an encoding method, when the accuracy of encoding, that is, how detailed the original information is to be reproduced, is determined, the compression rate is also determined.

  Therefore, if the encoding accuracy is determined, the code length per unit time is automatically fixed, which is convenient.

  From another point of view, the accuracy in encoding is convenient because it can be determined by back calculation based on the bit rate constraint.

When such an encoding method is employed, in order to increase or decrease the code length, for example, the order of prediction analysis may be increased or decreased, respectively. Conventionally, the order of prediction analysis has been fixed in advance to a specific order maximized within a range that satisfies the bit rate constraint.
Sei Imai, Kazuo Sumita, Chieko Furuichi, “Mel Log Spectrum Approximation (MLSA) Filter for Speech Synthesis”, IEICE Transactions, Vol. J66-A, No. 2, p. 122-129, 1983

  On the other hand, among the encoding methods, there is an encoding method called an entropy encoding method that performs encoding in consideration of the frequency of occurrence of numerical values included in a signal to be encoded.

  Comparing the entropy encoding method and the encoding method based on the above-described vector quantization, etc., on the premise that the encoding accuracy is equal, the former is more of the time during which the speech signal to be encoded continues. There are often times when the compression ratio is high.

  In other words, it can be said that adopting an encoding method based on vector quantization or the like in such a time zone wastes a limited bit rate.

  However, this does not mean that an encoding method based on vector quantization or the like is simply replaced with an entropy encoding method. The reason is as follows. That is, in the entropy encoding method, even if the encoding accuracy is constant, the compression rate is not uniquely determined. Therefore, the compression rate may be lower than the compression rate by an encoding method based on vector quantization or the like. That is, it is inconvenient because a time zone that does not satisfy the bit rate constraint may occur during the time that the audio signal to be encoded continues.

  As described above, if a coding method based on vector quantization or the like is adopted uniformly, a given bit rate cannot be fully utilized. On the other hand, if the entropy coding method is simply adopted instead to improve the compression rate, a time zone in which the bit rate constraint cannot be satisfied occurs.

  The present invention has been made in view of such circumstances, and provides an audio encoding device, an audio decoding device, an audio encoding method, an audio decoding method, and a program that make effective use of a limited bit rate as much as possible. For the purpose.

In order to achieve the above object, a speech encoding apparatus according to the first aspect of the present invention provides:
A prediction analysis unit that decomposes a speech signal into a prediction coefficient and a residual signal by a prediction analysis of a predetermined order;
A gain extraction unit for obtaining a gain of the residual signal;
Determining whether the residual signal is voiced or unvoiced and, if it is determined that the residual signal is voiced, a voiced and unvoiced discrimination and pitch extracting unit that extracts a pitch frequency from the residual signal;
An encoding unit that converts the pitch frequency into an entropy code when the prediction coefficient, the gain, the determination result, and the pitch frequency are extracted as a result of the determination;
It is determined whether or not the length of the entropy code exceeds an allowable length, and when it is determined that the length of the code exceeds the allowable length, the order of the prediction analysis in the prediction analysis unit is reduced to reduce the order of the series. A control unit that repeatedly executes an encoding operation;
A code transmission unit that transmits an entropy code that is repeatedly executed by the encoding unit under the control of the control unit and falls within an allowable length;
It is provided with.

  According to such an encoding device, an encoded audio signal for reproducing audio having the highest quality possible under the restriction can be generated in a situation where the amount of information transmission is restricted. This is because the higher the order of predictive analysis, the clearer the reproduced sound.

  The voiced / unvoiced discrimination / pitch extraction unit includes a low-pass filter that extracts a low-frequency part in advance from the residual signal, determines whether the low-frequency part is voiced sound or unvoiced sound, and the low-frequency part is voiced sound. If it is discriminated, it is desirable to extract the pitch frequency from the low frequency region.

  Since the pitch frequency, which is an amount that characterizes voiced sound, exists in a relatively low band, the accuracy of discrimination between voiced and unvoiced sound increases by passing the residual signal through a low-pass filter before the extraction of the pitch frequency.

  The prediction analysis unit decomposes the speech signal into a prediction coefficient and a residual signal by, for example, linear prediction analysis.

  The prediction analysis unit decomposes the speech signal into a prediction coefficient and a residual signal by, for example, MLSA (Mel Log Spectrum Approximation) analysis.

In order to achieve the above object, a speech decoding apparatus according to the second aspect of the present invention provides:
A receiving unit that receives an entropy code transmitted from the code transmitting unit of the speech encoding device according to the first aspect ;
A decoding unit that decodes the received entropy code and generates a prediction coefficient, a gain of the residual signal, a voiced / unvoiced discrimination result of the residual signal, and a pitch frequency in the case of voiced ;
When the voice signal is an unvoiced sound, noise having a gain equal to the residual signal gain is generated as an excitation signal. When the voice signal is a voiced sound, the noise has a gain equal to the residual signal gain. A signal generator for generating a pulse train having a frequency equal to the pitch frequency as an excitation signal;
A synthesis filter that restores speech by synthesizing the prediction coefficient and the excitation signal;
Is provided.

  When the speech decoding apparatus receives the prediction coefficient, the residual signal gain, the discrimination result of voiced or unvoiced sound, and the pitch frequency of voiced sound, it is necessary for speech restoration. In order to generate the excitation signal in the simplest and most reliable manner, it is appropriate to provide a signal generator having the above-described mechanism.

  According to the present invention, it is possible to encode and decode speech so as to maintain the sound quality of the original speech to the maximum, under the condition that the predetermined communication capacity is not exceeded.

  The speech encoding apparatus and speech decoding apparatus according to embodiments of the present invention will be described in detail below.

  FIG. 1 is a functional configuration diagram of the speech encoding device 111 according to the present embodiment.

  As shown in the figure, the speech encoding apparatus 111 includes a microphone 121, an A / D conversion unit 123, a prediction analysis unit 125, a gain extraction unit 127, a low-pass filter 129, a voiced / unvoiced discrimination / pitch extraction unit 131, and the like. , An encoding unit 133, a switch 135, a prediction analysis order adjustment unit 137, and a transmission unit 139. The prediction analysis unit 125 includes a prediction analysis inverse filter calculator 141.

  First, sound is input to the microphone 121. The voice is an analog signal. On the other hand, analysis and encoding performed later are discrete processes. Therefore, in order to prepare for this, the analog signal is converted into a digital audio signal by the A / D conversion unit 123 and sent to the prediction analysis unit 125.

  The prediction analysis unit 125 performs prediction analysis on the digital audio signal delivered from the A / D conversion unit 123. As the prediction analysis, for example, linear prediction analysis is used. Alternatively, MLSA (Mel Log Spectrum Approximation) analysis may be used. Both are known methods. The procedures of both analyzes will be described later in detail with reference to FIG.

  Precisely speaking, the prediction analysis performed by the prediction analysis unit 125 is a procedure for time-dividing a digital audio signal and calculating a prediction coefficient and a residual signal in the time interval for each time interval. The length of the time interval is preferably 5 ms, for example.

  In the following, it is assumed that the digital audio signal sent from the A / D conversion unit 123 to the prediction analysis unit 125 is time-divided into M time intervals. Also, let l be the number of digital audio signal data included in each time interval. Then, the entire digital audio signal contains (l × M) data.

As a whole, the prediction analysis unit 125 predicts the digital audio signal S _i = {s _{i, 0} ,..., S _{i, l−1} } (0 ≦ i ≦ M−1) in each time interval. A function of decomposing into a number of prediction coefficients equal to the order of analysis and residual signals D _i = {d _{i, 0} ,..., D _{i, l−1} } (0 ≦ i ≦ M−1). Have.

  More specifically, the prediction analysis unit 125 first calculates a prediction coefficient from the input digital audio signal. At this time, the order of the prediction analysis is a predetermined initial value.

Next, a prediction analysis inverse filter calculator 141 built in the prediction analysis unit 125 calculates a prediction analysis inverse filter from the prediction coefficient. Subsequently, a residual signal D _i (0 ≦ i ≦ M−1) is obtained as an output when the digital audio signal from the A / D converter 123 is input to the prediction analysis inverse filter.

  The prediction coefficient is sent to the encoding unit 133 as it is.

  On the other hand, the residual signal is not directly delivered to the encoding unit 133. When the residual signal is sent to the encoding unit 133 as it is and encoded, the amount of information is still too large, and the result is contrary to the audio compression assumed by the audio encoding device 111 according to the present embodiment. Because it becomes.

  Therefore, it is necessary to reduce the amount of information in advance by extracting only the essential features of the residual signal as much as possible, and then deliver the residual signal to the encoding unit 133.

The residual signal D _i (0 ≦ i ≦ M−1) generated by the prediction analysis unit 125 is delivered to the gain extraction unit 127 and the low-pass filter 129.

The gain extraction unit 127 obtains the gain, that is, the magnitude of the residual signal. Such is the magnitude of Determination There are various ways, e.g., as follows, to a value based on the mean square of the sample value and the gain G _i in the i-th time interval.
G _i = 10 × log ₁₀ {(d _{i, 0} ² +... + D _{i, l−1} ² ) / l}

  Here, the logarithm is based on the fact that the human auditory sense has logarithmic sensitivity to the volume of sound.

The gain G _i calculated in this way is delivered to the encoding unit 133.

  On the other hand, the low pass filter 129 extracts a low frequency component of the residual signal, for example, a component of 500 Hz to 1 kHz. This is to improve the accuracy of the next voiced / unvoiced discrimination. In other words, in order to discriminate between voiced and unvoiced sounds, components other than such low-frequency components are unnecessary, and may also cause a reduction in discrimination accuracy. Cut it into pieces.

The residual signal D _i = {d _{i, 0} ,..., D _{i, l−1} } (0 ≦ i ≦ M−1) is passed through the low-pass filter 129, so that the low frequency residual signal D _{Low, i} = {d _{Low, i, 0} ,..., D _{Low, i, l−1} } (0 ≦ i ≦ M−1). The low-frequency residual signal D _{Low, i} = {d _{Low, i, 0} ,..., D _{Low, i, l−1} } (0 ≦ i ≦ M−1) is used for the voiced / unvoiced discrimination and pitch extraction unit 131. To be handed over.

The voiced / unvoiced discrimination / pitch extraction unit 131 sends the discrimination result as to whether the low-frequency residual signal D _{Low, i} (0 ≦ i ≦ M−1) is a voiced sound or an unvoiced sound to the encoding unit 133. If it is determined that the sound is a voiced sound, the pitch frequency is also sent to the encoding unit 133 in addition to the determination result. These processes will be described in detail later with reference to FIG.

  Thus, from the residual signal, the gain, the discrimination result of voiced sound or unvoiced sound, and the pitch frequency in the case of voiced sound are extracted and sent to the encoding unit 133. It can be said that these extracted values and discrimination results essentially characterize the characteristics of the residual signal for a small amount of information in consideration of the characteristics of the audio signal. In this way, encoding only the feature amount of the residual signal means that the code length after encoding is reduced compared to the case where the entire residual signal is encoded as a whole. The degree of deterioration of the residual signal restored by the device is small in terms of auditory characteristics. Therefore, performing the above-described processing on the residual signal before encoding enables audio compression to the extent assumed by the audio encoding device 111 according to the present embodiment, and is restored by an audio decoding device described later. As a result, the degree of deterioration of the recorded voice with respect to the original voice is within an allowable limit.

  In the end, the encoding unit 133 receives the prediction coefficient from the prediction analysis unit 125, the gain from the gain extraction unit 127, and the voiced / unvoiced discrimination / pitch extraction unit 131 as a voiced / unvoiced discrimination result and voiced sound. If so, the pitch frequency is delivered. The encoding unit 133 encodes these together.

  The encoding unit 133 employs an entropy encoding method as an encoding method. The entropy encoding method has a disadvantage in that the compression rate cannot be predicted, but an extremely high compression rate may be realized depending on the distribution of the appearance frequency of elements included in the encoding target data. A high compression rate means that a signal that enables restoration of higher quality audio can be transmitted. This is because the amount of original information that can be transmitted can be increased, so that more feature amounts related to the original voice can be transmitted.

  Examples of the entropy encoding method include Huffman code and RangeCoder.

  In general, in audio signal communication, the amount of information that can be transmitted per unit time is constant. Therefore, as described above, the entropy encoding method having a nonuniformity in the compression rate seems to be unsuitable as an encoding method to be adopted for the audio signal communication.

  However, in this regard, in the time interval in which the compression rate is high, the high-quality speech signal is transmitted by utilizing this fact, and in the time interval in which the compression rate is low, the number of feature quantities to be encoded is set. It can be dealt with by reducing it until it fits in the given information capacity.

  By adopting such a policy, it is better to use the entropy encoding method than to use the normal encoding method with a constant compression rate, if all the time intervals are comprehensively determined. It can be said that it contributes to the reproduction of voice.

  The present embodiment embodies such a policy. For that purpose, in particular, the prediction analysis order adjustment unit 137 plays an important role.

  As described above, the prediction analysis unit 125 performs prediction analysis on the input digital audio signal, using a predetermined initial value as the order of prediction analysis.

  In general, the greater the order of prediction analysis, the greater the amount of information delivered to the encoding unit 133, and the more advantageous it is for faithfully reproducing the original audio signal.

  However, when the amount of information delivered to the encoding unit 133 increases, even if an entropy encoding method with uneven compression ratio is used as an encoding method, the code length after encoding becomes long on average. Certainly. Therefore, it can be said that there is an upper limit in the order of predictive analysis in order for the code length to be within a given information transmission capacity.

  However, since the entropy encoding method is an encoding method in which the compression ratio varies, the upper limit in the order of prediction analysis is not simply determined.

  In the present embodiment, such an upper limit is determined based on the case where the entropy encoding method achieves the highest compression rate, and is set as the above-described predetermined initial value in the prediction analysis.

  As mentioned above, if high compression is possible, it will be used to play back high-quality audio, while if it remains low compression, the amount of original information will be reduced as much as possible to minimize degradation of the playback audio. The policy adopted in the present embodiment is to suppress it.

  Therefore, at first, the prediction analysis is performed with the above-described initial value as the order of the prediction analysis in anticipation of the case where the compression ratio becomes the highest. Then, the encoding unit 133 actually performs entropy encoding, obtains a code length, and notifies the prediction analysis order adjustment unit 137 of the code length. At this time, the switch 135 is open, and the entropy code generated by the encoding unit 133 is not delivered to the transmission unit 139, so that no code is transmitted.

  The prediction analysis order adjustment unit 137 determines whether or not the code length notified from the encoding unit 133 satisfies a given information communication capacity limit. If it is determined that the restriction is satisfied, the switch 135 is instructed to permit transmission. Upon receiving the transmission permission instruction, the switch 135 is closed, and the entropy code generated by the encoding unit 133 is transferred to the transmission unit 139 and transmitted to the speech decoding apparatus 211 described later.

  As described above, since the initial value of the order of predictive analysis is determined based on the most convenient case for the entropy coding method, in many cases, the predictive analysis order adjusting unit 137 has the coding unit 133. It is determined that the code length notified from the above does not satisfy the restriction, and a transmission non-permission instruction is sent to the switch 135. The switch 135 that has received the transmission non-permission instruction remains open, and the entropy code is not delivered to the transmission unit 139 and is therefore not transmitted.

  When the entropy code length exceeds the predetermined code length in this way, the prediction analysis order adjustment unit 137 reduces the order of prediction analysis by 1 to the prediction analysis unit 125 and then performs prediction analysis again. Order to start over. When the order of prediction analysis decreases, the amount of information sent to the encoding unit 133 decreases, so the possibility that the code length of the entropy code generated by the encoding unit 133 is less than or equal to the predetermined code length is higher than in the previous case. .

  The code length of the generated entropy code is notified to the prediction analysis order adjustment unit 137 again. The prediction analysis order adjustment unit 137 performs the same determination as in the previous case, notifies the switch 135 of permission of transmission or notifies the switch 135 of permission of transmission and notifies the prediction analysis unit 125 of the order of prediction analysis. Is further reduced by 1 and the prediction analysis is performed again.

  If such a procedure is followed, in any case, transmission permission is issued from the prediction analysis order adjustment unit 137 to the switch 135, the entropy code is handed over to the transmission unit 139, and further transmitted to a speech decoding apparatus to be described later.

  In the present embodiment, the transmission unit 139 adopts a transmission method by wireless communication, but may be various other methods such as wired communication and communication using both wired and wireless.

  As described above, by adopting the entropy encoding method, the policy of the present embodiment is used in which the given information communication capacity is utilized to the maximum to help reproduce high-quality audio.

  For example, when the sampling frequency is 8 kHz, the initial value of the order of prediction analysis is set to 10, and when it overflows, the order is lowered by 1 like 9, 8,... Until the target code length is reached. Go.

  FIG. 2 is a functional configuration diagram of the speech decoding apparatus 211 according to the present embodiment.

  As shown in the figure, the speech decoding apparatus 211 includes a receiving unit 231, a decoding unit 233, a residual signal restoration unit 235, a synthesis filter calculation unit 237, a synthesis filter unit 239, and a D / A conversion unit 241. And a speaker 243.

  The reception unit 231 receives from the transmission unit 139 of the speech encoding device 111 of FIG. 1 the one in which the prediction coefficient and the residual signal information are encoded together (entropy code) by the wireless communication unit, and sends it to the decoding unit 233. hand over.

  The decoding unit 233 decodes the entropy code delivered from the receiving unit 231, and in each time segment, the prediction coefficient, the residual signal gain, the voiced / unvoiced discrimination result of the residual signal, and the pitch frequency in the case of voiced And generate. Note that information about the order of prediction analysis performed by the speech coding apparatus 111 after all, which is information necessary for speech synthesis, is not directly transmitted to the speech decoding apparatus 211. However, such information can be obtained by counting the number of decoded prediction coefficients.

  The gain, which is information decoded with respect to the residual signal, and the voiced / unvoiced discrimination result of the residual signal and the pitch frequency in the case of voiced are delivered to the residual signal restoration unit 235.

  The residual signal restoration unit 235 restores the residual signal based on the result of collecting the residual signal of the original speech into several feature amounts. In this sense, it can be said that the residual signal restoration unit 235 is a pseudo residual signal generation unit.

The pseudo residual signal generated by the residual signal restoration unit 235 is represented as D ′ _i = {d ′ _{i, 0} ,..., D ′ _{i, l−1} } (0 ≦ i ≦ M−1). The pseudo residual signal D ′ _i is a pulse train or noise. If the received voiced / unvoiced discrimination result is voiced sound, the residual signal restoration unit 235 generates a pulse train having the same pitch frequency as the received pitch frequency and having a magnitude corresponding to the received gain. On the other hand, if the received voiced / unvoiced discrimination result is an unvoiced sound, a noise value is obtained by multiplying a signal value sequence of size 1 having a random time interval prepared in advance by a size corresponding to the received gain. Generate a column. The procedure for generating such a pulse train or noise train will be described in detail later with reference to FIG. The pseudo residual signal D ′ _i is delivered to the synthesis filter unit 239 as an excitation signal.

  On the other hand, the prediction coefficient decoded by the decoding unit 233 is transferred to the synthesis filter calculation unit 237 and used to calculate a speech synthesis filter. The filter for speech synthesis is a filter having such a property that a speech signal is reproduced by inputting an excitation signal to the filter.

  The filter calculation result by the synthesis filter calculation unit 237 is sent to the synthesis filter unit 239. The synthesizing filter unit 239 determines its own specification according to the received filter calculation result. Alternatively, it may be considered that the synthesis filter unit 239 is generated by the synthesis filter calculation unit 237.

When the pseudo residual signal D ′ _i is input as an excitation signal to the synthesizing filter unit 239, an audio signal as digital data is restored. The audio signal restoration procedure described above will be described in detail later with reference to FIG.

  The reproduction signal output from the synthesizing filter unit 239 is transmitted to the speaker 243 after being converted into an analog audio signal by the D / A conversion unit 241. The speaker 243 actually emits sound according to the received analog signal.

  The speech encoding apparatus 111 and speech decoding apparatus 211 that have been described with reference to FIGS. 1 and 2 which are functional configuration diagrams so far are physically integrated from the viewpoint of usability. Is realized by the speech encoding / decoding device 311 shown in FIG. In the following description, a mobile phone is assumed as the speech encoding / decoding device 311.

  The speech encoding / decoding device 311 includes the microphone 121 already shown in FIG. 1 and the speaker 243 already shown in FIG. 2, and further includes an antenna 353 and operation keys 363.

  The speech encoding / decoding device 311 further includes a CPU 321, a ROM (Read Only Memory) 323, a storage unit 325, a speech processing unit 341, a wireless communication unit 351, and an operation key input processing unit 361. These are connected to each other via a system bus 371. The system bus 371 is a transmission path for transferring commands and data.

  The ROM 323 stores an operation program for speech encoding and decoding and an operating system necessary for overall control of the speech encoding / decoding device 311.

  In the present embodiment, the prediction analysis unit 125, the gain extraction unit 127, the low-pass filter 129, the voiced / unvoiced discrimination / pitch extraction unit 131, the switch 135, the prediction analysis order adjustment unit 137, and the residual signal restoration of FIG. The functions of the unit 235, the synthesis filter calculation unit 237, and the synthesis filter unit 239 are realized by numerical processing by the CPU 321 in FIG. The operation program stored in the ROM 323 includes a program for such numerical processing by the CPU 321.

  The CPU 321 encodes or decodes sound by executing an operation program or an operating system stored in the ROM 323.

As described above, the CPU 321 performs numerical calculations in accordance with the operation program stored in the ROM 323. For this purpose, a numerical sequence to be processed, for example, a digital audio signal S _i (0 ≦ i ≦ M−1) is stored, or a numerical sequence as a processing result, for example, a residual signal D _i (0 ≦ i ≦ M−1). -1) is required to be stored.

  The storage unit 325 includes a RAM (Random Access Memory) 331 and a hard disk 333, and includes the order of prediction analysis, digital speech signal, prediction coefficient, residual signal, gain, voiced / unvoiced discrimination result, and pitch of voiced sound. The frequency, prediction coefficient and residual signal information encoded together, pulse train, noise train, inverse filter calculation result, pseudo residual signal, etc. are stored.

  The CPU 321 has a built-in register (not shown), and according to an operation program read from the ROM 323, appropriately loads a numerical sequence to be processed from the storage unit 325 into the register, and stores the predetermined numerical sequence in the loaded numerical sequence. And the result is stored in the storage unit 325.

  The wireless communication unit 351 and the audio processing unit 341 function as follows when the audio encoding / decoding device 311 is the audio encoding device 111 (FIG. 1). That is, the sound input to the microphone 121 and converted into a digital signal by the A / D conversion unit 123 (FIG. 1) included in the sound processing unit 341 is encoded by the CPU 321, ROM 323, and storage unit 325 through the process shown in FIG. It becomes. Then, the wireless communication unit 351 uses the antenna 353 to function as the transmission unit 139 (FIG. 1) to the other party (another speech encoding and decoding device 311 on the receiving side) and the encoded prediction coefficient and encoding. Residual signal information is transmitted.

  The wireless communication unit 351 and the speech processing unit 341 function as follows when the speech encoding / decoding device 311 is the speech decoding device 211 (FIG. 2). That is, the wireless communication unit 351 receives the encoded prediction coefficient and the encoded residual signal information using the antenna 353 so as to function as the receiving unit 231 (FIG. 2). The received signal is decoded by the CPU 321, ROM 323, and storage unit 325 into a digital audio signal through the process shown in FIG. The digital audio signal is converted into an analog audio signal using a D / A conversion unit 241 (FIG. 2) provided in the audio processing unit 341, and is output from the speaker 243 as audio.

  The operation key input processing unit 361 receives an operation signal from the operation key 363 and inputs a key code signal corresponding to the operation signal to the CPU 321. The CPU 321 determines the operation content based on the input key code signal.

  The user may be able to customize the speech encoding / decoding device 311 so that the user can easily use the operation keys 363 with respect to the set variables and the like. For example, the initial value of the order of predictive analysis is determined in advance based on the considerations already described, and is described in the program stored in the ROM 323 in principle. However, this may be rewritten by the user using the operation keys 363. If the initial value is reduced, the advantage of entropy coding will not be fully exerted, but on average the number of trials will fit within a predetermined code length, so the processing speed will be improved and the user will be able to talk. It may be possible to realize an improvement in real-time feeling during the event.

  In addition, the operation key 363 specifies a transmission partner device among the other speech encoding / decoding devices 311 that are widely distributed when the speech encoding / decoding device 311 functions as the speech encoding device 111. It is also necessary to enter a number (telephone number, etc.).

(Predictive analysis procedure)
Hereinafter, the prediction analysis performed by the prediction analysis unit 125 of FIG. 1 will be described with reference to the flowchart shown in FIG. As prediction analysis, for example, linear prediction analysis and MLSA (Mel Log Spectrum Approximation) analysis are known. In FIG. 4, both analyzes are shown together with the latter in parentheses.

The storage unit 325 (FIG. 3) already stores digital audio signals (input waveforms) S _i = {s _{i, 0} ,..., Si _{, l−1} } (0 ≦ i ≦ M−1). Suppose that

  The CPU 321 (FIG. 3) uses a built-in counter register (not shown) for storing the input signal sample counter i, and sets i = 0 as an initial value (step S411 in FIG. 4).

The CPU 321 loads input signal samples S _i = {s _{i, 0} ,..., S _{i, l−1} } from the storage unit 325 (FIG. 3) into a built-in general-purpose register (not shown) ( Step S413 in FIG.

In the case of linear prediction analysis, the CPU 321 calculates linear prediction coefficients A _i = {a _{i, 1} ,..., A _{i, n} } from the input signal sample S _i (step S415). Where n is the order of linear predictive analysis. As a calculation method, any known method may be employed as long as the residual signal is evaluated to be sufficiently small based on a predetermined scale. For example, it is preferable to use a well-known calculation method that combines the calculation of the autocorrelation function and the Levinson-Durbin algorithm.

In the case of MLSA analysis, the CPU 321 first calculates cepstrum C _i = {c _{i, 0} ,..., C _{i, (l / 2) −1} } from the input signal sample S _i . Any known method may be employed for such calculation. In any method, procedures such as discrete Fourier transform, absolute value, logarithm, and inverse discrete Fourier transform are generally performed. Next, MLSA filter coefficients M _i = {m _{i, 0} ,..., M _{i, p−1} } are calculated from the obtained cepstrum C _i by any known method (step S415). In the case of MLSA analysis, p corresponds to the order of prediction analysis.

In the case of linear prediction analysis, linear prediction coefficients A _i = {a _{i, 1} ,..., A _{i, n} }, and in the case of MLSA analysis, MLSA filter coefficients M _i = {m _{i, 0} ,. m _{i, p−1} } is stored as a prediction coefficient in the storage unit 325 (step S417).

Subsequently, in the case of linear prediction analysis, an inverse prediction filter AIA _i for prediction analysis is calculated from the linear prediction coefficient A _i by an arbitrary known method. In the case of MLSA analysis, an arbitrary linear prediction filter A _i is calculated from the MLSA filter coefficient M _i . An inverse MLSA filter AIM _i for predictive analysis is calculated by a known method. (Step S419) These calculations correspond to the calculations performed by the prediction analysis inverse filter calculator 141 of FIG.

The input signal sample S _i = {s _{i, 0} ,..., S _{i, l−1} } is passed through the obtained prediction analysis inverse linear prediction filter AIA _i or prediction analysis inverse MLSA filter AIM _i. , Residual signal D _i = {d _{i, 0} ,..., D _{i, l−1} } is obtained (step S421 in FIG. 4). The residual signal D _i is stored in the storage unit 325 (step S423).

  Here, it is determined whether or not the input signal sample counter i has reached M−1 (step S425). If it has been reached (step S425; Yes), the process ends. On the other hand, if not reached (step S425; No), i is incremented by 1 (step S427) in order to perform processing on the input signal sample in the next time interval, and the processing after step S413 is repeated.

(Entropy code generation procedure)
Hereinafter, prediction analysis, gain extraction, voiced and unvoiced discrimination and pitch extraction, trial of entropy encoding, and generation of an entropy code to be actually transmitted, which are performed under the control of the prediction analysis order adjustment unit 137 of FIG. The procedure will be described with reference to the flowchart shown in FIG. Here, a case where linear prediction analysis is adopted as prediction analysis will be described, but the same applies when MLSA analysis is adopted.

The CPU 321 (FIG. 3) sets the input signal sample counter i to i = 0 (step S511 in FIG. 5), and the order n of the linear prediction coefficient A _i is set to n _initial which is a predetermined value based on the consideration already described. Set (step S513).

Then, according to the procedure shown in FIG. 4, from the input signal samples S _i = {s _{i, 0} ,..., S _{i, l−1} }, linear prediction coefficients A _i = {a _{i, 1} ,. A _{i, n} } and residual signal D _i = {d _{i, 0} ,..., D _{i, l−1} } are calculated (step S515 in FIG. 5).

A gain G _i is calculated from the residual signal D _i (step S517). G _i is, for example,
G _i = 10 × log ₁₀ {(d _{i, 0} ² +... + D _{i, l−1} ² ) / l}
It is calculated as follows.

Next, it is determined whether the sound in the i-th time interval is a voiced sound or an unvoiced sound. Whether a voiced sound, in other words, the residual signal D _i (or, a low pass filter 129 (FIG. 1) lowband residual signal D _Low after _passage, is a _i, in the following, D _{Low, i is} also simply referred to as D _i .) whether or not it has the property of pitch. If the residual signal _Di has periodicity, it can be said that it has a pitch property. Therefore, it is examined whether _Di has periodicity.

Any known method may be used to check for the presence or absence of periodicity.For example, it is possible to obtain a standardized autocorrelation function and check whether a sufficiently large maximum value exists. Is preferred. If such a maximum value exists, it can be said that periodicity also exists, and furthermore, it can be said that the time interval t _MAX that provides such a maximum value is the period. On the other hand, if there is no such maximum value, it can be said that there is no periodicity.

The autocorrelation function C (t) of the residual signal D _i is
C (t) = d _{i, 0} × d _{i, t}
+ D _{i, 1} xd _{i, t + 1}
+ ...
+ D _{i, l-1-t} × d _{i, l-1}
It is. As can be seen from this equation, t is an interval in which the number of elements in the residual signal D _i as a unit. Therefore, strictly speaking, the time interval to be considered here is that each element multiplied by the time interval sampled in t included in the residual signal D _i. Therefore, in this respect, care must be taken in obtaining the pitch frequency. However, since the time interval during which each element included in the residual signal D _i is normally sampled is constant, the time interval to be considered here is proportional to t. Therefore, hereinafter, when there is no possibility of confusion, the time interval to be considered here is simply denoted by t.

In the normalized autocorrelation function C (t), be any method as long as the method magnitude of the autocorrelation function C (t) is not depend on the size of the entire residual signal D _i For example, the normalization factor REG (t)
REG (t) = {(d _{i, 0} ² + ... + d _{i, l-1-t} ² )
× (d _{i, t} ² + ... + d _{i, l-1} ² )} ^0.5
And define the normalized autocorrelation function C _REG (t) as
C _REG (t) = C (t) / REG (t)
Is preferably defined.

The predetermined threshold C _th may be an arbitrary value as long as it is a numerical value useful for determining whether or not there is a clear maximum value in the normalized autocorrelation function C _REG (t), but is set to 0.5, for example. Is preferred.

As described above, in step S519, the normalized autocorrelation function C _REG (t) is calculated from the residual signal D _i, and the maximum value C _REG (t) where C _REG (t = t _MAX )> C _th (= 0.5) is obtained. = T _MAX ) exists.

If it exists, it can be said that the residual signal D _i has a property as a voiced sound (step S519; Yes). _Therefore, Flag _{VorUV, i} which is a variable representing voiced sound or unvoiced sound _{, i} is Flag _{VorUV, i} = “V “(Means voiced sound)” is set and stored in the storage unit 325. Further, the pitch frequency Pitch _i is calculated by taking the reciprocal of t _MAX which is the value of t that has caused the maximum value in the normalized autocorrelation function C _REG (t), and is stored in the storage unit 325 (step S521). Proceed to step S525.

When there is no t that causes a maximum value of C _REG (t)> C _th (= 0.5) in the normalized autocorrelation function C _REG (t) (step _S519 ; No), Flag _{VorUV, i} = ” UV "(meaning unvoiced sound) is set and stored in the storage unit 325 (step S523), and the process proceeds to step S525.

In step S525, the linear prediction coefficient A _i , the gain G _i , the pitch discrimination flag Flag _{VorUV, i} and the pitch frequency Pitch _i if present are collectively entropy by an entropy encoding method such as Huffman code or RangeCoder. Encode. Then, the code length of the generated entropy code is calculated.

  Subsequently, it is determined whether or not the calculated code length is equal to or less than a predetermined target code length in consideration of circumstances such as transmittable communication capacity (step S527). If an overflow has occurred, that is, if the calculated code length is larger than the target code length (step S527; No), the order n of the prediction analysis is reduced by 1 (step S529), and the process goes to step S515. Return and repeat the entropy coding trial.

  When the calculated code length is less than or equal to the target code length (step S527; Yes), the entropy code generated in step S525 is actually transmitted, so that the code is stored in the storage unit 325 in preparation for that. (Step S531).

  Subsequently, it is determined whether or not i is equal to or greater than (M−1) (step S533). If i has reached M−1 (step S533; Yes), the processing for all the time sections is completed, and the process is terminated. If i has not reached M−1 (step S533; No), in order to perform the process for the next time interval, i is increased by 1 (step S535), and the process returns to step S513.

(Pulse train or noise train generation procedure)
Hereinafter, processing performed by the residual signal restoration unit 235 of FIG. 2 will be described with reference to the flowchart shown in FIG.

  Processing in the i-th time segment (0 ≦ i ≦ M−1) will be described.

The CPU 321 (FIG. 3) loads the gain G _i and the voiced / unvoiced discrimination variable Flag _{VorUV, i} from the storage unit 325 (FIG. 3) to the general-purpose register (step S611 in FIG. 6).

It is determined whether or not the voiced / unvoiced discrimination variable Flag _{VorUV, i} is Flag _{VorUV, i} = “V” (step S613). That is, it is determined whether or not the original residual signal _Di is a voiced sound.

If it is a voiced sound (step S613; Yes), in step S521 of FIG. 5, Pitch _i is generated by the voiced / unvoiced discrimination and pitch extraction unit 131 (FIG. 1) of the voice encoding / decoding device 311 on the transmission side. Therefore, the pitch frequency Pitch _i should be stored in the storage unit 325 of the speech encoding / decoding device 311 on the receiving side through encoding / transmission / reception / decoding. Therefore, Pitch _i is loaded (step S615).

Subsequently, the residual signal is restored. That is, a pulse train D ′ _i = {d ′ _{i, 0} ,..., D ′ _{i, l−1} } having a gain G _i and a period of the pitch frequency Pitch _i is generated (Step 1 S617). This is the restored residual signal. The pulse train D ′ _i is generated assuming the same sampling interval as that of the original residual signal.

Since D ′ _i is generated according to the sampling interval of the original residual signal, the value of each element d ′ _{i, 0} ,..., D ′ _{i, l−1} is actually 0 or G, respectively. _Limited to one of _i . In addition, in these element sequences arranged in chronological order, G _i appears at every number interval corresponding to the pitch period that is the reciprocal of Pitch _i , and the values of the other elements are zero.

If it is determined in step S613 that the original residual signal is not a voiced sound (step S613; No), it is determined that the original residual signal is an unvoiced sound. Therefore, a signal value sequence D ′ _i = {d ′ _{i, 0} ,..., D ′ _{i, l−1} } that is appropriate as noise is generated according to the following procedure while reflecting the gain G _i. .

First, a basic noise sequence R _i = {r _{i, 0} ,..., R _{i, l-1} } having a size of ± 1 and a time interval of a random number is generated (step S619).

Here, _Ri is generated assuming that the sampling interval is the same as the sampling interval of the original residual signal. Therefore, in practice, the value of each element r _{i, 0} ,..., R _{i, l−1} is either 0, +1, or −1. Moreover, in these element sequences arranged in time order, +1 or −1 appears at random number intervals, and the values of the other elements are zero.

By multiplying the obtained basic noise sequence R _i by the loaded gain G _i , a noise sequence D ′ _i = {d ′ _{i, 0} ,..., D ′ _{i, l−1} } is generated. (Step S621).

Thus, D ′ _i = {d ′ _{i, 0} ,..., Which is a residual signal restored as a pulse train or a noise train, regardless of whether the original residual signal is a voiced sound or an unvoiced sound. D ′ _{i, l−1} } is generated. Since this is used later for reproduction of an audio signal, it is stored in the storage unit 325 (step S623).

(Procedure for audio signal restoration)
Hereinafter, the procedure of audio signal restoration by the synthesis filter calculation unit 237 and the synthesis filter unit 239 in FIG. 2 will be described with reference to the flowchart shown in FIG. Although the case where linear prediction analysis is adopted as the prediction analysis will be described, the procedure is the same in other cases, for example, when MLSA analysis is adopted.

  The CPU 321 (FIG. 3) sets the input signal sample counter to i = 0 in the counter register (step S711 in FIG. 7).

The CPU 321 loads the linear prediction coefficient A _i = {a _{i, 1} ,..., A _{i, n} } from the storage unit 325 (FIG. 3) to the general-purpose register (step S713 in FIG. 7).

Next, a synthesis filter CIA _i is calculated from the linear prediction coefficient A _i by any known method (step S715). This is an operation performed by the synthesis filter calculation unit 237 of FIG.

Subsequently, the pseudo residual signal D ′ _i = {d ′ _{i, 0} ,..., D ′ _{i, l−1} } is loaded and passed through the synthesis filter CIA _i , thereby obtaining the speech signal S ′ _i. = {S ′ _{i, 0} ,..., S ′ _{i, l−1} } is restored (step S717).

The restored audio signal S ′ _i is stored in the storage unit 325 (step S719).

  It is determined whether or not the input signal sample counter i has reached M−1 (step S721). If it has been reached (step S721; Yes), since all the audio signals to be restored have been restored, the process is terminated. If not reached (step S721; No), in order to restore the audio signal in the next time interval, i is incremented by 1 (step S723), and the processing after step S713 is repeated.

(Example of procedure for obtaining MLSA coefficients from cepstrum)
Figure 8 is a cepstrum _{C i = {c i, 0} , ···, c i, (l / 2) -1} MLSA filter coefficients from _{M i = {m i, 0} , ···, m i, p _-1 } is a flowchart illustrating an example of a specific procedure. By performing the calculations shown in steps S811 to S835, the MLSA filter coefficient is obtained. α is a numerical value for approximation, and α = 0.35 is preferable when the audio signal is sampled at 10 kHz. Further, β = 1−α ² . m _i (0 ≦ i ≦ p−1) is initialized to 0.

An example of the configuration of the MLSA filter using the MLSA filter coefficient obtained in this way is shown in FIG. P _{1 to} P ₄ are approximation coefficients. For example, it is preferable that P ₁ = 0.4999, P ₂ = 0.1067, P ₃ = 0.0117, and P ₄ = 0.0005656.

  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, the description has been made assuming that a mobile phone is used as the speech encoding / decoding device 311 shown in FIG. The present invention can also be applied in the same manner. For example, when the present invention is applied to a personal computer, if a voice input / output device, a communication device, or the like is added to the personal computer, it can have the function of a mobile phone as hardware. Then, if a computer program for causing a computer to execute the above-described processing is distributed by a recording medium or communication, the computer is installed and executed on the computer, thereby causing the computer to execute the speech encoding apparatus or the speech according to the present invention. It is also possible to function as a decoding device.

  That is, the said embodiment is for description and does not restrict | limit the scope of the present invention. Therefore, those skilled in the art can employ embodiments in which each or all of these elements are replaced with equivalent ones, and these embodiments are also included in the scope of the present invention.

It is a functional block diagram of the audio | voice coding apparatus provided with the prediction analysis order adjustment part based on embodiment of this invention. It is a functional block diagram of the speech decoding apparatus which concerns on embodiment of this invention. It is a figure which shows the physical structure of the audio | voice encoding / decoding apparatus which concerns on embodiment of this invention. It is a figure which shows the flow of a prediction analysis. It is a figure which shows the flow of entropy code generation. It is a figure which shows the flow which produces | generates a pulse train or a noise train. It is a figure which shows the flow which restore | restores an audio | voice signal. It is a figure which shows an example of the flow of calculation of an MLSA filter coefficient. It is a figure which shows an example of an MLSA filter.

Explanation of symbols

  DESCRIPTION OF SYMBOLS 111 ... Speech coding apparatus, 121 ... Microphone, 123 ... A / D conversion part, 125 ... Prediction analysis part, 127 ... Gain extraction part, 129 ... Low pass filter, 131 * ..Voiced / unvoiced discrimination and pitch extraction unit, 133 ... encoding unit, 135 ... switch, 137 ... prediction analysis order adjustment unit, 139 ... transmission unit, 141 ... inverse filter for prediction analysis Calculator 211... Speech decoding device, 231... Receiving unit, 233... Decoding unit, 235... Residual signal restoring unit, 237. Filter unit, 241... D / A conversion unit, 243... Speaker, 311... Voice encoding and decoding device, 321... CPU, 323. 331... RAM, 333 ... hard disk, 341 ... audio processing unit, 351 ... wireless communication unit, 353 ... antenna, 361 ... operation key input processing unit, 363 ... operation key, 371 ... system bus

Claims (9)

A prediction analysis unit that decomposes a speech signal into a prediction coefficient and a residual signal by a prediction analysis of a predetermined order;
A gain extraction unit for obtaining a gain of the residual signal;
Determining whether the residual signal is voiced or unvoiced and, if it is determined that the residual signal is voiced, a voiced and unvoiced discrimination and pitch extracting unit that extracts a pitch frequency from the residual signal;
An encoding unit that converts the pitch frequency into an entropy code when the prediction coefficient, the gain, the determination result, and the pitch frequency are extracted as a result of the determination;
It is determined whether or not the length of the entropy code exceeds an allowable length, and when it is determined that the length of the code exceeds the allowable length, the order of the prediction analysis in the prediction analysis unit is reduced to reduce the order of the series. A control unit that repeatedly executes an encoding operation;
A code transmission unit that transmits an entropy code that is repeatedly executed by the encoding unit under the control of the control unit and falls within an allowable length;
A speech encoding apparatus comprising: The voiced / unvoiced discrimination and pitch extraction unit,
A low-pass filter for extracting a low-frequency portion from the residual signal in advance,
Determining whether the low frequency part is voiced sound or unvoiced sound and extracting the pitch frequency from the low frequency part when the low frequency part is determined to be voiced sound;
The speech coding apparatus according to claim 1. The prediction analysis unit
The speech signal is decomposed into a prediction coefficient and a residual signal by linear predictive analysis.
The speech encoding apparatus according to claim 1 or 2 , characterized in that The prediction analysis unit
The audio signal is decomposed into a prediction coefficient and a residual signal by MLSA (Mel Log Spectrum Approximation) analysis.
The speech encoding apparatus according to claim 1 or 2 , characterized in that A receiving unit for receiving an entropy code transmitted from the code transmitting unit of the speech encoding device according to claim 1 ;
A decoding unit that decodes the received entropy code and generates a prediction coefficient, a gain of the residual signal, a voiced / unvoiced discrimination result of the residual signal, and a pitch frequency in the case of voiced ;
When the voice signal is an unvoiced sound, noise having a gain equal to the residual signal gain is generated as an excitation signal. When the voice signal is a voiced sound, the noise has a gain equal to the residual signal gain. A signal generator for generating a pulse train having a frequency equal to the pitch frequency as an excitation signal;
A synthesis filter that restores speech by synthesizing the prediction coefficient and the excitation signal;
A speech decoding apparatus comprising: A predictive analysis step that decomposes the speech signal into predictive coefficients and residual signals by predictive analysis;
A gain extracting step for obtaining a gain of the residual signal;
Determining whether the residual signal is voiced or unvoiced and, if it is determined that the residual signal is voiced, a voiced / unvoiced discrimination and pitch extraction step of extracting a pitch frequency from the residual signal;
An encoding step for converting the pitch frequency into an entropy code when the prediction coefficient, the gain, the determination result, and the pitch frequency are extracted as a result of the determination;
A code length examination step of calculating a length of the entropy code and determining whether the length exceeds an allowable length; and
When it is determined in the code length examination step that the length of the entropy code exceeds the allowable length, a subtraction step for reducing the order of prediction analysis in the prediction analysis step;
Consisting of
Until the entropy code falls within an allowable length, the prediction analysis step, the gain extraction step, the voiced unvoiced discrimination and pitch extraction step, the encoding step, the code length examination step, the subtraction step, Is repeated, and an entropy code within the allowable length is transmitted . A receiving unit step for receiving an entropy code encoded by the speech encoding method according to claim 6 ;
A decoding step of decoding the received entropy code to generate a prediction coefficient, a gain of the residual signal, a voiced / unvoiced discrimination result of the residual signal, and a pitch frequency in the case of voiced ;
When the voice signal is an unvoiced sound, noise having a gain equal to the residual signal gain is generated as an excitation signal. When the voice signal is a voiced sound, the noise has a gain equal to the residual signal gain. A signal generation step of generating a pulse train having a frequency equal to the pitch frequency as an excitation signal;
A synthesis step of restoring speech by synthesizing the prediction coefficient and the excitation signal;
A speech decoding method comprising: On the computer,
A predictive analysis step that decomposes the speech signal into predictive coefficients and residual signals by predictive analysis;
A gain extracting step for obtaining a gain of the residual signal;
Determining whether the residual signal is voiced or unvoiced and, if it is determined that the residual signal is voiced, a voiced / unvoiced discrimination and pitch extraction step of extracting a pitch frequency from the residual signal;
An encoding step for converting the pitch frequency into an entropy code when the prediction coefficient, the gain, the determination result, and the pitch frequency are extracted as a result of the determination;
A code length examination step of calculating a length of the entropy code and determining whether the length exceeds an allowable length; and
When it is determined in the code length examination step that the length of the entropy code exceeds the allowable length, a subtraction step for reducing the order of prediction analysis in the prediction analysis step;
The prediction analysis step, the gain calculation step, the voiced / unvoiced discrimination and pitch extraction step, the encoding step, the code length examination step, and the subtraction step until the entropy code falls within an allowable length. , and re-encoding step of repeating,
A code transmission step of transmitting the entropy code obtained by the re-encoding step;
A computer program that executes On the computer,
A receiving unit step for receiving an entropy code encoded by the speech encoding method according to claim 6 ;
A decoding step of decoding the received entropy code to generate a prediction coefficient, a gain of the residual signal, a voiced / unvoiced discrimination result of the residual signal, and a pitch frequency in the case of voiced ;
When the voice signal is an unvoiced sound, noise having a gain equal to the residual signal gain is generated as an excitation signal. When the voice signal is a voiced sound, the noise has a gain equal to the residual signal gain. A signal generation step of generating a pulse train having a frequency equal to the pitch frequency as an excitation signal;
A synthesis step of restoring speech by synthesizing the prediction coefficient and the excitation signal;
A computer program that executes

Images