JP4816107B2 - Speech coding apparatus and speech coding method - Google Patents

Description

The present invention relates to a speech encoding apparatus and speech encoding how.

  Many speech coding methods used in mobile phones, etc., calculate a linear prediction filter from an input speech signal, search for an excitation signal close to the speech signal, and parameters for both the excitation signal and the linear prediction coefficient Is output as a code (for example, see Non-Patent Document 1).

  FIG. 9 shows the configuration of a conventional speech encoding apparatus 300. The linear prediction analysis unit 30 calculates a linear prediction coefficient by performing linear prediction analysis of the input speech signal. Here, a certain length of the audio signal on the time axis is grouped and a linear prediction coefficient is calculated in the unit of the certain length. In order to reduce the amount of generated code of the speech encoding apparatus 300, the linear prediction coefficients are encoded with every other fixed length as a whole, so that the interpolating unit 31 uses the linear prediction coefficients thinned out before and after the linear prediction coefficients. It is generated by interpolation from the prediction coefficient.

  The linear prediction filter unit 32 generates a linear prediction filter from the linear prediction coefficient after interpolation. The adaptive code extracted from the adaptive codebook of the adaptive codebook search unit 34 and the noise code extracted from the noise codebook of the noise codebook search unit 35 are amplified by the amplifiers 36 and 37, respectively, and then combined by the combining unit 38. The synthesized signal is filtered by the generated linear prediction filter.

The error calculator 33 calculates an error between the filtered signal and the input audio signal. The speech encoding apparatus 300 outputs an index representing a noise code and an adaptive code when the error calculated by the error calculation unit 33 is minimized, and every other linear prediction coefficient as a code. In general, linear prediction coefficients are calculated at intervals of about 5 to 7 ms, but linear prediction coefficients in transmission codes are transmitted at intervals of 10 to 15 ms, and interpolation processing is often performed on the receiving side.
Speech coding standard JT-G729

  However, in the conventional speech coding method described above, the linear prediction coefficient generated by the interpolation process is calculated because the linear prediction coefficient is calculated in a unit of a fixed length and the interpolation process is performed regardless of the characteristics of the input speech. However, there are cases where it is not appropriate for the input voice (particularly during the transition period), leading to a reduction in sound quality.

  An object of the present invention is to improve sound quality by calculating a linear prediction coefficient of an audio signal in units of pitch waveforms.

In order to solve the above-mentioned problem, the invention described in claim 1 is a speech coding apparatus that encodes input speech using a linear prediction filter and an excitation signal, and extracts the pitch waveform from the input speech signal. And calculating a linear prediction coefficient from the input speech signal in units of the pitch waveform extracted by the extracting means, and when the pitch length of the pitch waveform is shorter than a predetermined value, a plurality of pitch waveforms Calculation means for calculating the linear prediction coefficient in a unit in which the linear prediction coefficients are calculated, generation means for generating the linear prediction filter using the linear prediction coefficient calculated by the calculation means, and the calculation means collects a plurality of pitch waveforms. Means for repeatedly extracting the excitation signal input to the linear prediction filter in accordance with the pitch length of each of the plurality of pitch waveforms. A selecting means for error between the speech signal excitation signal to a linear prediction filter that is generated is the the input synthesized speech which is calculated by the input as a driving signal by the generating means to select an excitation signal that minimizes the It is characterized by providing.

According to a second aspect of the invention, the speech coding apparatus according to claim 1, wherein the calculating means, when the pitch length is less than a pre-specified value, summarized pitch waveforms of a pre-specified number It is characterized by that.

Maximum The invention according to claim 3, in the speech coding apparatus according to claim 1, wherein the calculating means, when the pitch length is less than a pre-specified value, which does not exceed the fixed length specified in advance It is characterized by collecting pitch waveforms to a long length.

According to a fourth aspect of the present invention, in the speech coding apparatus according to any one of the first to third aspects, it is determined whether or not a change in the time axis of the pitch length of the pitch waveform is greater than a predetermined value. A determination means; and an interpolation means for performing interpolation processing of linear prediction coefficients thinned out at a predetermined interval when the change in the pitch length is determined to be equal to or less than a predetermined value by the determination means. When it is determined that the change in pitch length is larger than a predetermined value, linear prediction coefficient thinning and interpolation processing by the interpolation means are not performed.

According to a fifth aspect of the present invention, in the speech coding apparatus according to the fourth aspect , the determination unit determines whether or not there is a difference greater than a fixed value specified in advance between pitch lengths. It is said.

According to a sixth aspect of the present invention, in the speech coding apparatus according to the fourth aspect , the determination means has a difference between pitch lengths equal to or greater than a value obtained by multiplying a specific pitch length by a predetermined ratio. It is characterized by determining whether or not.

The speech encoding method according to claim 7 is a speech encoding method for encoding a speech signal using a linear prediction filter and an excitation signal, wherein a pitch waveform is extracted from the speech signal, and the extracted pitch waveform is extracted. The linear prediction coefficient is calculated from the speech signal in units, and when the pitch length of the pitch waveform is shorter than a predetermined value, the linear prediction coefficient is calculated in a unit of a plurality of pitch waveforms, and the calculation is performed. When the linear prediction filter is generated using the generated linear prediction coefficients and a plurality of pitch waveforms are collected, the excitation signal input to the linear prediction filter is matched with the pitch length of each of the plurality of pitch waveforms. Repeatedly extract and input the excitation signal as a drive signal to the generated linear prediction filter so that the error between the synthesized speech and the speech signal is minimized. It is characterized by selecting the excitation signal.

According to the present invention, by calculating a linear prediction coefficient from a speech signal in units of pitch waveforms, a transient waveform can be easily separated from a waveform for which linear prediction is performed, thereby improving the accuracy and sound quality of linear prediction. Ri is possible and do, the pitch length is less than a pre-specified value, performs a linear prediction analysis from together multiple pitch waveforms into one, and iterates the excitation signal in accordance with the pitch length of each Thus, the encoding efficiency can be improved.

  Furthermore, where there is a sudden change in pitch length, it is possible to efficiently find a place where deterioration is severe and improve sound quality by outputting the linear prediction coefficient without performing thinning and interpolation processing as it is. Become.

Embodiments of the present invention will be described below with reference to the drawings.
First, the configuration in the present embodiment will be described.

  FIG. 1 shows the configuration of a speech encoding apparatus 100 according to this embodiment. As shown in FIG. 1, the speech coding apparatus 100 includes a pitch extraction unit 1, a calculation position determination unit 2, a linear prediction analysis unit 3, an interpolation unit 4, a linear prediction filter unit 5, an error calculation unit 6, an adaptive codebook search. A unit 7, a noise codebook search unit 8, amplifiers 9 and 10, a synthesis unit 11, and an iterative processing unit 12.

The pitch extraction unit 1 extracts a pitch waveform from the input audio signal. The pitch is a unit for repeating a waveform as shown in FIG. When human speech waveforms are analyzed, repeated waveforms often exist continuously at a certain period (pitch period). Pitch extraction is a process of detecting a pitch period from an input audio signal and dividing the audio signal on the time axis for each pitch period. FIG. 3 shows a case where p _{0 to} p ₃ are extracted as pitch lengths.

  The method for detecting the pitch period is not particularly limited. For example, the method described in JP-A-11-45098 can be used. In this method, a voiced segment is extracted from an input voice signal, the extracted voiced segment is set as a processing target speech waveform, and a predetermined breakpoint on the time axis of the processing target speech waveform is used as a base point in advance. Copy the voice waveform of the defined section, translate the voice waveform of the copy section by a predetermined sampling point from the already determined breakpoint on the time axis, and copy section at each translation point for each translation Then, the correlation between the speech waveform and the processing target speech waveform is obtained, and the parallel movement point having the largest correlation is obtained as the next breakpoint. Then, the interval between the obtained break points is set as the pitch period of the processing target speech waveform.

  The calculation position determination unit 2 determines the calculation position of the linear prediction coefficient in order to calculate the linear prediction coefficient in pitch units.

ここで、｛ａ_i|i＝1,…,n｝が線形予測係数である。 The linear prediction analysis unit 3 performs linear prediction analysis of the input speech signal at the calculation position calculated by the calculation position determination unit 2 and calculates a linear prediction coefficient. The nth-order linear prediction filter is expressed as shown in Equation (1).

Here, {a _i | i = 1,..., N} is a linear prediction coefficient.

式（３）は、レビンソン・ダービンアルゴリズムを用いて解くことが可能である（非特許文献１参照）。 One method of linear prediction analysis is an autocorrelation method. In the autocorrelation method, for example, if m input signal samples used for linear prediction analysis are {s _i | i = 0,..., M−1}, the autocorrelation coefficient {r _i | i = 0,. , n} is calculated as shown in equation (2), and linear prediction coefficients {a _i | i = 1,..., n} that satisfy equation (3) are calculated.

Equation (3) can be solved using the Levinson-Durbin algorithm (see Non-Patent Document 1).

  When the pitch length is shorter than a predetermined value, linear prediction analysis is performed after combining a plurality of pitch waveforms into one in order to improve coding efficiency. As a simple method for combining a plurality of pitch waveforms, there is a method for combining a predetermined number of pitch waveforms into one. In this case, for example, if the pitch length is less than 3 ms, two pitches are combined. Further, pitch waveforms for a maximum length not exceeding a predetermined fixed length may be collected. That is, assuming that a fixed length designated in advance is c, the pitch waveforms are gathered by the maximum length below c.

  In this embodiment, in order to reduce the amount of generated code, linear prediction coefficients are encoded every other pitch. Therefore, the interpolation unit 4 interpolates the linear prediction coefficients to be thinned out from the preceding and following linear prediction coefficients. The linear prediction coefficient generated and interpolated is output to the linear prediction filter unit 5. As an interpolation method, for example, linear prediction coefficients before and after the linear prediction coefficient to be thinned out are converted into LSP (Line Spectrum Pair) coefficients, and the coefficients before and after the conversion are added and divided by 2, followed by linear prediction. Perform processing to return to coefficients.

At this time, the interpolation unit 4 compares the pitch lengths of the linear prediction coefficients before and after the linear prediction coefficient to be thinned out, determines whether the difference exceeds a predetermined fixed value, and determines the fixed value. If it exceeds, do not perform decimation. For example, the pitch length of the temporally previous pitch waveform thinning-out object and p _m-1, and the pitch length after temporally and p _{m + 1,} when a pre-designated fixed value is alpha, | p _{m When −1−} p _{m + 1} |> α, the thinning is not performed.

Whether the difference between the pitch lengths of the linear prediction coefficients before and after the linear prediction coefficient to be thinned out exceeds a specific ratio of a specific pitch length (a value obtained by multiplying a specific pitch length by a predetermined ratio). It may be determined whether or not thinning is performed when the predetermined ratio is exceeded. For example, the pitch length before the pitch waveform of the thinning target pitch waveform is set to _pm-1 , the pitch length after the time is set to _{pm + 1} , and a predetermined ratio is β (value of 0 to 1). Then, when | p _m-1 −p _{m + 1} |> p _m−1 · β, the thinning is not performed.

  The linear prediction filter unit 5 synthesizes (generates) a linear prediction filter from the linear prediction coefficient according to the equation (1). In addition, the linear prediction filter unit 5 performs filtering processing by the linear prediction filter on the adaptive code of the adaptive codebook in the adaptive codebook search processing (see FIG. 6), and outputs it to the error calculation unit 6. Further, the linear prediction filter unit 5 synthesizes the noise code of the noise codebook and the excitation signal finally obtained by the adaptive codebook search process (see FIG. 6) in the noise codebook search process (see FIG. 7). The processed signal is filtered by the linear prediction filter and output to the error calculation unit 6.

  The error calculation unit 6 calculates an error between the speech signal input to the speech encoding apparatus 100 and the signal after filtering processing in the linear prediction filter unit 5, and the adaptive codebook search unit 7 and the noise codebook search unit 8 is output.

  The adaptive codebook search unit 7 has an adaptive codebook storing the excitation signals used so far, takes out the adaptive code from the adaptive codebook, and performs filtering by the linear prediction filter unit 5 calculated by the error calculation unit 6 An adaptive code is selected such that the error between the processed adaptive code and the input speech signal is the smallest among the errors obtained so far (see FIG. 6). The adaptive codebook search unit 7 adds the finally obtained excitation signal to the adaptive codebook after the adaptive codebook search process (see FIG. 6) and the noise codebook search process (see FIG. 7). Then, the adaptive codebook is updated.

  The noise codebook search unit 8 has a noise codebook storing a white noise signal (noise code), extracts a noise code from the noise codebook, and performs filtering by the linear prediction filter unit 5 calculated by the error calculation unit 6. A noise signal is selected such that the error between the processed signal and the input voice signal is the smallest among the errors obtained so far (see FIG. 7). Here, the signal to be filtered by the linear prediction filter unit 5 is a signal obtained by adding the excitation signal determined in the adaptive codebook search process to the noise code extracted from the noise codebook.

  The amplifiers 9 and 10 amplify (adjust) the amplitude values of the adaptive code extracted from the adaptive codebook and the noise code extracted from the noise codebook, respectively, with a predetermined amplification factor.

  The synthesizer 11 synthesizes the amplified noise code extracted from the noise codebook and the amplified adaptive code determined as the excitation signal in the adaptive codebook search process.

  When the plurality of pitch waveforms are combined into one as described above, the iterative processing unit 12 adjusts the input excitation signal to the pitch length of each of the plurality of pitch waveforms as shown in FIG. The iterative process is performed, and the excitation signal after the iterative process is output to the linear prediction filter unit 5. FIG. 4 shows an example in which the excitation signal is repeated in accordance with each pitch length when the pitch waveform with the pitch length 1 and the pitch waveform with the pitch length 2 are combined into one.

  The speech coding apparatus 100 configured as described above has an adaptive codebook index and a noise codebook index finally obtained as excitation signals in the adaptive codebook search process and the noise codebook search process, and one pitch. Every other thinned linear prediction coefficient, a signal representing the pitch length, and a signal representing the amplification factor in the amplifiers 9 and 10 are output as encoded signals.

  FIG. 2 shows a configuration of speech decoding apparatus 200 according to the embodiment of the present invention. Speech decoding apparatus 200 is an apparatus for decoding the signal encoded by speech encoding apparatus 100. As shown in FIG. 2, adaptive codebook search unit 21, noise codebook search unit 22, amplifier 23, 24, a synthesis unit 25, an iterative processing unit 26, an interpolation unit 27, and a linear prediction filter unit 28.

  The adaptive codebook search unit 21 searches and extracts an adaptive code corresponding to the index of the input adaptive codebook from the adaptive codebook, and outputs it to the amplifier 23.

  The noise codebook search unit 22 extracts a noise code corresponding to the input index of the noise codebook from the noise codebook and outputs it to the amplifier 24.

  The amplifiers 23 and 24 amplify the input adaptive code and noise code, respectively, and output them to the synthesis unit 25. The synthesizer 25 synthesizes the adaptive code and the noise code input from the amplifiers 23 and 24, respectively.

  The iterative processing unit 26 generates an excitation signal by repeating the input adaptive code and noise code according to the input pitch length, and outputs the excitation signal to the linear prediction filter unit 28.

  When there is a linear prediction coefficient that has not been input as an encoded signal, the interpolation unit 27 performs interpolation processing on the linear prediction coefficient, and outputs the linear prediction coefficient after the interpolation processing to the linear prediction filter unit 28. For the interpolation processing in the interpolation unit 27, a method similar to the interpolation method in the interpolation unit 4 can be applied.

  The linear prediction filter unit 28 synthesizes (generates) a linear prediction filter from the input linear prediction coefficient according to Equation (1), and performs filtering processing on the input excitation signal using the generated linear prediction filter. To generate and output a synthesized speech.

Next, the operation in this embodiment will be described.
First, with reference to the flowchart of FIG. 5, the speech encoding process executed in speech encoding apparatus 100 will be described.

  First, a pitch waveform is extracted from the speech signal input to speech coding apparatus 100 (step S1), and a calculation position of a linear prediction coefficient is determined based on the extracted pitch waveform (step S2). At this time, if the pitch length is shorter than the specified value, a plurality of pitch waveforms are combined into one. Next, linear prediction analysis is performed on the input speech signal at the calculation position calculated in step S2, and a linear prediction coefficient is calculated (step S3). Next, a linear prediction coefficient to be thinned out is generated by interpolating from the preceding and subsequent linear prediction coefficients (step S4). Next, a linear prediction filter is synthesized from the linear prediction coefficients according to equation (1) (step S5).

  Next, adaptive codebook search processing and noise codebook search processing for searching for an excitation signal that minimizes an error from the input speech signal from the adaptive codebook and noise codebook are performed (steps S6 and S7). The adaptive codebook search process in step S6 and the noise codebook search process in step S7 will be described in detail later with reference to FIGS. 6 and 7, respectively.

  When the adaptive codebook search process and the noise codebook search process are completed, the excitation signal obtained by these processes is added to the adaptive codebook to update the adaptive codebook (step S8), and the adaptation representing the excitation signal The codebook index, the noise codebook index, the linear prediction coefficient thinned out every other pitch, and a signal representing the pitch length are output as encoded signals, and the speech encoding process ends.

  Next, the adaptive codebook search process (step S6 in FIG. 5) will be described with reference to the flowchart in FIG.

  First, the first adaptive code is extracted from the adaptive codebook and set as the processing target adaptive code (step S11). Next, it is determined whether or not the processing for all adaptive codes in the adaptive codebook has been completed (step S12). If it is determined in step S12 that the process has not been completed (step S12; NO), when a plurality of pitches are combined into one, the current adaptive code is repeated according to each pitch length. The repetitive process of taking out is performed (step S13).

  Next, a filtering process using the linear prediction filter synthesized in step S5 is performed on the current adaptive code to be processed (step S14), and an error between the filtered signal and the input speech signal is calculated. (Step S15). Next, it is determined whether or not the error calculated in step S15 is the smallest among the errors obtained after the start of the main search process (step S16).

  If it is determined in step S16 that the error is not minimum (step S16; NO), the next adaptive code in the adaptive codebook is set as a processing target (step S18), and steps S12 to S12 are performed for the adaptive code. The process of S17 is repeated.

  If it is determined in step S16 that the error is the minimum (step S16; YES), the adaptive code to be processed is set as an excitation signal candidate (step S17). Next, the next adaptive code in the adaptive codebook is set as a processing target (step S18), and the processes of steps S12 to S17 are repeated for the adaptive code.

  When the processing of steps S13 to S17 is completed for all the adaptive codes in the adaptive codebook (step S12; YES), the adaptive codebook search process is terminated, and finally the index of the adaptive code remaining as the excitation signal candidate is code. Is selected as the data of the digitized signal.

  Next, the noise codebook search process (step S7 in FIG. 5) will be described with reference to the flowchart in FIG.

  First, the first noise code is extracted from the noise code book and set as a noise code to be processed (step S21). Next, it is determined whether or not the processing for all noise codes in the noise codebook has been completed (step S22). If it is determined in step S22 that the process has not been completed (step S22; NO), when a plurality of pitches are combined into one, the current noise code is repeated according to each pitch length. An iterative process is performed (step S23).

  Next, the adaptive code finally set as the excitation signal in the adaptive codebook search process of FIG. 6 and the current noise code to be processed are synthesized (step S24), and the synthesized signal is synthesized in step S5. A filtering process using the linear prediction filter is performed (step S25). Then, an error between the filtered signal and the input audio signal is calculated (step S26). Next, it is determined whether or not the error calculated in step S26 is the smallest among the errors obtained after the start of the main search process (step S27).

  If it is determined in step S27 that the error is not minimum (step S27; NO), the next noise code in the noise codebook is set as a processing target (step S29), and steps S22 to S22 are performed on the noise code. The process of S28 is repeated.

  If it is determined in step S27 that the error is minimum (step S27; YES), the current noise code to be processed is set as an excitation signal candidate (step S28). Next, the next noise code of the noise code book is set as a processing target (step S29), and the processes of steps S22 to S28 are repeated for the noise code.

  When the processing of steps S23 to S28 is completed for all the noise codes of the noise codebook (step S22; YES), the noise codebook search process is terminated, and finally the noise code index remaining as the excitation signal candidate is encoded. Is selected as the data of the digitized signal.

  Next, speech decoding processing executed in the speech decoding apparatus 200 will be described with reference to the flowchart of FIG.

  First, an adaptive code corresponding to the index of the adaptive codebook included in the input encoded signal is extracted from the adaptive codebook (step T1), and the noise codebook included in the encoded signal is extracted from the noise codebook. Is extracted (step T2).

  Next, an excitation signal is generated by repeating an adaptive code and a noise code according to the input pitch length (step T3), and the generated excitation signal is added to the adaptive codebook. The book is updated (step T4).

  Next, when there is a linear prediction coefficient that has not been input as an encoded signal, interpolation processing of the linear prediction coefficient is performed (step T5). Next, a linear prediction filter is synthesized from the linear prediction coefficient according to the equation (1) (step T6).

  Next, the reproduced speech is synthesized by applying filtering processing to the created excitation signal using the linear prediction filter synthesized in step T6 (step T7), and the speech decoding processing is completed.

  As described above, according to the speech coding apparatus 100 and the speech decoding apparatus 200 of the present embodiment, by calculating the linear prediction coefficient of the speech signal in units of pitch waveforms, the waveform in the transient period is calculated from the waveform in which linear prediction is performed. Can be easily separated, and the accuracy and sound quality of linear prediction can be improved.

  In addition, when the pitch length is shorter than a predesignated value, a plurality of pitch waveforms are collected as one pitch waveform and then linear prediction analysis is performed, and by repeating the excitation signal according to each pitch length, Encoding efficiency can be improved.

  Furthermore, where there is a sudden change in the pitch length, it is possible to efficiently find a place where deterioration is severe and improve the sound quality by encoding as it is without performing thinning and interpolation processing of the linear prediction coefficient. It becomes.

The block diagram which shows the structure of the audio | voice coding apparatus which concerns on embodiment of this invention. The block diagram which shows the structure of the audio | voice decoding apparatus which concerns on embodiment of this invention. The figure which shows the waveform of an audio | voice signal, and the pitch extracted from an audio | voice signal. The figure which shows an excitation signal and the excitation signal repeated with pitch length. The flowchart which shows the audio | voice encoding process performed in the audio | voice encoding apparatus of this embodiment. The flowchart which shows an adaptive codebook search process. The flowchart which shows a noise codebook search process. The flowchart which shows the audio | voice decoding process performed in the audio | voice decoding apparatus of this embodiment. The block diagram which shows the structure of the conventional audio | voice encoding apparatus.

Explanation of symbols

1 Pitch extraction unit (extraction means)
2 Calculation position determination unit 3 Linear prediction analysis unit (calculation means)
4 Interpolation section (judgment means)
5 Linear prediction filter section (generation means)
6 Error calculation unit 7 Adaptive codebook search unit (selection means)
8 Noise codebook search section (selection means)
9, 10, 23, 24 Amplifier 11 Combining unit 12 Repetition processing unit (repetition means)
21 Adaptive Codebook Search Unit 22 Noise Codebook Search Unit 25 Synthesis Unit 26 Iterative Processing Unit 27 Interpolation Unit 28 Linear Prediction Filter Unit (Output Unit)
100 Speech coding apparatus 200 Speech decoding apparatus

Claims (7)

A speech encoding device that encodes input speech using a linear prediction filter and an excitation signal,
Extraction means for extracting a pitch waveform from the input audio signal;
A linear prediction coefficient is calculated from the input speech signal in units of the pitch waveform extracted by the extracting means, and when the pitch length of the pitch waveform is shorter than a predetermined value, a plurality of pitch waveforms are combined. Calculating means for calculating the linear prediction coefficient in units of
Generating means for generating the linear prediction filter using the linear prediction coefficient calculated by the calculating means;
When the calculation unit summarizes a plurality of pitch waveforms, a repetition unit that repeatedly extracts the excitation signal input to the linear prediction filter in accordance with the pitch length of each of the plurality of pitch waveforms;
A selection means for selecting an excitation signal that minimizes an error between the synthesized speech calculated by inputting the excitation signal as a drive signal to the linear prediction filter generated by the generation means and the input speech signal;
A speech encoding apparatus comprising:   2. The speech encoding apparatus according to claim 1, wherein, when the pitch length is shorter than a predetermined value, the calculating unit collects a predetermined number of pitch waveforms. 3.   2. The speech encoding apparatus according to claim 1, wherein, when the pitch length is shorter than a predesignated value, the calculation unit summarizes the pitch waveform up to a maximum length not exceeding a predesignated fixed length. . Determination means for determining whether or not a change in the time axis of the pitch length of the pitch waveform is greater than a predetermined value;
Interpolating means for performing interpolation processing of linear prediction coefficients thinned out at predetermined intervals when the determining means determines that the change in pitch length is equal to or less than a predetermined value;
4. If the determination means determines that the change in pitch length is greater than a predetermined value, linear prediction coefficient thinning and interpolation processing by the interpolation means are not performed. The speech encoding device according to item.   5. The speech encoding apparatus according to claim 4, wherein the determination unit determines whether there is a difference equal to or greater than a predetermined fixed value between pitch lengths.   5. The speech encoding apparatus according to claim 4, wherein the determination unit determines whether there is a difference between pitch lengths that is equal to or greater than a value obtained by multiplying a specific pitch length by a predetermined ratio. . A speech encoding method for encoding a speech signal with a linear prediction filter and an excitation signal,
Extracting a pitch waveform from the audio signal;
A linear prediction coefficient is calculated from the speech signal in units of the extracted pitch waveform, and when the pitch length of the pitch waveform is shorter than a predetermined value, the linear prediction is in units of a plurality of pitch waveforms. Calculate the coefficient,
Generating the linear prediction filter using the calculated linear prediction coefficient;
When collecting a plurality of pitch waveforms, the excitation signal input to the linear prediction filter is repeatedly extracted according to the pitch length of each of the plurality of pitch waveforms,
A speech encoding method, wherein an excitation signal that minimizes an error between a synthesized speech calculated by inputting an excitation signal as a drive signal to the generated linear prediction filter and the speech signal is selected.

Images