FORM 2 THE PATENTS ACT, 1970 (39 of 1970) & THE PATENTS RULES, 2003 COMPLETE SPECIFICATION [See section 10, Rule 13] SPECTRAL SMOOTHING DEVICE, ENCODING DEVICE, DECODING DEVICE, COMMUNICATION TERMINAL DEVICE, BASE STATION DEVICE, AND SPECTRAL SMOOTHING METHOD; PANASONIC CORPORATION, A CORPORATION ORGANIZED AND EXISTING UNDER THE LAWS OF JAPAN, WHOSE ADDRESS IS 1006, OAZA KADOMA, KADOMA-SHI, OSAKA 5718501, JAPAN THE FOLLOWING SPECIFICATION PARTICULARLY DESCRIBES THE INVENTION AND THE MANNER IN WHICH IT IS TO BE PERFORMED. DESCRIPTION Technical Field The present invention relates to a spectrum smoothing apparatus, a coding apparatus, a decoding apparatus, a communication terminal apparatus, a base station apparatus and a spectrum smoothing method smoothing spectrum of speech signals. Background Art When speech/audio signals are transmitted in a packet communication system typified by Internet communication and a mobile communication system, a compression/coding technique is often used to improve the transmission rate of speech/audio signals. Furthermore, in recent years, in addition to a demand for simply encoding speech/audio signals at low bit rates, there is an increasing demand for a technique to encode speech/audio signals in high quality. To meet this demand, studies are underway to develop various techniques to perform orthogonal transformation (i.e. time-frequency transformation) of a speech signal to extract frequency components (i.e. spectrum) of the speech signal and apply various processing such as linear transformation and non-linear transformation to the calculated spectrum to improve the quality of the decoded signal (see, for example, patent literature 1). According to the method disclosed in patent literature 1, first, a frequency spectrum contained in a speech signal of a certain time length is analyzed, and then non-linear transformation processing to emphasize greater spectrum power values is applied to the analyzed spectrum. Next, linear smoothing processing for the spectrum subjected to non-linear transformation processing, is performed in the frequency domain. After this, inverse non-linear transformation processing is performed to cancel non-linear transformation characteristics, and, furthermore, inverse smoothing processing is performed to cancel smoothing characteristics, so that noise components included in the speech signal over the entire band are suppressed. Thus, with the method disclosed in patent literature 1, all samples of a spectrum acquired from a speech signal are subjected to non-linear transformation processing and then the spectrum is smoothed, so that the speech signal is acquired in good quality. Patent literature 1 introduces transformation methods such as power transform and logarithmic transform as examples of non-linear processing. Citation List Patent Literature PTL 1 Japanese Patent Application Laid-Open No. 2002-244695 PTL 2 WO 2007/037361 Non-Patent Literature Yuichiro TAKAMIZAWA, Toshiyuki NOMURA and Masao IKEKAWA, "High-Quality and Processor-Efficient Implementation of and MPEG-2 AAC Encoder", IEICE TRANS. INF. &SYST., VOL.E86-D, No.3 MARCH 2003 Summary of Invention Technical Problem However, with the method disclosed in patent literature 1, non-linear transformation processing needs to be performed for all samples of a spectrum acquired from a speech signal, and therefore there is a problem that the amount of calculation processing is enormous. Furthermore, if only part of samples of a spectrum are extracted to reduce the amount of calculation processing, sufficiently high speech quality cannot be always achieved by simply performing spectrum smoothing after non-linear transformation. Based upon a configuration for performing non-linear transformation of a spectrum value calculated from a speech signal and then smoothing the spectrum, it is an object of the present invention to provide a spectrum smoothing apparatus, a coding apparatus, a decoding apparatus, a communication terminal apparatus, a base station apparatus and a spectrum smoothing method, whereby good speech quality is maintained and the amount of calculation processing can be reduced substantially. Solution to Problem The spectrum smoothing apparatus according to the present invention employs a configuration to include: a time-frequency transformation section that performs a time-frequency transformation of an input signal and generates a frequency component; a subband dividing section that divides the frequency component into a plurality of subbands; a representative value calculating section that calculates a representative value of each divided subband by calculating an arithmetic mean and by using a multiplication calculation using a calculation result of the arithmetic mean; a non-linear transformation section that performs a non-linear transformation of representative values of the subbands; and a smoothing section that smoothes the representative values subjected to the non-linear transformation in the frequency domain. The spectrum smoothing method according to the present invention includes: a time-frequency transformation step of performing a time-frequency transformation of an input signal and generates a frequency component; a subband division step of dividing the frequency component into a plurality of subbands; a representative value calculation step of calculating a representative value of each divided subband by calculating an arithmetic mean and by using a multiplication calculation using a calculation result of the arithmetic mean; a non-linear transformation step of performing a non-linear transformation of representative values of the subbands; and a smoothing step of smoothing the representative values subjected to the non-linear transformation in the frequency domain. Advantageous Effects of Invention With the present invention, it is possible to maintain good speech quality and reduce the amount of calculation processing substantially. Brief Description of Drawings FIG.l provides spectrum overviews showing an overview of processing according to embodiment 1 of the present invention; FIG.2 is a block diagram showing a principal-part configuration of a spectrum smoothing apparatus according to embodiment 1; FIG.3 is a block diagram showing a principal-part configuration of a representative value calculating section according to embodiment 1; FIG.4 is an overview showing a configuration of subbands and subgroups of an input signal according to embodiment 1; FIG.5 is a block diagram showing a configuration of a communication system having a coding apparatus and decoding apparatus according to embodiment 2 of the present invention; FIG.6 is a block diagram showing an inner principal-part of the coding apparatus according to embodiment 2 shown in FIG.5; FIG.7 is a block diagram showing an inner principal-part configuration of the second layer coding section according to embodiment 2 shown in FIG.6; FIG.8 is a block diagram showing a principal-part configuration of the spectrum smoothing apparatus according to embodiment 2 shown in FIG.7; FIG.9 shows a diagram for explaining the details of the filtering processing in the filtering section according to embodiment 2 shwon in FIG.7; FIG.10 is a flowchart for explaining the steps of processing for searching for optimal pitch coefficient Tp' with respect to subband SBP in the search section according to embodiment 2 shwon in FIG.7; FIG. 11 is a block diagram showing an inner principal-part configuration of the decoding apparatus according to embodiment 2 shown in FIG.5; and FIG. 12 is a block diagram showing an inner principal-part configuration of the second layer decoding section according to embodiment 2 shown in FIG. 11. Description of Embodiments Embodiments of the present invention will be described in detail with reference to the accompanying drawings. (Embodiment 1) First, an overview of the spectrum smoothing method according to an embodiment of the present invention will be described using FIG. 1. FIG. 1 shows spectrum diagrams for explaining an overview of the spectrum smoothing method according to the present embodiment. FIG. 1A shows a spectrum of an input signal. With the present embodiment, first, an input signal spectrum is divided into a plurality of subbands. FIG. 1B shows how an input signal spectrum is divided into a plurality of subbands. The spectrum diagram of FIG. 1 is for explaining an overview of the present invention, and the present invention is by no means limited to the number of subbands shown in the drawing. Next, a representative value of each subband is calculated. To be more specific, samples in a subband are further divided into a plurality of subgroups. Then, an arithmetic mean of absolute spectrum values is calculated per subgroup. Next, a geometric mean of the arithmetic mean values of individual subgroups is calculated per subband. This geometric mean value is not an accurate geometric mean value yet, and, at this point, a value that is obtained by simply multiplying individual groups' arithmetic mean values may be calculated, and an accurate geometric mean value may be found after non-linear transformation (described later). The above processing is to reduce the amount of calculation processing, and it is equally possible to find an accurate geometric mean value at this point. A geometric mean value found this way may be used as a representative value of each subband. FIG.1C shows representative values of individual subbands over an input signal spectrum shown with dotted lines. For ease of explanation, FIG.1C shows accurate geometric mean values as representative values, instead of values obtained by simply multiplying arithmetic mean values of individual subgroups. Next, referring to each subband's representative value, non-linear transformation (for example, logarithmic transform) is performed for a spectrum of an input signal such that greater spectrum power values are emphasized, and then smoothing processing is performed in the frequency domain. Afterward, inverse non-linear transformation (for example, inverse logarithmic transform) is performed, and a smoothed spectrum is calculated in each subband. FIG.1D shows a smoothed spectrum of each subband over an input signal spectrum shown with dotted lines. By means of this processing, it is possible to perform spectrum smoothing in the logarithmic domain while reducing speech quality degradation and reducing the amount of calculation processing substantially. Now, a configuration of a spectrum smoothing apparatus providing the above advantage, according to an embodiment of the present invention, will be described. The spectrum smoothing apparatus according to the present embodiment smoothes an input spectrum, and outputs the spectrum after the smoothing (hereinafter "smoothed spectrum") as an output signal. To be more specific, the spectrum smoothing apparatus divides an input signal every N samples (where N is a natural number), and performs smoothing processing per frame using N samples as one frame. Here, an input signal that is subject to smoothing processing is represented as "xn" (n=0, ...,N-1). FIG.2 shows a principal-part configuration of spectrum smoothing apparatus 100 according to the present embodiment. Spectrum smoothing apparatus 100 shown in FIG.2 is primarily formed with time-frequency transformation processing section 101, subband dividing section 102, representative value calculating section 103, non-linear transformation section 104, smoothing section 105 and inverse non-linear transformation section 106. Time-frequency transformation processing section 101 applies a fast Fourier transform (FFT) to input signal xn and finds a frequency component spectrum Sl(k) (hereinafter "input spectrum"). Then, time-frequency transformation processing section 101 outputs input spectrum Sl(k) to subband dividing section 102. Subband dividing section 102 divides input spectrum Sl(k) received as input from time-frequency transformation processing section 101, into P subbands (where P is an integer equal to or greater than 2). Now, a case will be described below where subband dividing section 102 divides input spectrum Sl(k) such that each subband contains the same number of samples. The number of samples may vary between subbands. Subband dividing section 102 outputs the spectrums divided per subband (hereinafter "subband spectrums"), to representative value calculating section 103. Representative value calculating section 103 calculates a representative value for each subband of an input spectrum divided into subbands, received as input from subband dividing section 102, and outputs the representative value calculated per subband, to non-linear transformation section 104. The processing in representative value calculating section 103 will be described in detail later. FIG.3 shows an inner configuration of representative value calculating section 103. Representative value calculating section 103 shown in FIG.3 has arithmetic mean calculating section 201, and geometric mean calculating section 202. First, subband dividing section 102 outputs a subband spectrum to arithmetic mean calculating section 201. Arithmetic mean calculating section 201 divides each subband of the subband spectrum received as input into Q subgroups of subgroup 0, subgroup Q-l, etc. (where Q is an integer equal to or greater than 2). Now, a case will be described below where Q subgroups are each formed with R samples (R is an integer equal to or greater than 2). Although a case will be described below where Q subgroups are all formed with R samples, the number of samples may vary between subgroups. FIG.4 shows a sample configuration of subbands and subgroups. FIG.4 shows, as an example, a case where the number of samples to constitute one subband is eight, the number of subgroups Q to constitute one subband is two and the number of samples R in one subgroup is four. Next, for each of the Q subgroups, arithmetic mean calculating section 201 calculates an arithmetic mean of the absolute values of the spectrums (FFT coefficients) contained in each subgroup, using equation In equation 1, AVElq is an arithmetic mean of the absolute values of the spectrums contained in subgroup q, and BSq is the index of the leading sample in subgroup q. Next, arithmetic mean calculating section 201 outputs arithmetic mean value spectrums calculated per subband, AVElq (q=0~Q-l) (subband arithmetic mean value spectrums), to geometric mean calculating section 202. Geometric mean calculating section 202 multiplies arithmetic mean value spectrums AVElq (q=0~Q-l) of all subbands received as input from arithmetic mean calculating section 201, as shown in equation 2, and calculates a representative spectrum, AVE2P (p=0~P-l), for each subband. In equation 2, P is the number of subbands. Next, geometric mean calculating section 202 outputs calculated subband representative value spectrums AVE2P (p=0~P-l) to non-linear transformation section 104. Non-linear transformation section 104 applies non-linear transformation having a characteristic of emphasizing greater representative values, to subband representative value spectrums AVE2P, received as input from geometric mean calculating section 202, using equation 3, and calculates first subband logarithmic representative value spectrums, AVE3P (p=0~.P-l). A case will be described here where logarithmic transform is performed as non-linear transformation processing. Next, a second subband logarithmic representative value spectrum, AVE4P (p=0~P-l), is calculated by multiplying calculated first subband logarithmic representative value spectrum, AVE3P (p=0~P-l) by the reciprocal of the number of subgroups, Q, using equation 4. Although in the processing of equation 2 in geometric mean calculating section 202 subband arithmetic mean value spectrums AVElp of individual subbands are simply multiplied, in the processing of equation 4 in non-linear transformation section 104, a geometric mean is calculated. With the present embodiment, transformation into the logarithmic domain is performed using equation 3, and then multiplication by the reciprocal of the number of subgroups, Q, is performed using equation 4. By this means, radical root calculation, which involves a large amount of calculation, can be replaced by simple division. Furthermore, when the number of subgroups, Q, is a constant, the radical root calculation can be replaced by simple multiplication, by calculating the reciprocal of Q in advance, so that the amount of calculation can be reduced further. Next, non-linear transformation section 104 outputs second subband logarithmic representative value spectrums AVE4P (p=0~P-l) calculated using equation 4, to smoothing section 105. Referring back to FIG.2 again, smoothing section 105 smoothes second subband logarithmic representative value spectrums AVE4P (p=0~P-l) received as input from non-linear transformation section 104, in the frequency domain, using equation 5, and calculates logarithmic smoothed spectrums AVE5P (p=0~P-l). Equation 5 represents smoothing filtering processing, and, in this equation 5, MALEN is the order of smoothing filtering and W1 is the smoothing filter weight. Furthermore, in equation 5 provides a method of calculating a logarithmic smoothed spectrum when subband index p is p> =(MA_LEN-l)/2 and p<=P-l-(MA_LEN-l)/2. When subband index p is at the top or near the last, spectrums are smoothed using equation 6 and equation 7 taking into account the boundary conditrions. Furthermore, smoothing section 105 performs smoothing based on simple moving average, as smoothing processing by smoothing filtering processing, as described above (when W; is 1 for all i's, smoothing is performed based on moving average). For the window function (weight), Hanning window or other window functions may be used. Next, smoothing section 105 outputs calculated smoothed spectrums AVE5P (p=0~P-l) to inverse non-linear transformation section 106. Inverse non-linear transformation section 106 performs inverse logarithmic transformation as inverse non-linear transformation for logarithmic smoothed spectrums AVE5P (p=0~P-l) received as input from smoothing section 105. Inverse non-linear transformation section 106 performs inverse logarithmic transformation for logarithmic smoothed spectrums AVE5P (p=0~P-l) using equation 8, and calculates smoothed spectrum AVE6P (p=0~P-l). Furthermore, inverse non-linear transformation section 106 calculates a smoothed spectrum of all samples using the values of samples in each subband as the values of linear domain smoothed spectrum AVE6P (p=0~P-l). Inverse non-linear transformation section 106 outputs the smoothed spectrum values of all samples as a processing result of spectrum smoothing apparatus 100. The spectrum smoothing apparatus and spectrum smoothing method according to the present invention have been described. As described above, with the present embodiment, subband dividing section 102 divides an input spectrum into a plurality of subbands, representative value calculating section 103 calculates representative value per subband using an arithmetic mean or geometric mean, non-linear transformation section 104 performs non-linear transformation having a characteristic of emphasizing greater values to each representative value, and smoothing section 105 smoothes representative values subjected to non-linear transformation per subband in the frequency domain. Thus, all samples of a spectrum are divided into a plurality of subbands, and, for each subband, a representative value is found by combining an arithmetic mean with multiplication calculation or geometric mean, and then smoothing is performed after the representative value is subjected to non-linear transformation, so that it is possible to maintain good speech quality and reduce the amount of calculation processing substantially. As described above, the present invention employs a configuration for calculating representative values of subbands by combining arithmetic means and geometric means of samples in subbands, so that it is possible to prevent speech quality degradation that can occur due to the variation of the scale of sample values in a subband when average values in the linear domain are used simply as representative values of subbands. Although the fast Fourier transform (FFT) has been explained as an example of time-frequency transformation processing with the present embodiment, the present invention is by no means limited to this, and other time-frequency transformation methods besides the fast Fourier transform (FFT) are equally applicable. For example, according to patent literature 1, upon calculation of perceptual masking values (see FIG.2), the modified discrete cosine transform (MDCT), not the fast Fourier transform (FFT), is used to calculate frequency components (spectrum). Thus, the present invention is applicable to configurations using the modified discrete cosine transform (MDCT) and other time-frequency transformation methods in a time-frequency transformation processing section. In the configuration described above, geometric mean calculating section 202 multiplies an arithmetic mean value spectrum AVElq (q=0~Q-l), and does not calculate radical roots. That is to say, strictly speaking, geometric mean calculating section 202 does not calculate geometric mean values, because, as explained above, in non-linear transformation section 104, transformation into the logarithmic domain is performed using equation 3 as non-linear transformation processing and then multiplication by the reciprocal of the number of subgroups Q is performed using equation 4, so that it is possible to replace radical root calculation by simple division (multiplication) and consequently reduce the amount of calculation. Consequently, the present invention is not necessarily limited to the above configuration. The present invention is equally applicable to, for example, a configuration for multiplying, in geometric mean calculating section 202, arithmetic mean value spectrums AVElq (q=0~Q-l) by the values of arithmetic mean value spectrums per subband, and then calculating a radical root of the number of subgroups and outputting the calculated radical root to non-linear transformation section 104 as subband representative value spectrums AVE2P (p=0~P-l). Either way, smoothing section 105 is able to acquire a representative value having been subjected to non-linear transformation, per subband. In this case, the calculation of equation 4 in non-linear transformation section 104 may be omitted. A case has been described above with the present embodiment where a representative value of each subband is calculated by, first, calculating an arithmetic mean value of a subgroup, and next finding a geometric mean value of the arithmetic mean values of all subgroups in a subband. However, the present invention is by no means limited to this and is equally applicable to a case where, for example, the number of samples to constitute a subgroup is one, that is, a case where a geometric mean value of all samples in a subband is used as a representative value of the subband without calculating an arithmetic mean value of each subgroup. In this configuration again, as described above, rather than calculating an accurate geometric mean value, it is possible to calculate a geometric mean value in the logarithmic domain by performing non-linear transformation and then performing multiplication by the reciprocal of the number of subgroups. In the above description, all samples in a subband have the same spectrum value in inverse non-linear transformation section 106. However, the present invention is by no means limited to this, and it is equally possible to provide an inverse smoothing processing section after inverse non-linear transformation section 106 so that the inverse smoothing processing section may assign weight to samples in each subband and perform inverse smoothing processing. This inverse smoothing processing needs not be completely opposite to smoothing section 105. Although a case has been described with the above description where non-linear transformation section 104 performs inverse logarithmic transformation as inverse non-linear transformation processing and inverse non-linear transformation section 106 performs inverse logarithmic transformation as inverse non-linear transformation processing, this is by no means limiting, and it is equally possible to use power transform and others and perform inverse processing of non-linear transformation as inverse non-linear transformation processing. However, given that calculation of a radical root can be replaced by simple division (multiplication) by multiplying the reciprocal of the number of subgroups Q using equation 4, the fact that non-linear transformation section 104 performs logarithmic transform as non-linear transformation, should be credited for the reduction of the amount of calculation. Consequently, if processing that is different from logarithmic transform is performed as non-linear transformation processing, it is then equally possible to calculate a representative value per subband by calculating a geometric mean value of arithmetic mean values of subgroups and apply non-linear processing to the representative values. Furthermore, as for the number of subbands and the number of subgroups, if, for example, the sampling frequency of an input signal is 32 kHz and one frame is 20 msec long, that is, if an input signal is comprised of 640 samples, it is possible to, for example, set the number of subbands to eighty, the number of subgroups to two, the number of samples per subgroup to four, and the order of smoothing filtering to seven, for example. The present invention is by no means limited to this setting and is equally applicable to cases where different values are applied. The spectrum smoothing apparatus and spectrum smoothing method according to the present invention are applicable to any and all of spectrum smoothing devices or components that perform smoothing in the spectral domain, including speech coding apparatus and speech coding method, speech decoding apparatus and speech decoding method, and speech recognition apparatus and speech recognition method. For example, although, with the bandwidth enhancement technique disclosed in patent literature 2, processing for calculating a spectral envelope from LPCs (Linear Predictive Coefficients), and, based on this calculated spectral envelope, removing the spectral envelope from the lower band spectrum, is used to calculate parameters for generating a higher band spectrum, it is equally possible to use a smoothed spectrum calculated by applying the spectrum smoothing method according to the present invention to a lower band spectrum instead of the spectral envelope used in spectral envelope removing processing in patent literature 2. Furthermore, although a configuration has been explained with the present embodiment where an input spectrum Sl(k) is divided into P subbands (where P is an integer equal to or greater than 2) all having the same number of samples, the present invention is by no means limited to this and is equally applicable to a configuration in which the number of samples varies between subbands. Fro example, a configuration is possible in which subbands are divided such that a subband on the lower band side has a smaller number of samples and a subband on the higher band side has a greater number of samples. Generally speaking, in human perception, frequency resolution decreases in the higher band side, so that more efficient spectrum smoothing is made possible with the above configuration. The same applies to subgroups to constitute each subband. Although a case has been described above with the present embodiment where Q subgroups are all formed with R samples, the present invention is by no means limited to this, and is equally applicable to configurations where subgroups are divided such that a subgroup on the lower band side has a smaller number of samples and a subgroup on the higher band side has a larger number of samples. Although weighted moving average has been described as an example of smoothing processing with the present embodiment, the present invention is by no means limited to this and is equally applicable to various smoothing processing. For example, as described above, in a configuration in which the number of samples varies between subbands (that is, the number of samples increases in the higher band), it is possible to make the number of taps in a moving average filter not the same between the left and the right and increase the number of taps in the higher band. When the number of samples increases in subbands in the higher band, it is possible to perform perceptually more adequate smoothing processing by using a moving average filter having a small number of taps in the higher band side. The present invention is applicable to cases using a moving average filter that is asymmetrical between the left and the right and has a greater number of taps on the higher band side. (Embodiment 2) A configuration will be described now with the present embodiment where the spectrum smoothing processing explained with embodiment 1 is used in preparatory processing upon band enhancement coding disclosed in patent literature 2. FIG.5 is a block diagram showing a configuration of a communication system having a coding apparatus and decoding apparatus according to embodiment 2. In FIG.5, the communication system has a coding apparatus and decoding apparatus that are mutually communicable via a transmission channel. The coding apparatus and decoding apparatus are usually mounted in a base station apparatus and communication terminal apparatus for use. Coding apparatus 301 divides an input signal every N samples (where N is a natural number) and performs coding on a per frame basis using N samples as one frame. The input signal to be subject to coding is represented as xn (n=0, ..., N-l). n is the (n+l)-th signal component in the input signal divided every N samples. Input information having been subjected to coding (coded information) is transmitted to decoding apparatus 303 via transmission channel 302. Decoding apparatus 303 receives the coded information transmitted from coding apparatus 301 via transmission channel 302, and, by decoding this, acquires an output signal. FIG.6 is a block diagram showing an inner principal-part configuration of coding apparatus 301. If input signal sampling frequency is SRinput, down-sampling processing section 311 down-samples the input signal sampling frequency from SRinput to SRbase (SRbase