8 2. LITERATURE SURVEY The available radio spectrum for the wireless radio communication is very limited hence to accommodate maximum number of users the speech is compressed. The speech compression techniques are means of reducing the bandwidth needed to represent the human voice. In mobile communication the acceptable voice bandwidth is 3 khz and the frequencies above 3 khz contain the energy of the voice signal, which are needed for the application other than high fidelity. By eliminating these high frequencies we can save the bandwidth. The basic requirement of a speech compression technique is to retain the quality of the speech. The bandwidth required to transmit a speech signal is directly proportional to the speech quality. High quality speech requires maximum bandwidth or bit rate. The quality of the speech signal is measured by comparing with the toll quality which is provided by the modern telephone carrier system under favorable condition [4]. The approximate S/N ratio of a toll quality speech required is 30db. There are different speech compression techniques and the simplest technique is the pulse code modulation technique [PCM]. PCM: This is the standard technique used by the telephone companies [5]. Here the speech is sampled at the sampling rate of 8 khz and each sample is represented by 8 bits. Thus uses 64 khz of bandwidth which is not preferred for wireless communication because of wider bandwidth requirement. The resulting digital signal due to PCM with wider bandwidth will not fit on ordinary telephone channel. Hence data compression is required to transmit the data through the telephone line. The S/N ratio provided by a single PCM encoded data is about 38db. If the quantizer step size varies itself to the strength of the input speech signal then such type of quantizers are called adaptive PCM. The conversion of high rate data to lower rate data virtually always entails a loss of fidelity or increase in distortion. Hence the coding technique chosen must
9 give good fidelity. The adaptive vector quantizers are preferred because of the low quantization noise. The quantizers can be of memory type or memoryless type. In case of memoryless quantizer the present quantized value does not depend on the previously quantized value and this methodology is used in PCM [5] and if the present quantized value depends on the past value then it is memory type and is used in predictive coding. According to the Shannon s rate distortion theory devoted to data compression states that A better performance can be achieved by vector quantizers than scalar quantizers. The quantization error can be minimized by using vector quantizers [6]. The vector quantizers with and without memory are preferable compared to other quantizers for both speech and image quantization. Differential PCM: When a message signal is sampled at a rate of 8 khz, then the present sample value can be predicted by the previous sample value. In this method only the difference between adjacent sample values are encoded and transmitted. Usually DPCM [5] are adaptive type where in the step size adjust with the input signal and are adaptive PCM. The APCM digitize the continuous signal at the rate of 32kb/s to 40kb/s. Delta modulation: This is the simplest coding technique. Here the difference between adjacent values are encoded into a one bit word. In this technique the sampling rate is equal to the bit rate. If the quantizer step size varies with the I/P voice signal, then it is designated as ADM and operate at 16 to 32 kb/s. The quantization noise is audible at 16 kb/s. Digital Vocoders: In this technique the I/P speech signal is analysed and parameters necessary for the synthesis are extracted. Only the extracted parameters are transmitted from the transmitter to the receiver. The advantage of Vocoders are due to its low bit rate and the disadvantage is due to the synthetic type of speech generated. This synthetic type of speech is not prefered in wireless communication[7].
10 Linear Predictive Coders [LPC]: These are the complex speech coders which generate the quality speech signal compared to Vocoder and are called hybrid coders. The nature of the speech signal is retained by the hybrid coder at the output. The different coding techniques like CELP, MP-LPC (Multi pulse linear predictive coding), APC (adaptive predictive coding) and ATC (Adaptive transform coding) requires coding delay of 40 to 60 ms and due to the larger coding delay echo cancellation becomes impossible[8]. The Vocoders like VSELP, RPE-LTP have lesser bit rate but are not toll quality speech coders. One of the widely used hybrid coder is CELP coder. The CELP coder uses pitch predictor along with LPC predictor. In case of the CELP coders the processing delay is more because the processing is done for every frame of size 32 vectors and each vector is of 5 samples, which is over come using the LD-CELP [9]. The LD-CELP coder does not use the pitch predictor as conventional CELP because of its sensitivity to the channel noise. Hence the pitch predictor of conventional type is replaced by higher order LPC predictor. The excitation gain is updated by a 10 th order adaptive linear predictor based on the logarithmic gain of previously quantized and scaled excitation vector. The autocorrelation functions are calculated by hybrid windowing method. The code index is gray coded because of its robustness to the channel error. The LD- CELP uses backward gain adaptive vector quantization for excitation and the LPC predictors are updated once every four vectors. The predictor gain coefficients are calculated using the recursive windowing method because of its low computational complexity and improved S/N ratio. The recursive windowing was not suitable for fixed point processor hence hybrid windowing method was chosen [9]. The CELP coder has been modified to get very good toll quality at 16 kb/sec. This coder is similar to that of LD-CELP and has the following features like fast LPC quantization, 3-tap
11 pitch prediction with efficient open loop pitch search and predictor tap quantization, backward adaptive excitation gain and a trained excitation codebook with small size [11]. The main drawback of modified CELP coder is its computational delay due to its open loop prediction. The improved form of CELP known as VCELP [Vector quantization code excited linear prediction]and is proposed for GSM system [13]. The VCELP has low computation complexity because it uses two different code books. These two codebooks are orthogonal codebooks but the two codebook search are interdependent hence increasing the search time. This drawback is overcome in FVQ-CELP [Fast vector quantization CELP]. In real time implementation, the LD-CELP algorithm [12] requires coding delay less than 20ms and using the Matlab stimulation requires 20 million operations per second for full duplex operation. The codebook used in hybrid speech coders can be ordered type or disorder type [14]. In case of ordered type the computation time required is less but number of bits required to code the non-redundant residual speech data is more. In case of ordered code book the residual speech signal obtained is directly compared with templates in the codebook, hence number of templates required in the codebook must be more. This inturn increases the entries in the codebook, increasing the bits required to code. In case of disordered codebook the residual speech is compared with the set of speech templates in additive and subtractive manner until we get the desired template present in the codebook. The codebook has limited number of entries and hence the number of bits required to code are less. The advantage of disordered codebook is because of its limited number of bits required to implement the code. But the disadvantage is due to its computation time. The gain codebook can be optimized by predicting the gain value adaptively and quantizing the same [11]. The long term predictor is omitted and a synthesis filter of order
12 50 is chosen in LD-CELP. But as the order of the filter increases the computation complexity increases non-linearly which can be overcome by optimizing the gain value. In digital communication the noise in the channel along with the channel bit errors can degrade the performance of speech coders hence this drawback can be overcome by using the channel coders. The practical channel used to transmit the message packet can be of Gaussian type or Rayleigh fading type. This fading effect can be nullified by using the convolution coding technique [15] for full-rate. The speech compressed data has to be protected before transmitting the same into the channel. This can be implemented by using joint source and channel coding technique [16]. In this thesis the channel is assumed to be memoryless type and error correcting code is selected by using rate adaptive type. If the information about channel is known then the joint coding technique is preferred because it has minimum coding delay. The joint coding technique is not preferred for the channels with memory because of the maximum feedback time. This drawback can be overcome in the adaptive error control [17] scheme where the noise level of the channel is checked. If the channel is erroneous then Convolutional code is used else Diffused code is used. This is practically not feasible because of the cost. The channel coding technique can be used not only for error detection and correction but also to overcome the cochannel interference. Generally the co-channel interference can be overcome by power controlling technique or by antenna array or by sectoring. These techniques are not feasible for the mobile unit, hence we can use channel coding technique to overcome the co-channel interference [18]. Linear combining codes can be used alone or with small array of antennas, which is preferred for the smaller size of the mobile unit. By this technique the capacity of the cellular system can be improved as the small array of antennas works as smart antenna with high directivity. The high directivity antennas inturn improves the S/I ratio and also S/N ratio. In the
13 above coding techniques the throughput is maintained constant irrespective of the S/N of the channel. This can be overcome by changing the coding gain adaptively according to the radio link condition [19]. This was mainly designed for the CDMA system. The above code mentioned is designed for full-rate codec. The present speech coding standard used in the CDMA2000 network is SMV [Selectable mode Vocoder]. The SMV provides multiple modes of operation that are selected based on input speech characteristics. The SMV for wideband CDMA is based on four codec s which are operated in full-rate or half-rate. The full-rate and half-rate are based on the CELP algorithm. The disadvantage of this methodology is due to its high processing delay compared to the LD- CELP which is less than 20ms. The CELP algorithm has got higher delay because the coder works on a frame of 160 speech samples (20 ms) and requires a look-ahead of 80 samples (10 ms) if noise-suppression option is used. So the algorithmic delay for the coder is 30 ms. In CDMA the forward link information is spread using Walsh codes. The Walsh codes are binary digits and different Walsh codes are reserved for different functions like synch, paging Etc. and rest for traffic. The reverse link carries information like access signaling and traffic information. The channel coders used are convolutional encoders. It uses half-rate convolutional encoder and outputs two bits of encoded data for every input bit whenever the information rate is lower than 9600 bps. Each code symbol at 4800 bps rate shall be repeated 1 time. Each code symbol at 2400 bps rate shall be repeated 3 times. Each code symbol at 1200 bps data rate shall be repeated 7 times. For all the data rates this result in a constant modulation symbol rate of 19200 modulation symbols per second. The channel coding technique used is the convolutional coding irrespective to the S/N in the channel. Hence this drawback is overcome in the proposed method.
14 The speech processing technique used in the GSM system is CELP. The most common speech coding scheme is code exited linear prediction which is mainly used in the GSM standard. In this coding the modeling is divided into two stages i.e. linear predictive stage that models the spectral envelop and code based model to encode the residual data. The pitch predictor used in the conventional CELP is very sensitive to the channel error and this disadvantage can be overcome by the proposed method using a low delay CELP coder [12]. The pitch predictor in the conventional CELP is replaced by a LPC predictor of high order in LD- CELP. The coding delay is reduced in LD-CELP as it uses only one vector for processing. The speech coded data is encoded using the channel coding technique to avoid the loss of information due to transmission errors. The channel coding techniques used by the GSM system are block coding technique and convolutional coding technique irrespective of the S/N ratio. Hence the number of redundancy bits added are more irrespective to the noise level in the channel. The proposed method overcomes this drawback by checking the noise level in the channel and then decides the channel coding technique to be used. If the channel is erroneous then we use convolutional coding technique else we use cyclic coding technique. The adaptive coding technique designed for the CELP coder is adaptive multirate code [20]. This codec is designed to operate on both full-rate traffic channels and half-rate traffic channels. The present source coding technique used in the GSM system is the CELP coder. The main drawback of the CELP coder is its coding delay which is overcome by the proposed method used in this thesis. 2.1 SUMMARY In this chapter the literature survey for different source and channel coding techniques is done. A brief description of the source coding techniques and the channel coding techniques
15 used by the GSM system and CDMA system is briefed. The drawback of the existing source coding techniques used in GSM system and CDMA system is highlighted in this chapter.