A method for dynamically allocating bits for the adaptive and fixed codebook of an analysis-by-synthesis speech codec is proposed. The bit allocation is based on the closed loop weighted mean squared error. The different bit allocations identify various codebook configurations used by the adaptive and fixed codebooks of the codec. Unlike open loop approaches where the decision on a codebook configuration is made once per frame, in the closed loop approach the decision on which codebook configuration should be used is made in each subframe. A variable length code is used for coding the codebook configuration. The factorial packing codebook is used as the fixed codebook. The technique is a part of 8.5 kbps mode of EVRC-WB speech coding standard.

Factorial pulse coding, a method which is known to efficiently code an information signal using unit magnitude pulses, involves computation of combinatorial functions. These computations are highly complex as they require many multiply and divide operations on multi-precision numbers, especially when the length of a signal is large or many unit magnitude pulses are used for coding. In this paper, we propose a very low complexity method for approximation of these combinatorial functions. The approximate functions satisfy a property which preserves unique decode-ability of the factorial packing encoding/decoding algorithm. The low complexity computation enables use of factorial packing in encoding/decoding of 144 MDCT coefficients using 28 unit magnitude pulses for the audio coding mode of the EVRC-WB speech coding standard without affecting the number of bits required for coding.

The method of a 1/sup st/ order long-term predictor (LTP) filter, using a sub-sample resolution delay, is extended to a multi-tap LTP filter, or, equivalently, the conventional integer-sample resolution multi-tap LTP filter is extended to use sub-sample resolution delay. Defining the delay with sub-sample resolution enables this novel multi-tap LTP filter to explicitly model delay values that have a fractional component. The filter coefficients, largely freed from implicitly modeling the effect of delays that have a fractional component, seek to maximize the prediction gain of the LTP filter by modeling the frequency dependent gain. This is in contrast to a conventional multitap LTP filter, which applies a single model to tackle the dual tasks of representing the non-integer valued delays and the frequency dependent gain. Experimental results are presented for narrowband and wideband speech. This technique is part of the 3GPP2 source-controlled variable-rate multimode wideband speech codec (VMR-WB) rate set 1 standard.

Codebook searches in analysis-by-synthesis speech coders typically involve minimization of a perceptually weighted squared error signal. Minimization of the error over multiple codebooks is often done in a sequential manner, resulting in the choice of overall excitation parameters being sub-optimal. In this paper, we propose a joint excitation parameter optimization framework in which the associated complexity is slightly greater than the traditional sequential optimization, but with significant quality improvement. Moreover, the framework allows joint optimization to be easily incorporated into existing pulse codebook systems with little or no impact on the codebook search algorithms. This technique is part of the 3GPP2 source-controlled variable-rate multimode wideband speech codec (VMR-WB) rate set 1 standard.

A joint source-channel coding system for transmitting speech on a bandlimited additive white Gaussian noise (AWGN) channel is presented. The proposed method uses a hybrid of digital and analog modulation techniques. The digital part of the system consists of a Federal Standard 1016 code-excited linear predictive (FS 1016 CELP) speech coder followed by a rate-3/5 parallel concatenated (turbo) error correcting code. The analog part, which transmits the quantization error due to the FS 1016 CELP coder, consists of a linear encoder and decoder. The advantage of the proposed system is that it achieves excellent rate-distortion/capacity performance that is common in digital systems while maintaining a "graceful degradation" characteristic that is common in analog systems. Comparisons are made with three purely digital systems and an analog system-with all systems operating at the same overall rate. A formal listening test shows that, at high channel SNR, the improved performance of the proposed hybrid system (versus the purely digital systems) is noticeable to the average listener. Finally, an informal listening test indicates that at low channel SNR (where the error correcting code breaks down) the decoded speech of the hybrid system is still intelligible.

We consider the problem of transmitting a band-limited Gaussian source on an additive band-limited Gaussian noise channel. The well-known "threshold effect" dictates that the more powerful a code is, the more sensitive it is to the exact knowledge of the channel noise. A code is said to be robust if it is asymptotically optimal for a wide range of channel noise. Thus, robust codes have a "graceful degradation" characteristic and are free of the threshold effect. It is demonstrated that robust codes exist whenever the source and channel bandwidths are equal. In the unequal-bandwidth case, a collection of nearly robust joint source-channel codes is constructed using a hybrid digital-analog (HDA) coding technique. For designing nearly robust codes, a matched tandem code whose channel encoder's output is partially/fully matched to its input is proposed and the existence of an asymptotically optimal matched tandem code is shown. The nearly robust codes achieve the Shannon limit (theoretically optimum distortion) and have a less severe threshold effect. Finally, for the case of two different noise conditions, the distortion regions of these codes are determined.

A signal/noise KLT based approach for enhancing speech degraded by colored noise is proposed. The noisy speech frames are classified into speech-dominated frames and noise-dominated frames. In the speech-dominated frames, the signal KLT matrix is used and in the noise-dominated frames, the noise KLT matrix is used. The approach does not require noise whitening and hence works well even with narrowband noise. A two-dimensional objective measure which captures both the speech distortion and the noise shaping characteristics of the algorithm is proposed. This measure indicates that the proposed method performs better noise shaping than a modified version of the signal subspace approach proposed by Ephraim and Van Trees and the standard spectral subtraction method. Informal listening tests show that the proposed algorithm does not suffer from the problem of residual musical noise and performs better noise masking than the signal subspace approach.

A joint source-channel coding system for transmitting speech on a bandlimited additive white Gaussian noise (AWGN) channel is presented. The proposed method uses a hybrid of digital and analog modulation techniques. The digital part of the system consists of an FS-1016 CELP speech coder followed by a rate-3/5 parallel concatenated (turbo) error correcting code. The analog part, which transmits the quantization error due to the FS-1016 CELP coder consists of a linear encoder and decoder. The advantage of the proposed system is that it achieves excellent rate-distortion/capacity performance that is common in digital systems while maintaining a "graceful degradation" characteristic that is common in analog systems. Comparisons are made with three purely digital systems and two purely analog systems. A formal listening test shows that, at high channel SNR, the improved performance of the proposed hybrid system versus the purely digital systems is noticeable to the average listener.

A signal/noise Karhunen-Loeve transform (KLT) based approach for enhancing speech degraded by colored noise is proposed. The noisy speech frames are classified into speech-dominated frames and noise-dominated frames. In the speech-dominated frames, the signal KLT matrix is used and in the noise dominated frames, the noise KLT matrix is used. The approach does not require noise whitening and hence works well even with narrowband noise. A two-dimensional objective measure which captures both the speech distortion and the noise shaping characteristics of the algorithm is proposed. This measure indicates that the proposed method performs better noise shaping than a modified form of the signal subspace approach proposed by Ephraim and Van Trees (1995) and the standard spectral subtraction method. Informal listening tests show that the proposed algorithm does not suffer from the problem of residual musical noise and performs better noise masking than the signal subspace approach.

We consider joint source-channel coding for a memoryless Gaussian source and an additive white Gaussian noise (AWGN) channel. For a given code defined by an encoder-decoder pair (/spl alpha/, /spl beta/), its dual code is obtained by interchanging the encoder and decoder: (/spl beta/, /spl alpha/). It is shown that if a code (/spl alpha/, /spl beta/) is optimal at rate p channel uses per source sample and if it satisfies a certain uniform continuity condition, then its dual code (/spl beta/, /spl alpha/) is optimal for rate 1//spl rho/ channel uses per source sample. Further, it is demonstrated that there is a code which is optimal but its dual code is not optimal. Finally, using random coding, we show that there is an optimal code which has an optimal dual. The duality concept is also presented for the cases of (i) binary memoryless equiprobable source and binary-symmetric channel (BSC), and (ii) colored Gaussian source and additive colored Gaussian noise (ACGN) channel.