In this research, a new coder based on analysis-by-synthesis technique is developed for coding wideband audio signals at low bit rates. The proposed coder utilizes both the advantages of low-delay coded-excited linear prediction (LD-CELP) coding and adaptive differential pulse code modulation (ADPCM) coding. Instead of using a vector-by- vector gain adaptation approach as in the conventional CELP-based coders, the proposed coder utilizes a sample-by-sample gain adaptive excitation model. With this model, the excitation gain is adapted in a sample-by-sample fashion so that instantaneous response to rapid time-varying signal can be achieved. The approach is similar to the Jayant-Type adaptive quantizer in ADPCM coder but it is working based on the analysis-by-synthesis principle. A non-linear gain adapter driven by a neural network is also developed. The neural network is optimized by a training procedure based on minimizing the coding error. The proposed coder also exploits the masking property of human auditory system. Although psychoacoustic effect has been widely applied to dynamic bit-allocation of the subband or transform coders, it cannot be used in time-domain coders such as LD-CELP. In this research, perceptual masking is applied to generate a psychoacoustic weighting filter for time-domain coders. The quantization noise can be shaped according to the masking threshold of the input audio signal. The performance of the proposed coder working at different bit rates is evaluated by objective and subjective tests. The results show that the new coder can achieve transparency coding of CD quality audio signals at 1.5 bid/sample and can still maintain good quality at 1.0 bid/sample.