Abstract: In this paper, based on LPC adaptive forward-backward quantization, a novel variable-rate MELP speech coder is proposed, in which linear prediction is done by using either the current (forward LPC) or previously decoded (backward LPC) speech frames. The backward LPC scheme shall be applied, i.e., the LPC coefficients based on the previously decoded optimal speech frame are used to encode the current frame and only the time sequence code shall be transmitted to the decoder, so, average LPC bit number becomes smaller. Computer simulation shows significant average overall bit rate reduction is achieved without compromising the decoded speech quality.