Time-domain flexible matrix quantization for very-low-rate speech coding