Efficient Speech Representation Learning with Low-Bit Quantization