Pixel-based pixel value ordering (PPVO) represents an efficient solution for low bit-rate and high-fidelity reversible data hiding in digital images. The main novelty of this approach is the selection between two distinct prediction values for each potential host pixel. This paper maintains and refines this feature and adapts it for audio reversible data hiding. For each audio sample value, three possible prediction values are considered: a simple average of neighboring samples around the current position and two weighted averages. Differences between neighboring samples are used to select between the two weighted predictors from the three possibilities. The PPVO framework is then used to select the final predictor for each individual sample from the two remaining predictors. The data embedding algorithm is also adapted to allow for multi-bit embedding, which better exploits the prediction error histogram of audio sample values. Experimental are provided, the proposed PPVO-based approach outperforms other existing audio reversible data hiding scheme, most notably at medium bit-rates.