Speech quality evaluation of neural audio codecs

This paper presents speech quality results to characterize the state of the art and technological advance of recent neural audio codecs targeting low bitrates. Audio quality was evaluated in one clean speech experiment (in French). Degradation Mean Opinion Score (DMOS) results are reported and discussed for neural audio codecs (LPCNet, Lyra V2, EnCodec, AudioCraft, AudioDec, Descript Audio Codec) – traditional codecs (Opus, EVS) are also included as performance yardsticks. We also discuss observed codec complexity to complement subjective test results.