LACE: Layer-Wise Compression for Dynamic Frame Rate Codecs
LACE lets each quantization layer choose its own audio frame boundaries instead of forcing one shared compression map.
The paper argues that residual embeddings can change at different speeds across codec layers, making shared segmentation a poor fit. Its layer-wise compression is paired with union alignment and boundary anchors so TTS durations remain consistent across layers. On LibriTTS, the authors report a better reconstruction rate-quality tradeoff than earlier dynamic frame rate methods, plus faster TTS inference with competitive synthesis quality. Code is released through the ESPnet3 codec recipe. ArXiv · AI/CL/LG's note
The paper argues that residual embeddings can change at different speeds across codec layers, making shared segmentation a poor fit. Its layer-wise compression is paired with union alignment and boundary anchors so TTS durations remain consistent across layers. On LibriTTS, the authors report a better reconstruction rate-quality tradeoff than earlier dynamic frame rate methods, plus faster TTS inference with competitive synthesis quality. Code is released through the ESPnet3 codec recipe. ArXiv · AI/CL/LG's note
score 4