Listening Samples

Here, we post audio samples for comparative listening between our DNN model and MP3. The following samples were obtained using a ResGLU-type autoencoder having 1.8M parameters. The loss function for the network training comprises a perceptual loss term based on a psychoacoustic model (PAM-1). We also approximate the quantization process using white noise with uniform distribution.

References

[1] Byun  Joon,  Shin  Seungmin,  Park  Youngcheol,  Sung Jongmo,  and  Beack  Seungkwon,  “Development  of  a psychoacoustic  loss  function  for  the  deep  neural  network (DNN)-based speech coder,” in Proceedings of the Annual Conference of the International Speech Communication Association (Interspeech), 2021, pp. 1694–1698.

[2]  Valero Laparra Johannes Balle and Eero P. Simoncelli, “End-to-end optimization of nonlinear transform codes for perceptual quality,”  in Picture Coding Symposium, 2016.

<Sample-1> Encoded at 48kbps, fs=32kHz

<Sample-2> Encoded at 56kbps, fs=32kHz

<Sample-3> Encoded at 64kbps, fs=32kHz