Listening Samples

This page presents speech samples for comparative listening between our DNN model and the commercial speech/audio coders, including AMR-WB and OPUS, operating at bitrates from 9kbps to 24kbps. 

Our model's loss function for the network training comprises perceptual loss terms based on both frame-wise and subframe-wise psychoacoustic models (PAM-1). We also approximate the quantization process using white noise with uniform distribution.

References

[1] Joon Byun, Seungmin Shin,  Youngcheol Park,  Jongmo Sung,  and Seungkwon  Beack ,  “Development  of  a psychoacoustic  loss  function  for  the  deep  neural  network (DNN)-based speech coder,” Interspeech 2021.

[2] Seungmin Shin, Joon Byun,  Youngcheol Park,  Jongmo Sung,  and Seungkwon  Beack ,  “Deep Neural Network (DNN) Audio Coder Using A Perceptually Improved Training Method,” IEEE ICASSP 2022.

<Sample-1> Encoded at 9kbps, fs=16kHz

<Sample-2> Encoded at 13kbps, fs=16kHz

<Sample-3> Encoded at 20kbps, fs=16kHz

<Sample-4> Encoded at 24kbps, fs=16kHz