Exploiting time-frequency patterns with LSTM-RNNs for low-bitrate audio restoration

Perceptual audio coding is heavily and successfully applied for audio compression. However, perceptual audio coders may inject audible coding artifacts when encoding audio at low bitrates. Low-bitrate audio restoration is a challenging problem, which tries to recover a high-quality audio sample close to the uncompressed original from a low-quality encoded version. In this paper, we propose a novel data-driven method for audio restoration, where temporal and spectral dynamics are explicitly captured by a deep time-frequency-LSTM recurrent neural networks. Leveraging the captured temporal and spectral information can facilitate the task of learning a nonlinear mapping from the magnitude spectrogram of low-quality audio to that of high-quality audio. The proposed method substantially attenuates audible artifacts caused by codecs and is conceptually straightforward. Extensive experiments were carried out and the experimental results show that for low-bitrate audio at 96 kbps (mono), 64 kbps (mono), and 96 kbps (stereo), the proposed method can efficiently generate improved-quality audio that is competitive or even superior in perceptual quality to the audio produced by other state-of-the-art deep neural network methods and the LAME-MP3 codec.