Scholarlink

<p>Audio super-resolution (ASR), also known as bandwidth extension (BWE), restores low-resolution or frequency-distorted speech signals by recovering high-frequency components. However, existing approaches frequently encounter difficulties capturing the high-frequency harmonic relations accurately while balancing inference speed with computational efficiency. This paper introduces the Depthwise Convolutional Gated Recurrent Network (DC-GRN) model, which attempts to address these challenges by focusing on varying sampling rates and improving high-frequency reconstruction. The DC-GRN model incorporates depthwise convolutions and gated recurrent units (GRUs) to simultaneously capture the temporal and spatial dependencies in audio, enabling high-quality reconstruction with reduced computational cost. The model is trained using a direct mapping learned with the MAE loss function to reconstruct degraded inputs from 2k, 4k, 8k, and 16k to target frequencies of 24k and 48k. Experimental results show that the proposed DC-GRN ASR model, tested on the VCTK dataset, upsamples 2k degraded signals to 24&#xa0;kHz, achieving a Log-Spectral Distance (LSD) of 1.21 and a ViSQOL score of 3.46, outperforming existing methods. For 48&#xa0;kHz upsampling from 8k signals, the model achieves an LSD of 0.61 and a ViSQOL score of 3.17, surpassing baselines like AP-BWE and VM-ASR. Additionally, DC-GRN achieves a real-time factor of 0.0034 (198.74×), with 3.43&#xa0;M parameters and 2.71G FLOPS, making it a strong candidate for practical audio super-resolution (ASR) applications.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Audio super-resolution based on depthwise convolutional gated recurrent network

  • 期刊论文

  • 发表日期：

    2025年12月10日

查看全文
  • Ikram Azaz,
  • Tao Zhang,
  • Yasir Iqbal,
  • Xin Zhao,
  • Yanzhang Geng

摘要

Audio super-resolution (ASR), also known as bandwidth extension (BWE), restores low-resolution or frequency-distorted speech signals by recovering high-frequency components. However, existing approaches frequently encounter difficulties capturing the high-frequency harmonic relations accurately while balancing inference speed with computational efficiency. This paper introduces the Depthwise Convolutional Gated Recurrent Network (DC-GRN) model, which attempts to address these challenges by focusing on varying sampling rates and improving high-frequency reconstruction. The DC-GRN model incorporates depthwise convolutions and gated recurrent units (GRUs) to simultaneously capture the temporal and spatial dependencies in audio, enabling high-quality reconstruction with reduced computational cost. The model is trained using a direct mapping learned with the MAE loss function to reconstruct degraded inputs from 2k, 4k, 8k, and 16k to target frequencies of 24k and 48k. Experimental results show that the proposed DC-GRN ASR model, tested on the VCTK dataset, upsamples 2k degraded signals to 24 kHz, achieving a Log-Spectral Distance (LSD) of 1.21 and a ViSQOL score of 3.46, outperforming existing methods. For 48 kHz upsampling from 8k signals, the model achieves an LSD of 0.61 and a ViSQOL score of 3.17, surpassing baselines like AP-BWE and VM-ASR. Additionally, DC-GRN achieves a real-time factor of 0.0034 (198.74×), with 3.43 M parameters and 2.71G FLOPS, making it a strong candidate for practical audio super-resolution (ASR) applications.