Scholarlink

<p>To solve the problem that the speech separation effect needs to be improved under the condition of low signal-to-noise ratio, this paper proposes a Fourier Transformer Network (FNet) speech separation model without parametric Fourier transform based on TDANet architecture. Since the Fourier transform can better characterize the characteristics and structure of the signal, the self-attention layer in the Transformer encoder is replaced by the parameter-free Fourier transform. The encoder uses partial convolution (PConv) to extract features with different temporal resolutions, and the decoder uses the local attention (LA) layer to generate a set of learnable parameters to adaptively modulate the acoustic features of FNet output to reconstruct the target speech waveform. Based on the TIMIT, LibriSpeech, VoxCeleb1 speech dataset, and NoiseX-92 noise dataset, using SI-SNRi and SDR evaluation metrics, evaluate the speech separation performance of the proposed model and baseline model under different signal-to-noise ratio conditions, and analyze the training speed of the proposed model. Research has shown that the model proposed in this article has good separation performance under low signal-to-noise ratio conditions. When the noise is machinegun and the signal-to-noise ratio is − 10&#xa0;dB, the scale-invariant signal-to-noise ratio improvement (SI-SNRi) and signal-to-distortion ratio (SDR) of FNet are 12.33 and 12.72, respectively, which are 0.39 and 0.51 higher than the baseline model Sepformer, and the training speed of the model is only 15% of Sepformer. It can be seen that the model proposed in this article has good speech separation performance, especially for situations with low signal-to-noise ratio, and the model has good generalization ability.</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

Research on End-to-End Speech Separation Based on FNet

  • 期刊论文

  • 发表日期：

    2026年7月14日

查看全文
  • Rongqi Liu,
  • Ping Xue,
  • Chaofeng Lan,
  • Yingqi Chen,
  • Yuqiao Wang

摘要

To solve the problem that the speech separation effect needs to be improved under the condition of low signal-to-noise ratio, this paper proposes a Fourier Transformer Network (FNet) speech separation model without parametric Fourier transform based on TDANet architecture. Since the Fourier transform can better characterize the characteristics and structure of the signal, the self-attention layer in the Transformer encoder is replaced by the parameter-free Fourier transform. The encoder uses partial convolution (PConv) to extract features with different temporal resolutions, and the decoder uses the local attention (LA) layer to generate a set of learnable parameters to adaptively modulate the acoustic features of FNet output to reconstruct the target speech waveform. Based on the TIMIT, LibriSpeech, VoxCeleb1 speech dataset, and NoiseX-92 noise dataset, using SI-SNRi and SDR evaluation metrics, evaluate the speech separation performance of the proposed model and baseline model under different signal-to-noise ratio conditions, and analyze the training speed of the proposed model. Research has shown that the model proposed in this article has good separation performance under low signal-to-noise ratio conditions. When the noise is machinegun and the signal-to-noise ratio is − 10 dB, the scale-invariant signal-to-noise ratio improvement (SI-SNRi) and signal-to-distortion ratio (SDR) of FNet are 12.33 and 12.72, respectively, which are 0.39 and 0.51 higher than the baseline model Sepformer, and the training speed of the model is only 15% of Sepformer. It can be seen that the model proposed in this article has good speech separation performance, especially for situations with low signal-to-noise ratio, and the model has good generalization ability.