Toward Complex-Valued Neural Networks for Waveform Generation

Dimensions overall score 9.0

• ComVo, a complex-valued neural vocoder, is introduced, utilizing native complex arithmetic in both generator and discriminator.method 95%
  Evidencepartial

  We present ComVo, a Complex-valued neural Vocoder whose generator and discriminator use native complex arithmetic.

  Implicationmissing

  Implication not extracted yet.

  Verificationpartial

  partial

• Phase quantization is introduced to discretize phase values and regularize the training process.method 90%
  Evidencepartial

  To guide phase transformations in a structured manner, we introduce phase quantization, which discretizes phase values and regularizes the training process.

  Implicationmissing

  Implication not extracted yet.

  Verificationpartial

  partial

• A block-matrix computation scheme is proposed to improve training efficiency.method 90%
  Evidencepartial

  Finally, we propose a block-matrix computation scheme to improve training efficiency by reducing redundant operations.

  Implicationmissing

  Implication not extracted yet.

  Verificationpartial

  partial

• ComVo achieves higher synthesis quality than comparable real-valued baselines.result 85%
  Evidencepartial

  Experiments demonstrate that ComVo achieves higher synthesis quality than comparable real-valued baselines

  Implicationmissing

  Implication not extracted yet.

  Verificationpartial

  partial

• The block-matrix scheme reduces training time by 25%.result 80%
  Evidencepartial

  its block-matrix scheme reduces training time by 25%

  Implicationmissing

  Implication not extracted yet.

  Verificationpartial

  partial

• Current real-valued networks limit the ability to capture the inherent structure of complex spectrograms.limitation 80%
  Evidencepartial

  This separation limits their ability to capture the inherent structure of complex spectrograms.

  Implicationmissing

  Implication not extracted yet.

  Verificationpartial

  partial

• iSTFT-based vocoders avoid learned upsampling stages, reducing computational cost.technical 75%
  Evidencepartial

  They predict a complex-valued spectrogram and then synthesize the waveform via iSTFT, thereby avoiding learned upsampling stages that can increase computational cost.

  Implicationmissing

  Implication not extracted yet.

  Verificationpartial

  partial

• Current iSTFT-based vocoders use real-valued networks that process real and imaginary parts independently, limiting their ability to capture inherent structure of complex spectrograms.limitation 90%
  Evidencepartial

  However, current approaches use real-valued networks that process the real and imaginary parts independently. This separation limits their ability to capture the inherent structure of complex spectrograms.

  Implicationpartial

  Directly stated in abstract as a limitation of existing approaches

  Verificationpartial

  partial

• ComVo achieves higher synthesis quality than comparable real-valued baselines.result 85%
  Evidencepartial

  Experiments demonstrate that ComVo achieves higher synthesis quality than comparable real-valued baselines

  Implicationpartial

  Directly stated in abstract with experimental support mentioned

  Verificationpartial

  partial

• ComVo's block-matrix computation scheme reduces training time by 25%.result 90%
  Evidencepartial

  its block-matrix scheme reduces training time by 25%

  Implicationpartial

  Direct numeric claim with specific percentage improvement stated in abstract

  Verificationpartial

  partial

• ComVo uses native complex arithmetic in both generator and discriminator, enabling adversarial training with structured feedback in complex-valued representations.method 85%
  Evidencepartial

  We present ComVo, a Complex-valued neural Vocoder whose generator and discriminator use native complex arithmetic. This enables an adversarial training framework that provides structured feedback in complex-valued representations.

  Implicationpartial

  Direct description of method's core innovation in abstract

  Verificationpartial

  partial

• ComVo introduces phase quantization to discretize phase values and regularize training process.method 85%
  Evidencepartial

  we introduce phase quantization, which discretizes phase values and regularizes the training process

  Implicationpartial

  Direct description of specific technical innovation in abstract

  Verificationpartial

  partial

Competitive landscape

ComVo is a complex-valued neural vocoder that enhances waveform generation with structured feedback and improved training efficiency.