[Hackathon 7th No.55] Add fft_conv1d to PaddleSpeech -part

[DrRyanHuang]

PR types

New features

PR changes

APIs

Describe

fft_conv1d is much faster than conv1d when kernel_size >= 128

Implementation:

• Standard conv1D applies convolutional kernels to the input signal using a sliding window approach.
• fft_conv1d leverages the Fast Fourier Transform (FFT) to accelerate the convolution operation, ideally reducing the computational complexity from $$O(N \cdot K)$$ to $$O(N \log N)$$, where $$N$$ denotes the signal length and $$K$$ represents the kernel length.

Performance:

• Standard conv1D: Typically performs well with small kernels or short input signals.
  Simple to implement and easy to debug.
• fft_conv1d: Generally offers performance improvements when the kernel size is large (e.g., above 256) and the stride is 1. However, the introduction of FFT and IFFT incurs additional memory overhead.

Result:

• origin PR: [Hackathon 7th No.55] Add audiotools to PaddleSpeech #3900
• RFC: 【Hackathon 7th No.55】在 PaddleSpeech 中实现 audiotools community#1017

[paddle-bot]

Thanks for your contribution!

[zxcd]

What does this api do?

[zxcd]

paddle do not have .strides release api. Also for as_strided support >2.6, but can be combined with view and reshape .

[zxcd]

v2.6 and above cannot be used?

[zxcd]

LGTM