Skip to content

Fast and differentiable time domain all-pole filter in PyTorch.

License

Notifications You must be signed in to change notification settings

DiffAPF/torchlpc

Folders and files

NameName
Last commit message
Last commit date
Jan 24, 2025
May 4, 2025
May 4, 2025
Jul 8, 2023
Apr 13, 2024
Jul 9, 2023
Oct 19, 2024
Sep 10, 2024
May 4, 2025

Repository files navigation

TorchLPC

PyPI version

torchlpc provides a PyTorch implementation of the Linear Predictive Coding (LPC) filter, also known as all-pole filter. It's fast, differentiable, and supports batched inputs with time-varying filter coefficients.

Given an input signal x R T and time-varying LPC coefficients A R T × N with an order of N , the LPC filter is defined as:

y t = x t i = 1 N A t , i y t i .

Usage

import torch
from torchlpc import sample_wise_lpc

# Create a batch of 10 signals, each with 100 time steps
x = torch.randn(10, 100)

# Create a batch of 10 sets of LPC coefficients, each with 100 time steps and an order of 3
A = torch.randn(10, 100, 3)

# Apply LPC filtering
y = sample_wise_lpc(x, A)

# Optionally, you can provide initial values for the output signal (default is 0)
zi = torch.randn(10, 3)
y = sample_wise_lpc(x, A, zi=zi)

Installation

pip install torchlpc

or from source

pip install git+https://github.com/DiffAPF/torchlpc.git

Derivation of the gradients of the LPC filter

The details of the derivation can be found in our preprints12. We show that, given the instataneous gradient L y t where L is the loss function, the gradients of the LPC filter with respect to the input signal x and the filter coefficients A can be expresssed also through a time-varying filter:

L x t = L y t i = 1 N A t + i , i L x t + i

L A = | L x 1 0 0 0 L x 2 0 0 0 L x t | | y 0 y 1 y N + 1 y 1 y 0 y N + 2 y T 1 y T 2 y T N | .

Gradients for the initial condition y t | t 0

The initial conditions provide an entry point at t = 1 for filtering, as we cannot evaluate t = . Let us assume A t , : | t 0 = 0 so y t | t 0 = x t | t 0 , which also means L y t | t 0 = L x t | t 0 . Thus, the initial condition gradients are

L y t = L x t = i = 1 t N A t + i , i L x t + i for  N < t 0.

In practice, we pad N and N × N zeros to the beginning of L y and A before evaluating L x . The first N outputs are the gradients to y t | t 0 and the rest are to x t | t > 0 .

Time-invariant filtering

In the time-invariant setting, A t , i = A 1 , i t [ 1 , T ] and the filter is simplified to

y t = x t i = 1 N a i y t i , a = A 1 , : .

The gradients L x are filtering L y with a backwards in time, same as in the time-varying case. L a is simply doing a vector-matrix multiplication:

L a T = L x T | y 0 y 1 y N + 1 y 1 y 0 y N + 2 y T 1 y T 2 y T N | .

This algorithm is more efficient than 3 because it only needs one pass of filtering to get the two gradients while the latter needs two.

TODO

  • Use PyTorch C++ extension for faster computation.
  • Use native CUDA kernels for GPU computation.
  • Add examples.

Related Projects

  • torchcomp: differentiable compressors that use torchlpc for differentiable backpropagation.
  • jaxpole: equivalent implementation in JAX by @rodrigodzf.

Citation

If you find this repository useful in your research, please cite our work with the following BibTex entries:

@inproceedings{ycy2024diffapf,
    title={Differentiable All-pole Filters for Time-varying Audio Systems},
    author={Chin-Yun Yu and Christopher Mitcheltree and Alistair Carson and Stefan Bilbao and Joshua D. Reiss and György Fazekas},
    booktitle={International Conference on Digital Audio Effects (DAFx)},
    year={2024},
    pages={345--352},
}

@inproceedings{ycy2024golf,
    title     = {Differentiable Time-Varying Linear Prediction in the Context of End-to-End Analysis-by-Synthesis},
    author    = {Chin-Yun Yu and György Fazekas},
    year      = {2024},
    booktitle = {Proc. Interspeech},
    pages     = {1820--1824},
    doi       = {10.21437/Interspeech.2024-1187},
}

Footnotes

  1. Differentiable All-pole Filters for Time-varying Audio Systems.

  2. Differentiable Time-Varying Linear Prediction in the Context of End-to-End Analysis-by-Synthesis.

  3. Singing Voice Synthesis Using Differentiable LPC and Glottal-Flow-Inspired Wavetables.

About

Fast and differentiable time domain all-pole filter in PyTorch.

Topics

Resources

License

Citation

Stars

Watchers

Forks

Packages

No packages published