---
id: julius
version: "0.2.8"
license: MIT License
license_treatment: permissive
maintenance: active
---

# julius — Nice DSP sweets: resampling, FFT Convolutions. All with PyTorch, differentiable and with CUDA support.

License: permissive · Maintenance: active · Downloads: 2.7M/mo

## What it is and what it does

Julius is a PyTorch-based DSP library that makes audio and signal processing operations differentiable and GPU-accelerated. It implements sinc resampling, FFT-based convolutions for large kernels, FIR filter banks (lowpass, highpass, bandpass), and frequency-band decomposition in mel-scale space. All operations run on CUDA and are compatible with TorchScript, making them suitable for training neural networks that incorporate signal processing or for real-time inference pipelines.

The library is designed for cases where you need DSP operations to be part of a differentiable computation graph or where GPU acceleration matters. Its resampling is faster than resampy even on CPU and negligible on GPU for typical sample-rate ratios. FFT convolutions outperform standard convolution for kernels >= 128 samples, especially with many channels or large batch sizes.

Use it for:

- Train end-to-end neural networks that include learnable audio resampling or filtering layers.
- Accelerate batch audio preprocessing on GPU by replacing CPU-based librosa or scipy operations.
- Implement parametric EQ or frequency-band-dependent processing as differentiable modules.
- Build real-time audio inference pipelines using TorchScript-compiled DSP operations.
- Perform large-kernel convolutions on long audio signals faster than torch.nn.Conv1d.

## Worth the install?

AI-flagged interpretation of the facts above — verify before relying.

Julius provides differentiable, GPU-accelerated digital signal processing for audio and 1D signals using PyTorch, including resampling, FFT convolutions, and frequency-domain filtering.

Yes. Julius is actively maintained, has no known vulnerabilities, installs with minimal friction (torch only), and fills a genuine gap for GPU-accelerated, differentiable DSP in PyTorch workflows. Install it if you need signal processing as part of a neural network or GPU pipeline; skip it if you only do offline audio analysis on CPU.

## Install

pip install julius
uv add julius
poetry add julius

## Installing julius

Before you install: Low friction: pure Python wheel with only torch as a runtime dependency. Actively maintained with a recent release (72 days old), last commit 2026-06-03, and 461 repository stars.

License in practice: MIT license (permissive) — you can use, modify, and distribute julius freely in commercial and private projects with minimal restrictions.

Quickstart:

pip install julius

import julius
import torch

signal = torch.randn(6, 4, 1024)
resampled = julius.resample_frac(signal, 100, 70)
low_freqs = julius.lowpass_filter(signal, 0.1)

Requires Python >= 3.9.0 and torch installed.

Verify before relying:

- Whether FFTConv1d performance advantage over torch.nn.Conv1d holds for your specific tensor sizes and hardware.
- TorchScript compatibility status and any known limitations beyond what the description states.

## Package facts

- License: MIT License (permissive)
- Python support: supports_current
- Install friction: low
- Maintenance: active
- Downloads: 2.7M/month (top 5,000 on PyPI)
- Known vulnerabilities: none known

## Tags

pytorch audio dsp, differentiable signal processing, gpu accelerated resampling, fft convolution pytorch, audio filtering cuda, mel-scale frequency bands, torchscript dsp, audio-dsp, gpu-accelerated, differentiable

[View on SkillFed](https://skillfed.io/packages/julius) · [View on PyPI](https://pypi.org/project/julius/)