Oven logo

Oven

soxr0.5.0.post1

Published

High quality, one-dimensional sample-rate conversion library

Python-SoXR

GitHub PyPI conda-forge Packaging status Read the Docs

High quality, one-dimensional sample-rate conversion library for Python.

Keywords: Resampler, Audio resampling, Samplerate conversion, DSP(Digital Signal Processing)

Python-SoXR is a Python wrapper of libsoxr.

Installation

pip install soxr

If installation fails, upgrade pip with python -m pip install --upgrade pip and try again.

in Conda environment

conda install -c conda-forge soxr-python

Note: Conda packge name is soxr-python, not python-soxr.

Basic usage

import soxr

y = soxr.resample(
    x,          # input array ā€“ mono(1D) or multi-channel(2D of [frame, channel])
    48000,      # input samplerate
    16000       # target samplerate
)

If input is not numpy.ndarray, it will be converted to numpy.ndarray(dtype='float32').
dtype should be one of float32, float64, int16, int32.

Output is numpy.ndarray with same dimension and data type of input.

Streaming usage

Use ResampleStream for real-time processing or very long signal.

import soxr

rs = soxr.ResampleStream(
    44100,              # input samplerate
    16000,              # target samplerate
    1,                  # channel(s)
    dtype='float32'     # data type (default = 'float32')
)

eof = False
while not eof:
    # Get chunk
    ...

    y_chunk = rs.resample_chunk(
        x,              # input aray ā€“ mono(1D) or multi-channel(2D of [frame, channel])
        last=eof        # Set True at end of input
    )

Output frame count may not be consistent. This is normal operation.
(ex. [0, 0, 0, 186, 186, 166, 186, 186, 168, ...])

šŸ“ More code examples

Benchmark

Sweep, impulse, speed compairsion with other resamplers for Python.

https://colab.research.google.com/drive/1_xYUs00VWYOAXShB85W1MFWaUjGHfO4K?usp=sharing

Speed comparison summary

Downsampling 10 sec of 48000 Hz to 44100 Hz.
Ran on Google Colab.

LibraryTime on CPU (ms)
soxr (HQ)10.8
torchaudio13.8
soxr (VHQ)14.5
scipy.signal.resample21.3
lilfilter24.7
julius31
resampy (kaiser_fast)108
samplerate (sinc_medium)223
resampy (kaiser_best)310
samplerate (sinc_best)794

Technical detail

For technical details behind resampler, see libsoxr docs.

Python-SoXR package comes with modified version of libsoxr. See changes here.
These changes do not apply to dynamic-linked builds (e.g. conda-forge build).
To check the version of libsoxr, use soxr.__libsoxr_version__.

Credit and License

Python-SoXR is LGPL v2.1+ licensed, following libsoxr's license.

OSS libraries used

libsoxr (LGPLv2.1+)

The SoX Resampler library
https://sourceforge.net/projects/soxr/

Python-SoXR is a Python wrapper of libsoxr.

PFFFT (BSD-like)

PFFFT: a pretty fast FFT.
https://bitbucket.org/jpommier/pffft/

libsoxr dependency.