1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
|
# Python-SoXR
[](https://github.com/dofuuz/python-soxr) [](https://pypi.org/project/soxr/) [](https://anaconda.org/conda-forge/soxr-python) [](https://repology.org/project/python:soxr/versions) [](https://python-soxr.readthedocs.io)
High quality, one-dimensional sample-rate conversion library for Python.
- Homepage: https://github.com/dofuuz/python-soxr
- Documentation: https://python-soxr.readthedocs.io
- PyPI: https://pypi.org/project/soxr/
Keywords: Resampler, Audio resampling, Samplerate conversion, DSP(Digital Signal Processing)
Python-SoXR is a Python wrapper of [libsoxr](https://sourceforge.net/projects/soxr/).
## 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
```python
import soxr
y = soxr.resample(
x, # input array – 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.
```python
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 – 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, ...])
📝 [More code examples](https://dofuuz.github.io/dsp/2024/05/26/sample-rate-conversion-in-python.html)
## 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.
Library | Time on CPU (ms)
------------------------ | ----------------
soxr (HQ) | 10.8
torchaudio | 13.8
soxr (VHQ) | 14.5
scipy.signal.resample | 21.3
lilfilter | 24.7
julius | 31
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.
- https://sourceforge.net/p/soxr/wiki/Home/
- http://sox.sourceforge.net/SoX/Resampling ([archive](https://web.archive.org/web/20230626144127/https://sox.sourceforge.net/SoX/Resampling))
- https://sourceforge.net/p/soxr/code/ci/master/tree/src/soxr.h
Python-SoXR package comes with [modified version](https://github.com/dofuuz/soxr) of libsoxr. [See changes here](https://github.com/dofuuz/soxr/compare/0.1.3...master).
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.
|
