audio_degrader
Latest version: 1.3.1
Audio degradation toolbox in python, with a command-line tool. It is useful to apply controlled degradations to audio.
Installation
pip install audio-degrader
The program depends on pysox, so you might need to install sox (and libsox-fmt-mp3 for mp3 encoding). Go to https://github.com/rabitt/pysox to have more details about it.
Available degradations
convolution,impulse_response,level: Convolve input with specified impulse response
parameters:
impulse_response: Full path, URL (requires wget), or relative path (see -l option)
level: Wet level (0.0=dry, 1.0=wet)
example:
convolution,impulse_responses/ir_classroom.wav,1.0
dr_compression,degree: Apply dynamic range compression
parameters:
degree: Degree of compression. Presets from 0 (soft) to 3 (hard)
example:
dr_compression,0
equalize,central_freq,bandwidth,gain: Apply a two-pole peaking equalisation (EQ) filter
parameters:
central_freq: Central frequency of filter in Hz
bandwidth: Bandwith of filter in Hz
gain: Gain of filter in dBs
example:
equalize,100,50,-10
gain,value: Apply gain expressed in dBs
parameters:
value: Gain value [dB]
example:
gain,6
mix,noise,snr: Mix input with a specified noise. The noise can be specified with its full path, URL (requires wget installed), or relative to the resources directory (see -l option)
parameters:
noise: Full or relative path (to resources dir) of noise
snr: Desired Signal-to-Noise-Ratio [dB]
example:
mix,sounds/ambience-pub.wav,6
mp3,bitrate: Emulate mp3 transcoding
parameters:
bitrate: Quality [bps]
example:
mp3,320k
normalize: Normalize amplitude of audio to range [-1.0, 1.0]
parameters:
example:
normalize
pitch_shift,pitch_shift_factor: Apply pitch shifting
parameters:
pitch_shift_factor: Pitch shift factor
example:
pitch_shift,0.9
resample,sample_rate: Resample to given sample rate
parameters:
sample_rate: Desired sample rate [Hz]
example:
resample,8000
speed,speed: Change playback speed
parameters:
speed: Playback speed factor
example:
speed,0.9
time_stretch,time_stretch_factor: Apply time stretching
parameters:
time_stretch_factor: Time stretch factor
example:
time_stretch,0.9
trim_from,start_time: Trim audio from a given start time
parameters:
start_time: Trim start [seconds]
example:
trim_from,0.1
Usage of python package
import audio_degrader as ad
audio_file = ad.AudioFile('input.wav', './tmp_dir')
for d in ad.ALL_DEGRADATIONS.values():
print ad.DegradationUsageDocGenerator.get_degradation_help(d)
degradations = ad.ParametersParser.parse_degradations_args([
'normalize',
'gain,6',
'dr_compression,3',
'equalize,500,10,30'])
for d in degradations:
audio_file.apply_degradation(d)
audio_file.to_wav('output.wav')
audio_file.delete_tmp_files()Usage of command-line tool
The script audio_degrader is installed along with the python package.
# e.g. mix with restaurant08.wav with snr=10db, then amplifies 6db, then compress dynamic range
$ audio_degrader -i input.mp3 -d mix,https://github.com/hagenw/audio-degradation-toolbox/raw/master/AudioDegradationToolbox/degradationData/PubSounds/restaurant08.wav,10 gain,6 dr_compression,3 -o out.wav
# for more details:
$ audio_degrader --help
A small set of sounds and impulse responses are installed along with the script, which can be listed with:
$ audio_degrader -l
# these relative paths can be used directly in the script too:
$ audio_degrader -i input.mp3 -d mix,sounds/applause.wav,-3 gain,6 -o out.wav
Applications
- Evaluate Music Information Retrieval systems under different degrees of degradations
- Prepare augmented data for training of machine learning systems
It is similar to the Audio Degradation Toolbox in Matlab by Sebastian Ewert and Matthias Mauch (for Matlab).
Some examples
# Mix input with a sound / noise (e.g. using installed resources)
$ audio_degrader -i input.wav -d mix,sounds/applause.wav,-3 -o out.wav
# Instead of paths, we can also use URLs
$ audio_degrader -i input.wav -d mix,https://www.pacdv.com/sounds/ambience_sounds/airport-security-1.mp3,-3 -o out.wav
# Microphone recording style
$ audio_degrader -i input.wav -d gain,-15 mix,sounds/ambience-pub.wav,18 convolution,impulse_responses/ir_smartphone_mic_mono.wav,0.8 dr_compression,2 equalize,50,100,-6 normalize -o out.wav
# Resample and normalize
$ audio_degrader -i input.mp3 -d resample,8000 normalize -o out.wav
# Convolution (again impulse responses can be resources, full paths or URLs)
$ audio_degrader -i input.wav -d convolution,impulse_responses/ir_classroom_mono.wav,0.7 -o out.wav
$ audio_degrader -i input.wav -d convolution,http://www.cksde.com/sounds/month_ir/FLANGERSPACE%20E001%20M2S.wav,0.7 -o out.wav
Audio formats
Input
audio_degrader relies on pysox for reading, so any format accepted by pysox should be ok.
Output
audio_degrader output format is always WAV pcm stereo with 32 bits per sample (sample rate from original audio file).
This output wav file can be easily coverted into another format with e.g. ffmpeg:
$ ffmpeg -i out.wav -b:a 320k out.mp3
$ ffmpeg -i out.wav -ac 2 -ar 44100 -acodec pcm_s16le out_formatted.wav