AI algorithms are mostly data-driven, and the quality of the data determines the quality of the model to some extent. This leads to the inherent shortcoming of deep learning, and data augmentation is an effective way to solve this problem.
This repo supports audio data augmentations such as :
- reverberation
- background noise
- distortion
- packet loss simulation
- farfield effect
- speed perturbation
After those time domain augmentations, one can apply feature extraction step.
To install the released stable version, enter the REPL mode
] add SpeechAugmentor
Pkg.add("SpeechAugment")To install the development version, enter the REPL mode
] add https://github.com/sonosole/SpeechAugment.jl.gitusing WAV
using SpeechAugment
# 1. read a wav file as a speech example
batchsize = 8;
data,fs = wavread("/XXPath/ASpeechExample.wav");
# 2. init all the augmentation functions you want
echo = initAddEcho(fs, (0.05,0.4), (3.0,3.2,2.5,3.5,2.0,3.0));
noise = initAddNoise("XXPathFullOfNoiseWAVs", 2, (5,15));
clip = initClipWav((0.5,2.0));
drop = initDropWav(fs, (0.09,0.15));
far = initFarfieldWav(fs, (0.4,0.9));
speed = initSpeedWav((0.8,1.2));
# 3. make a function list or array
fnlist = [echo noise clip drop far speed];
# 4. augment #batchSize audios
wavs = Vector(undef, batchsize)
for i = 1:batchsize
wavs[i] = copy(data)
end
wavs = augmentWavs(fnlist, wavs)
for i = 1:batchsize
wavwrite(wavs[i], "A$i.wav",Fs=16000,nbits=32)
end
# there is also a function called `augmentWav`
# it augments one audio into multiple audios.
audios = augmentWav(fnlist, data, batchsize)
for i = 1:batchsize
wavwrite(audios[i], "B$i.wav",Fs=16000,nbits=32)
endinitAddEcho(fs::Number, T₆₀Span::NTuple{2,Number}, roomSpan::NTuple{6,Number}) -> addecho(wav::Array)fssampling rateT₆₀Spaneffective reverberation time e.g. (minT60, maxT60)roomSpanroom size e.g. (MinL, MaxL, MinW, MaxW, MinH, MaxH)
initAddNoise(path::String, period::Int, dBSpan::NTuple{2,Number}) -> addnoise(speech::Array)patha path only full of noise WAVsperiodevery #period it would change another noise wav.dBSpanspan of SNR e.g. (mindB, maxdB)
initClipWav(clipSpan::NTuple{2,Number}) -> clipwav(wav::Array)clipSpanhow much it would clip a wav e.g. (0.5,2.0)
initDropWav(fs::Real, ratioSpan::NTuple{2,Number}) -> dropwav(wav::Array)fssampling rateratioSpanspan of droping ratio e.g. (0.02, 0.09). 1.0 is the uplimit.
initFarfieldWav(fs::Real, maxvalueSpan::NTuple{2,Number}) -> farfieldwav(wav::Array)fssampling ratemaxvalueSpanranges from (0.0,1.0). Smaller means farther away. (0.2, 0.9) is recommended.
initSpeedWav(speedSpan::NTuple{2,Number}) -> speedwav(wav::Array)speedSpanrange of speed perturbation. (0.85, 1.15) is recommended.
All the NTuple{2,Number} parameters should follow the small on the left and the big on the right i.e. (minvalue, maxvalue). To precisely control the extent of augmentation, the below functions could be used:
- addEcho
- addNoise
- clipWav
- dropWav
- farfieldWav
- speedWav
For details, check the documentation or enter the help?> mode.
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent