In audio data augmentation, we often add a noise to a clean audio. In this tutorial, we will introduce you how to do using python.
For example, we may see:
In each augmentation, one noise file is randomly selected in the MUSAN database and added to the recording with [0, 5, 10, 15]
dB SNR.
In order to understand what is SNR, you can read:
Compute Audio SNR (Signal-to-Noise Ratio) in Python – Python Tutorial
How to mix a clean audio with a noise file in python?
Here we will create a python function to implement it.
For example:
import numpy
import librosa
import soundfile
def add_noise_snr(audio, noise, snr):
clean_db = 10 * numpy.log10(numpy.mean(audio ** 2) + 1e-4)
audio_length = audio.shape[0]
if noise.shape[0] <= audio_length:
shortage = audio_length - noise.shape[0]
noise = numpy.pad(noise, (0, shortage), 'wrap')
noise_db = 10 * numpy.log10(numpy.mean(noise ** 2) + 1e-4)
add_noise = numpy.sqrt(10 ** ((clean_db - noise_db - snr) / 10)) * noise
mix_audio = audio + add_noise
return mix_audio
def read_audio(file_audio):
audio, sr = librosa.load(file_audio, sr=8000, mono=True)
return audio
def save_wav(audio, fx, sr = 8000):
soundfile.write(fx, audio, sr, "PCM_16")In this code, add_noise_snr(audio, noise, snr) function will add noise data to a clean audio data base on a snr value.
We can use this function as follows:
wav_file = r"speech-us-gov-0010.wav" audio = read_audio(wav_file) noise_file = "music-jamendo-0039.wav" noise = read_audio(noise_file) audio_noise = add_noise_snr(audio, noise, snr = 0) save_wav(audio_noise, "audio_noise_0.wav") audio_noise = add_noise_snr(audio, noise, snr = 5) save_wav(audio_noise, "audio_noise_5.wav") audio_noise = add_noise_snr(audio, noise, snr = 10) save_wav(audio_noise, "audio_noise_10.wav")
Then, we will create 3 mixed audio files.
You should notice: the larger of the value of SNR, the clearer of the clean audio in mixed audio file.