demo_denoise_subspace.py

import numpy as np
import os
from scipy.io import wavfile
import pyroomacoustics as pra

frame_len = 80
mu = 30         # higher value can give more suppression but more distortion
thres = 0.01    # threshold between (signal+noise) and noise

# How many frames to look back for the noise floor estimate.
lookback = 20

# How many samples to skip when estimating the covariance matrices
# with past samples.
skip = 1


"""
preparation of noisy input signal
"""
snr = 5         # SNR of input signal
speech_file = "./input/arctic_a0010.wav"
noise_file = "./input/exercise_bike.wav"

# load WAV files, should have same sampling rates!
fs_s, audio = wavfile.read(speech_file)
fs_n, noise = wavfile.read(noise_file)
assert fs_s == fs_n, "audio and noise wav files should have same sampling rate!"

# truncate to same length
noise = noise[:len(audio)]

# weight noise according to desired SNR
signal_level = np.linalg.norm(audio)
noise_level = np.linalg.norm(noise)
noise_fact = signal_level / 10**(snr / 20)
noise_weighted = noise * noise_fact / noise_level

# add signal and noise
noisy_signal = audio + noise_weighted
noisy_signal /= np.abs(noisy_signal).max()
noisy_signal -= noisy_signal.mean()


"""
noise reduction via subspace method
"""
# import or create `noisy_signal`
denoised_signal = pra.denoise.subspace.apply_subspace(
    noisy_signal, frame_len=frame_len, mu=mu,
    lookback=lookback, skip=skip, thresh=thres)

# output results
outputDir = './output'
os.makedirs(outputDir, exist_ok=True)

print("Output noisy speech.")
wavfile.write("{}/noisy_speech.wav".format(outputDir), fs_s, noisy_signal)

print("Output denoised speech.")
wavfile.write("{}/denoised_speech.wav".format(outputDir), fs_s, denoised_signal)