692d1131c0f77f85a9068f22ae5e7bf26352a0b2,examples/noise_reduction_spectral_subtraction.py,,,#,29

Before Change

signal_norm = signal / np.abs(signal).max()

min_val = -80
max_val = -40
plt.figure()
plt.subplot(3, 1, 1)
plt.specgram(noisy_signal[:n-hop], NFFT=256, Fs=fs_s, vmin=min_val, vmax=max_val)
plt.title("Noisy Signal")
plt.subplot(3, 1, 2)
plt.specgram(processed_audio[hop:n], NFFT=256, Fs=fs_s, vmin=min_val, vmax=max_val)
plt.title("Denoised Signal")
plt.subplot(3, 1, 3)

After Change

beta = 30       // An overestimation factor to "push" the suppression towards db_reduc.
alpha = 7       // An exponential factor to tune the suppresion (see documentation of "SpectralSub").

plot_spec = True


Prepare input file

fs_s, signal = wavfile.read(os.path.join(os.path.dirname(__file__),
                                         "input_samples",
                                         "cmu_arctic_us_aew_a0001.wav"))
fs_n, noise = wavfile.read(os.path.join(os.path.dirname(__file__),
                                        "input_samples",
                                        "doing_the_dishes.wav"))
if fs_s != fs_n:
    raise ValueError("Signal and noise WAV files should have same sampling "
                     "rate for this example.")

// truncate to same length
if len(noise) < len(signal):
    raise ValueError("Length of signal file should be longer than noise "
                     "file for this example.")
noise = noise[:len(signal)]

// weight noise according to desired SNR
signal_level = np.linalg.norm(signal)
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 = signal + noise_weighted
noisy_signal /= np.abs(noisy_signal).max()
noisy_signal -= noisy_signal.mean()
wavfile.write(os.path.join(os.path.dirname(__file__), "output_samples",
                           "denoise_input.wav"), fs_s,
              noisy_signal.astype(np.float32))


Create STFT and SCNR objects

hop = nfft // 2
window = pra.hann(nfft, flag="asymmetric", length="full")
stft = pra.transform.STFT(nfft, hop=hop, analysis_window=window,
                          streaming=True)

scnr = SpectralSub(nfft, db_reduc, lookback, beta, alpha)
lookback_time = hop/fs_s * lookback
print("Lookback : %f seconds" % lookback_time)


Process as in real-time

// collect the processed blocks
processed_audio = np.zeros(signal.shape)
n = 0
while noisy_signal.shape[0] - n >= hop:

    // SCNR in frequency domain
    stft.analysis(noisy_signal[n:(n+hop), ])
    gain_filt = scnr.compute_gain_filter(stft.X)

    // back to time domain
    processed_audio[n:n+hop, ] = stft.synthesis(gain_filt*stft.X)

    // update step
    n += hop



Save and plot spectrogram

wavfile.write(os.path.join(os.path.dirname(__file__),
                           "output_samples",
                           "denoise_output_SpectralSub.wav"), fs_s,
              pra.normalize(processed_audio).astype(np.float32))
print("Noisy and denoised file written to: "%s"" %
      os.path.join(os.path.dirname(__file__), "output_samples"))

signal_norm = signal / np.abs(signal).max()

if plot_spec:
    min_val = -80
    max_val = -40
    plt.figure()
    plt.subplot(3, 1, 1)
    plt.specgram(noisy_signal[:n-hop], NFFT=256, Fs=fs_s,
                 vmin=min_val, vmax=max_val)
    plt.title("Noisy Signal")
    plt.subplot(3, 1, 2)
    plt.specgram(processed_audio[hop:n], NFFT=256, Fs=fs_s,
                 vmin=min_val, vmax=max_val)
    plt.title("Denoised Signal")
    plt.subplot(3, 1, 3)
    plt.specgram(signal_norm[:n-hop], NFFT=256, Fs=fs_s,
                 vmin=min_val, vmax=max_val)
    plt.title("Original Signal")
    plt.tight_layout(pad=0.5)
    plt.show()

Instances