I am trying a simple task > calculate the cepstrum of a small audio sample, lift it and do the inverse process to build the audio file back. I obviously made a mistake as it does not work well import numpy as np from scipy.signal import lfilter import soundfile as sf import librosa import IPython.display as ipd# for audio output import matplotlib.pyplot as plt # matplotlib is for python graphs and display # Define the complex cepstrum function def complex_cepstrum(x, nfft=None): if nfft is None: nfft = len(x) X = np.fft.fft(x, n=nfft) logX = np.log(np.abs(X)) c = np.fft.ifft(logX, n=nfft) return c.real, c.imag, np.angle(X) # Define the liftering function def lifter(c_real, c_imag, lifter_order=22): c = c_real + 1j * c_imag c[0] = 0 # Set the DC component to zero lifter = 1 + (lifter_order / 2) * np.sin(np.pi * np.arange(len(c)) / lifter_order) return lifter * c.real, lifter * c.imag # Define the inverse complex cepstrum function def inverse_complex_cepstrum(c_real, c_imag, phase=None, n=None): if n is None: n = len(c_real) logX = c_real + 1j * c_imag if phase is not None: X = np.exp(logX + 1j * phase) else: X = np.exp(logX) x = np.fft.ifft(X, n=n).real return x # Load an audio sample x, fs = librosa.load('audio.wav', sr=44100, mono=True) # Pre-emphasis preemph = 0.97 x = lfilter([1, -preemph], [1], x) # Compute the complex cepstrum c_real, c_imag, phase = complex_cepstrum(x) # Apply liftering c_liftered_real, c_liftered_imag = lifter(c_real, c_imag) # Reconstruct the audio signal x_reconstructed = inverse_complex_cepstrum(c_liftered_real, c_liftered_imag, phase=phase) # Write the reconstructed signal to a file sf.write('audio_rec.wav', x_reconstructed, 44100, 'PCM_24') I note that I have searched quite some time before reaching you here. It generates a huge amount of (white) noise on top of the audio file which I can't understand where this comes from. If anyone has a direction where to look, thanks a lot. P Continue reading...
