computeMelSpectrogram.py

Go to the documentation of this file.

# -*- coding: utf-8 -*-
 
import numpy as np
 
from pyACA.computeSpectrogram import computeSpectrogram
from pyACA.ToolPreprocAudio import ToolPreprocAudio
from pyACA.ToolComputeHann import ToolComputeHann
from pyACA.ToolFreq2Mel import ToolFreq2Mel
from pyACA.ToolMel2Freq import ToolMel2Freq
 
 
 
def computeMelSpectrogram(x, f_s, afWindow=None, bLogarithmic=True, iBlockLength=4096, iHopLength=2048, iNumMelBands=128, fMaxInHz=None):
 
    if not fMaxInHz:
        fMaxInHz = f_s / 2
 
    # Pre-process: down-mix, normalize, zero-pad
    x = ToolPreprocAudio(x)
 
    if afWindow is None:
        # Compute window function for FFT
        afWindow = ToolComputeHann(iBlockLength)
 
    assert(afWindow.shape[0] == iBlockLength), "parameter error: invalid window dimension"
 
    # Compute spectrogram (in the real world, we would do this block by block)
    [X, f, t] = computeSpectrogram(x, f_s, None, iBlockLength, iHopLength)
 
    # Compute Mel filters
    H, f_c = generateMelFb_I(iBlockLength, f_s, iNumMelBands, fMaxInHz)
 
    M = np.matmul(H, X)
 
    if bLogarithmic:
        # Convert amplitude to level (dB)
        M = 20 * np.log10(M + 1e-12)
 
    return M, f_c, t
 
 
def generateMelFb_I(iFftLength, f_s, iNumFilters, f_max):
 
    # initialization
    f_min = 0
    f_max = min(f_max, f_s / 2)
    f_fft = np.linspace(0, f_s / 2, iFftLength // 2 + 1)
    H = np.zeros((iNumFilters, f_fft.size))
 
    # compute center band frequencies
    mel_min = ToolFreq2Mel(f_min)
    mel_max = ToolFreq2Mel(f_max)
    f_mel = ToolMel2Freq(np.linspace(mel_min, mel_max, iNumFilters + 2))
 
    f_l = f_mel[0:iNumFilters]
    f_c = f_mel[1:iNumFilters + 1]
    f_u = f_mel[2:iNumFilters + 2]
 
    afFilterMax = 2 / (f_u - f_l)
 
    # compute the transfer functions
    for c in range(iNumFilters):
        H[c] = np.logical_and(f_fft > f_l[c], f_fft <= f_c[c]) * \
            afFilterMax[c] * (f_fft-f_l[c]) / (f_c[c]-f_l[c]) + \
            np.logical_and(f_fft > f_c[c], f_fft < f_u[c]) * \
            afFilterMax[c] * (f_u[c]-f_fft) / (f_u[c]-f_c[c])
 
    return H, f_c
 
 
 
def computeMelSpectrogramCl(cPath):
    from pyACA.ToolReadAudio import ToolReadAudio
 
    # read audio file
    [f_s, x] = ToolReadAudio(cPath)
    
    # compute feature
    [M, f, t] = computeMelSpectrogram(x, f_s)
 
    return M, f, t
 
 
if __name__ == "__main__":
    import argparse
 
    # add command line args and parse them
    parser = argparse.ArgumentParser(description='Compute key of wav file')
    parser.add_argument('--infile', metavar='path', required=False,
                        help='path to input audio file')
 
    # retrieve command line args
    args = parser.parse_args()
    cPath = args.infile
 
    # only for debugging
    if __debug__:
        if not cPath:
            cPath = "../ACA-Plots/audio/sax_example.wav"
 
    # call the function
    computeMelSpectrogramCl(cPath)