def analyze_audio(infile):
# Load the audio
(y, sr) = librosa.load(infile, target_sr=TARGET_SR)
# Generate a mel spectrogram
S = librosa.melspectrogram(y, sr, window_length=FFT_WINDOW, hop_length=HOP_SIZE, mel_channels=MEL_BINS)
# Generate per-band onsets
onsets = numpy.empty( (S.shape[0]+1, S.shape[1]-1) )
for i in range(MEL_BINS):
onsets[i,:] = librosa.beat.onset_strength(y, sr, window_length=FFT_WINDOW, hop_length=HOP_SIZE, S=S[i:(i+1),:])
pass
# Generate the global onset profile
onsets[-1,:] = librosa.beat.onset_strength(y, sr, window_length=FFT_WINDOW, hop_length=HOP_SIZE, S=S)
# Per-band onset correlation
P = 0.0
for t in xrange(0, onsets.shape[1] - BEAT_WINDOW - BEAT_CLEAR):
P = P + numpy.dot(numpy.diag(onsets[:, t]), onsets[:, (t+BEAT_CLEAR):(t+BEAT_WINDOW+BEAT_CLEAR)])
pass
return P
def __call__(self, y) -> np.ndarray:
melspec = librosa.melspectrogram(
y,
sr=self.sr,
n_mels=self.n_mels,
fmin=self.fmin,
fmax=self.fmax,
**self.kwargs,
)
melspec = librosa.power_to_db(melspec).astype(np.float32)
return melspec
def audio_onset_correlation(infile):
# Load the audio
(y, sr) = librosa.load(infile, target_sr=PARAMETERS['TARGET_SR'])
# Generate a mel spectrogram
S = librosa.melspectrogram(y, sr, window_length=PARAMETERS['FFT_WINDOW'], hop_length=PARAMETERS['HOP_SIZE'], mel_channels=PARAMETERS['MEL_BINS'])
# Generate per-band onsets
onsets = numpy.empty( (S.shape[0], S.shape[1]-1) )
for i in range(PARAMETERS['MEL_BINS']):
onsets[i,:] = librosa.beat.onset_strength(y, sr, window_length=PARAMETERS['FFT_WINDOW'], hop_length=PARAMETERS['HOP_SIZE'], S=S[i:(i+1),:])
pass
# Per-band onset correlation
P = 0.0
for t in xrange(0, onsets.shape[1] - PARAMETERS['BEAT_WINDOW'] - PARAMETERS['BEAT_CLEAR']):
P = P + numpy.dot(numpy.diag(onsets[:, t]), onsets[:, (t+PARAMETERS['BEAT_CLEAR']):(t+PARAMETERS['BEAT_WINDOW']+PARAMETERS['BEAT_CLEAR'])])
pass
return P
def continuousGridSearch( datasetDirectory, csvName ):
''' test '''
downsamplingFactors = np.array([1])
frameSizes = np.array([1024])
hopSizeScales = np.array([8])
windows = [np.hanning]
offsets = np.array([100])
''' For continuous
downsamplingFactors = np.array([1])
frameSizes = np.array([1024, 2048, 4096])
hopSizeScales = np.array([8, 4])
windows = [np.hanning]
offsets = np.array([20]) '''
# Get subdirectories for the input folder, corresponding to different MIDI files
directories = [os.path.join( datasetDirectory, folder ) for folder in os.listdir(datasetDirectory) if os.path.isdir(os.path.join(datasetDirectory, folder)) and folder[0] is not '.']
# The variations on the MIDI files
filenames = []
for shift in xrange( 0, 60, 10 ):
filenames += ['0-' + str(shift) + 'ms.wav']
filenames += ['1-' + str(shift) + 'ms.wav']
nFiles = len( filenames )
# Calculate number of tests about to be run
nTests = np.product( [len(dimension) for dimension in (directories, downsamplingFactors, frameSizes, hopSizeScales, windows, offsets)] )
print "About to run " + str( nTests ) + " tests."
# Keep track of which test is being run
testNumber = 0
startTime = time.time()
# Store the parameters corresponding to each result
gridSearchResults = collections.defaultdict(list)
# The data, being manipulated each step of the way
audioData = {}
audioDataDownsampled = {}
spectrograms = {}
ODFOutput = {}
synchronizationScores = {}
# Test to plot histograms
allAccuracies = np.zeros( nTests )
# Can we calculate the spectrogram from the previous spectrogram?
previousHopSizeScale = 0
for directory in directories:
# Read in wav data for each file
for file in filenames: audioData[file], fs = librosa.load( os.path.join( directory, file ) )
for downsamplingFactor in downsamplingFactors:
for file in filenames: audioDataDownsampled[file] = scipy.signal.decimate( audioData[file], downsamplingFactor )
for frameSize, window, hopSizeScale in itertools.product( frameSizes, windows, hopSizeScales ):
if hopSizeScale < previousHopSizeScale and np.mod( previousHopSizeScale, hopSizeScale ) == 0:
# Instead of calculating a new spectrogram, just grab the frames
newHopRatio = previousHopSizeScale/hopSizeScale
for file in filenames: spectrograms[file] = spectrograms[file][:, ::newHopRatio]
else:
# Calculate spectrograms - should not re-calculate if just the hop size changes.
for file in filenames: spectrograms[file] = librosa.melspectrogram( audioDataDownsampled[file], fs, window_length=frameSize, hop_length=frameSize/hopSizeScale, mel_channels=40)
previousHopSizeScale = hopSizeScale
# Get the onset detection function
for file in filenames: ODFOutput[file] = librosa.beat.onset_strength( audioDataDownsampled[file], fs, window_length=frameSize, hop_length=frameSize/hopSizeScale, S=spectrograms[file])
for offset in offsets:
# Compute the synchronization score for the syncrhonized and unsynchronized files
for n in xrange( nFiles/2 ):
synchronizationScores[n] = synchronizationScore.getScore( ODFOutput[filenames[2*n]], ODFOutput[filenames[2*n + 1]], offset=offset )
# Add in the ratio of the scores, we will take the per-MIDI-file-average later.
print "{} -> {}, {:.3f}% done in {:.3f} minutes".format( (directory, downsamplingFactor, frameSize, hopSizeScale, window.__name__, offset), synchronizationScores.values(), (100.0*testNumber)/nTests, (time.time() - startTime)/60.0)
testNumber += 1
gridSearchResults[(downsamplingFactor, frameSize, hopSizeScale, window.__name__, offset)] += [np.array(synchronizationScores.values())]
# Write out CSV results
csvWriter = csv.writer( open( csvName, 'wb' ) )
for parameters, results in gridSearchResults.items():
resultArray = np.array( results )
resultArray = (resultArray.T/np.max( resultArray, axis=1 )).T
csvWriter.writerow( list( parameters ) + list(np.mean( resultArray, axis=0 )) + [np.mean( np.diff( resultArray ), axis=0 )] + [np.mean( np.diff( resultArray ) )] + [np.mean( np.sum( np.diff( resultArray ) < 0, axis=1 ) )/(0.5*nFiles - 1)] )
