def sortFiles(self):
# Get list of wav files, recursively
fileList = utility.getFiles(self.snippetDirectory, '.wav')
# Array of per-file, per-note tonalities. Each file gets a score for each pitch
tonalities = np.zeros((len(fileList), self.nNotes))
print "Getting tonalities..."
for n in np.arange(len(fileList)):
# Get audio data
audioData, fs = utility.getWavData(fileList[n])
# Make sure the sampling rate matches...
assert fs == self.fs
# Get the tonality scores for all notes for this file
tonalities[n] = self.getTonality(audioData)
# Hash for the fundamental frequencies of each file
fileFrequencies = {}
# Find the best file for each note
print "Finding best candidates for notes"
for n in np.arange(self.nNotes):
# Get the tonality scores
tonalitiesForThisNote = tonalities[:, n]
# Get the sorted indices of tonality scores
tonalitiesSort = np.argsort(tonalitiesForThisNote)[::-1]
# Keep track of which sorted array index we're getting the frequency for
sortedIndex = 0
# What frequency is the note we're looking for?
targetHz = utility.midiToHz(self.baseNote + n)
# What's the detected Hz of the note?
detectedHz = 1.0
# Until we find an audio file whose YIN detected pitch is sufficiently close
while (
(targetHz / detectedHz) <
(1 - PITCH_TOLERANCE) or (targetHz / detectedHz) >
(1 + PITCH_TOLERANCE)) and sortedIndex < tonalitiesSort.shape[
0] and sortedIndex < PITCHES_TO_RUN:
# If this file has not been YIN analyzed yet, analyze it
if not fileFrequencies.has_key(tonalitiesSort[sortedIndex]):
audioData, fs = utility.getWavData(
fileList[tonalitiesSort[sortedIndex]])
# ... and store it so that you don't have to calculate it next time
fileFrequencies[tonalitiesSort[
sortedIndex]] = self.yinPitchDetect(audioData)
detectedHz = fileFrequencies[tonalitiesSort[sortedIndex]]
# Check the next file next time
sortedIndex += 1
# If we didn't run out of files, copy out the file that we found (with the closest pitch)
if sortedIndex < len(fileList):
shutil.copy(
fileList[tonalitiesSort[sortedIndex - 1]],
os.path.join(self.destinationDirectory,
str(n + self.baseNote) + ".wav"))
else:
shutil.copy(
fileList[tonalitiesSort[0]],
os.path.join(self.destinationDirectory,
str(n + self.baseNote) + ".wav"))
def sortFiles( self ):
# Get list of wav files, recursively
fileList = utility.getFiles( self.snippetDirectory, '.wav' )
# Array of per-file, per-note tonalities. Each file gets a score for each pitch
tonalities = np.zeros( ( len( fileList ), self.nNotes ) )
print "Getting tonalities..."
for n in np.arange( len( fileList ) ):
# Get audio data
audioData, fs = utility.getWavData( fileList[n] )
# Make sure the sampling rate matches...
assert fs == self.fs
# Get the tonality scores for all notes for this file
tonalities[n] = self.getTonality( audioData )
# Hash for the fundamental frequencies of each file
fileFrequencies = {}
# Find the best file for each note
print "Finding best candidates for notes"
for n in np.arange( self.nNotes ):
# Get the tonality scores
tonalitiesForThisNote = tonalities[:,n]
# Get the sorted indices of tonality scores
tonalitiesSort = np.argsort( tonalitiesForThisNote )[::-1]
# Keep track of which sorted array index we're getting the frequency for
sortedIndex = 0
# What frequency is the note we're looking for?
targetHz = utility.midiToHz( self.baseNote + n )
# What's the detected Hz of the note?
detectedHz = 1.0
# Until we find an audio file whose YIN detected pitch is sufficiently close
while ((targetHz/detectedHz) < (1 - PITCH_TOLERANCE) or (targetHz/detectedHz) > (1 + PITCH_TOLERANCE)) and sortedIndex < tonalitiesSort.shape[0] and sortedIndex < PITCHES_TO_RUN:
# If this file has not been YIN analyzed yet, analyze it
if not fileFrequencies.has_key( tonalitiesSort[sortedIndex] ):
audioData, fs = utility.getWavData( fileList[tonalitiesSort[sortedIndex]] )
# ... and store it so that you don't have to calculate it next time
fileFrequencies[tonalitiesSort[sortedIndex]] = self.yinPitchDetect( audioData )
detectedHz = fileFrequencies[tonalitiesSort[sortedIndex]]
# Check the next file next time
sortedIndex += 1
# If we didn't run out of files, copy out the file that we found (with the closest pitch)
if sortedIndex < len( fileList ):
shutil.copy( fileList[tonalitiesSort[sortedIndex-1]], os.path.join( self.destinationDirectory, str(n + self.baseNote) + ".wav" ) )
else:
shutil.copy( fileList[tonalitiesSort[0]], os.path.join( self.destinationDirectory, str(n + self.baseNote) + ".wav" ) )
# Run function as script
if __name__ == "__main__":
import sys
import os
import utility
if len(sys.argv) < 2:
print "Usage: %s filename.mp3|filename.wav" % sys.argv[0]
sys.exit(-1)
# Wav or mp3?
basename, extension = os.path.splitext( sys.argv[1] )
if extension == '.mp3':
print "Getting mp3 data ..."
audioData, fs = utility.getMp3Data( sys.argv[1] )
elif extension == '.wav':
print "Getting wav data ..."
audioData, fs = utility.getWavData( sys.argv[1] )
else:
print "Not .wav or .mp3."
sys.exit(-1)
hop = 1024
frameSize = 4096
spectrogram = utility.getSpectrogram( audioData, hop=hop, frameSize=frameSize )
print "Separating harmonic and percussive components ..."
start = time()
seperator = HarmonicPercussiveSeparator( spectrogram )
end = time()
print "Took ", end - start
utility.plotSpectrogram( spectrogram )
utility.plotSpectrogram( seperator.harmonicSpectrogram )
utility.plotSpectrogram( seperator.percussiveSpectrogram )
# The data, being manipulated each step of the way
audioData = {}
audioDataDownsampled = {}
spectrograms = {}
ODFOutput = {}
# 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 - should try downsampling factors
for file in filenames:
audioData[file], fs = utility.getWavData(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] = utility.getSpectrogram(
audioDataDownsampled[file],
hop=frameSize / hopSizeScale,
