def audio_to_midi_melody(infile, outfile):
#Load audio and define parameters
audio = es.MonoLoader(filename = infile)()
winSize = 2048
hopSize = 128
Fs = 44100
pend = np.size(audio) - winSize #ultimo valor de ventana para calcular
pin =0 #Apuntador
totalTime = np.size(audio)/float(Fs)
#Instanciar funciones de essentia
w = es.Windowing(type = 'hann', size = winSize)
yin = es.PitchYinFFT(frameSize = winSize, minFrequency = 20, maxFrequency = 2000)
spectrum = es.Spectrum() #FUncion para calcular espectro por frame
calcEnergy = es.Energy()
# predMel = es.PredominantMelody(numberHarmonics = 15, filterIterations = 3, frameSize = 2048, hopSize = 128, minFrequency = 20, maxFrequency = 2000, minDuration = 50)
predMel = es.PredominantPitchMelodia(numberHarmonics = 15, filterIterations = 3, frameSize = 2048, hopSize = 128, minFrequency = 20, maxFrequency = 2000, minDuration = 50)
onsetDet = es.OnsetDetection(method = 'flux')
#Inicializar vectores
vectSize = np.ceil(np.size(audio))/hopSize
peakDet = es.PeakDetection(threshold = 0.15, maxPosition = vectSize, range = vectSize, maxPeaks = 300) #has to be normalized
melodyV = []
confV = []
energy = []
freqBands = []
for frame in es.FrameGenerator(audio, winSize, hopSize):
spec = spectrum(w(frame)) #Calcular el espectro
melody, conf = yin(spec) #Calcular melodia con algoritmo yin
melody = 0 if conf < 0.5 else melody #0 para valores con poca confidence
melodyV.append(melody)
ener = calcEnergy(frame) #Calcular energia por frame
energy.append(ener)
freqBands.append(es.FrequencyBands()(spec))
#onsets[i] = onsetDet(spec, spec)
#Convertir valores a float32 compatible con Essentia
melodyV = arrayCast(melodyV)
confV = arrayCast(confV)
energy = arrayCast(energy)
predMelody, confid = predMel(audio) #Calcular melodia con otro algoritmo
melodyV = freq2cent(melodyV)
predMelody = freq2cent(predMelody)
energy = np.log10(energy+.0000001)
energy = energy/float(max(energy)) #Normalize Energy
smoothEnergy = medfilt(arrayCast(energy), .7) #Usando filtro para alisar curva, valor en ms
onsets, amplitudes = peakDet(smoothEnergy) #detectar picos
noteOnsets = arrayCast(segmentByFreq(melodyV.copy(), totalTime))
bpm, bpmAmpl = es.NoveltyCurveFixedBpmEstimator()(es.NoveltyCurve()(arrayCast(freqBands)))
saveMIDI(outfile, noteOnsets, predMelody, bpm[0], Fs, hopSize)
#Graficar
plt.subplot(3,1,1)
plt.plot(np.linspace(0, totalTime, np.size(melodyV)), melodyV, 'b')
plt.plot(np.linspace(0, totalTime, np.size(predMelody)), predMelody, 'r')
plt.vlines(noteOnsets*hopSize/Fs, min(melodyV), max(melodyV))
plt.subplot(3,1,2)
plt.plot(np.linspace(0, totalTime, np.size(energy)),energy)
plt.vlines(onsets*hopSize/Fs, min(energy), max(energy))
plt.subplot(3,1,3)
plt.plot(np.linspace(0, totalTime, np.size(quantizeNote(cambioNota(melodyV)))),quantizeNote(cambioNota(melodyV)))
plt.show()
confV = []
energy = []
freqBands = []
for frame in es.FrameGenerator(audio, winSize, hopSize):
spec = spectrum(w(frame)) #Calcular el espectro
melody, conf = yin(spec) #Calcular melodia con algoritmo yin
melody = 0 if conf < 0.5 else melody #0 para valores con poca confidence
melodyV.append(melody)
ener = calcEnergy(frame) #Calcular energia por frame
energy.append(ener)
freqBands.append(es.FrequencyBands()(spec))
#onsets[i] = onsetDet(spec, spec)
#Convertir valores a float32 compatible con Essentia
melodyV = arrayCast(melodyV)
confV = arrayCast(confV)
energy = arrayCast(energy)
predMelody, confid = predMel(audio) #Calcular melodia con otro algoritmo
melodyV = freq2cent(melodyV)
predMelody = freq2cent(predMelody)
energy = np.log10(energy+.0000001)
energy = energy/float(max(energy)) #Normalize Energy
smoothEnergy = medfilt(arrayCast(energy), .7) #Usando filtro para alisar curva, valor en ms
onsets, amplitudes = peakDet(smoothEnergy) #detectar picos
noteOnsets = arrayCast(segmentByFreq(melodyV.copy()))