def det(tf):
# fmt: off
sp.check_call([
b"ffmpeg", b"-nostdin", b"-hide_banner", b"-v", b"fatal", b"-ss",
b"13", b"-y", b"-i",
fsenc(sys.argv[1]), b"-map", b"0:a:0", b"-ac", b"1", b"-ar", b"22050",
b"-t", b"300", b"-f", b"f32le",
fsenc(tf)
])
# fmt: on
with open(tf, "rb") as f:
d = np.fromfile(f, dtype=np.float32)
try:
# 98% accuracy on jcore
c = vamp.collect(d, 22050, "beatroot-vamp:beatroot")
cl = c["list"]
except:
# fallback; 73% accuracy
plug = "vamp-example-plugins:fixedtempo"
c = vamp.collect(d, 22050, plug, parameters={"maxdflen": 40})
print(c["list"][0]["label"].split(" ")[0])
return
# throws if detection failed:
bpm = float(cl[-1]["timestamp"] - cl[1]["timestamp"])
bpm = round(60 * ((len(cl) - 1) / bpm), 2)
print(f"{bpm:.2f}")
def process(frames):
#assert frames == client.blocksize
#assert (len(l_f) == len(r_f))
l_f = client.inports[0].get_array()
r_f = client.inports[1].get_array()
'''
l_f = np.frombuffer(l.get_buffer(), dtype=float)
r_f = np.frombuffer(r.get_buffer(), dtype=float)
'''
m_f = np.add(l_f, r_f)
m_f = np.true_divide(m_f, 2.0)
try:
data = vamp.collect(m_f, client.samplerate, "vamp-libxtract:loudness")
vector = data['vector']
L = vector[1]
if L > 0.0:
data = vamp.collect(m_f, client.samplerate, "pyin:pyin")
L = data['list']
print L[1]['values']
except:
pass
def analyzeAudio(audio_array, sample_rate, channels):
print "Found Vamp plugins:"
for name in vamp.list_plugins():
print " " + name
# convert audio array to float format expected by vamp
audio_array2 = audio_array / float(numpy.iinfo(numpy.int16).max)
audio_array2 = audio_array2.reshape((len(audio_array2)/channels,channels))[:,0]
if False:
# chuncked
results = []
for subarray in numpy.array_split(audio_array2, numpy.ceil(audio_array2.shape[0] / float(BUFFER_SIZE)), axis=0):
if DEBUG: print "subarray size: " + str(subarray.shape[0])
#assert(subarray.shape[0] == BUFFER_SIZE)
results += [vamp.collect(subarray, sample_rate, VAMP_PLUGIN)]
else:
# all at once
results = vamp.collect(audio_array2, sample_rate, VAMP_PLUGIN)
if DEBUG:
print "Vamp plugin output:"
for result in results["list"][:15]:
print " %s" % result
results = [(float(row["timestamp"]), float(row["duration"]), int(row["values"][0])) for row in results["list"]]
return results
def test_process_summary_param():
buf = input_data(blocksize * 10)
rdict = vamp.collect(buf, rate, plugin_key, "input-summary", {"produce_output": False})
assert "vector" in rdict
step, results = rdict["vector"]
assert len(results) == 0
rdict = vamp.collect(buf, rate, plugin_key, "input-summary", {"produce_output": True})
assert "vector" in rdict
step, results = rdict["vector"]
assert len(results) > 0
def test_process_summary_param():
buf = input_data(blocksize * 10)
rdict = vamp.collect(buf, rate, plugin_key, "input-summary",
{"produce_output": False})
assert ("vector" in rdict)
step, results = rdict["vector"]
assert len(results) == 0
rdict = vamp.collect(buf, rate, plugin_key, "input-summary",
{"produce_output": True})
assert ("vector" in rdict)
step, results = rdict["vector"]
assert len(results) > 0
def extract_F0_pYIN_vamp(folder, fn, sr=22050, H=221, N=2048):
''' Given an audio file, use the pYIN Vamp Plug-in to extract the F0
and the confidence of each frame to be voiced.
'''
# load wave file
x, fs = librosa.load(os.path.join(folder, fn), sr=sr)
# pYIN parameters
param = {'threshdistr': 2, 'outputunvoiced': 2, 'precisetime': 0}
'''
outputs:
'voicedprob' outputs the probability of each frame to be voiced
'smoothedpitchtrack' outputs the smoothed pitch track
parameters:
'outputunvoiced': 0 (No), 1 (Yes as zeros), 2 (Yes as negative frequencies)
'''
pYIN_f0_output = vamp.collect(x,
sr,
"pyin:pyin",
output='smoothedpitchtrack',
parameters=param,
step_size=H,
block_size=N)['vector']
pYIN_voiced_prob = vamp.collect(x,
sr,
"pyin:pyin",
output='voicedprob',
parameters=param,
step_size=H,
block_size=N)['vector']
time_step = float(pYIN_f0_output[0])
F0 = pYIN_f0_output[1]
voiced_prob = pYIN_voiced_prob[1]
timestamp = np.arange(start=0,
stop=(len(F0) - 0.5) * time_step,
step=time_step) #[:-1]
traj = np.vstack([timestamp, F0, voiced_prob]).transpose()
if not os.path.exists(os.path.join(folder, 'pYIN')):
os.mkdir(os.path.join(folder, 'pYIN'))
pd.DataFrame(traj).to_csv(os.path.join(folder, 'pYIN', fn[:-3] + 'csv'),
header=None)
#print("{} F0 curve saved to {}".format(fn, os.path.join(folder, 'pYIN', fn[:-3] + 'csv')))
return traj
def load_audio_data_from_url(self, url, audio_path):
"""
' 목적 : 음원의 유튜브 url로부터 음원의 코드정보(by vamp plugin)와 박자정보를 추출하는 함수
' 리턴값 : 1. 예측된 코드와 코드의 출현시간값 리스트 / 2. 예측된 박자의 시간값 리스트 / 3. 다운로드 받은 오디오 파일명 / 4. vamp 플러그인이 박자를 성공적으로 추출했는지 확인하는 Boolean값
"""
# 사용자가 앱을 수행하다 중단했을시, 제거되지 않고 남아있는 파일을 제거하여 youtube_dl 오류방지
files = glob.glob(audio_path + "/*.mp4")
for x in files:
if not os.path.isdir(x):
os.remove(x)
ydl_opts = {
'format': 'best',
'outtmpl': audio_path + '/%(title)s.%(ext)s',
}
with youtube_dl.YoutubeDL(ydl_opts) as ydl:
ydl.download([url])
files = glob.glob(audio_path + "/*.mp4")
audio_file_name = ''
signal, sr = 0, 0
for x in files:
if not os.path.isdir(x):
filename = os.path.splitext(x)
audio_file_name = filename[0].split(audio_path + '/')[1]
signal, sr = librosa.load(x)
os.remove(x)
# 악보 구조를 생성하는데에 필요한 박자 데이터 추출
predicted_beat_list = vamp.collect(signal, sr,
"beatroot-vamp:beatroot")
isEmptyBeatList = False
if (predicted_beat_list['list'] == []):
isEmptyBeatList = True
# 악보를 생성하는데에 필요한 코드 데이터 추출
predicted_chord_list = vamp.collect(signal, sr, "nnls-chroma:chordino")
chord_list = []
for i in predicted_chord_list['list']:
i['timestamp'] = float(i['timestamp'])
chord_list.append(i)
beat_list = []
for i in predicted_beat_list['list']:
beat_list.append(float(i['timestamp']))
return chord_list, beat_list, audio_file_name, isEmptyBeatList
def FeatureThread():
global currentTempo, tempoGram, predictTempo
global chords
rb = RingBuffer(8)
while True:
os.system('jack_capture_ms -d 1000 out.wav > /dev/null 2>&1')
y, sr = librosa.load('out.wav', sr=client.samplerate)
onset_env = librosa.onset.onset_strength(y, sr=sr)
tempo = librosa.beat.tempo(onset_envelope=onset_env, sr=sr)
rb.append(int(numpy.around(tempo)))
tempogram = numpy.array(rb.get())
unique, counts = numpy.unique(tempogram, return_counts=True)
max_unique_counts = numpy.asarray((unique, counts)).T
currentTempo, tempoGram, predictTempo = tempo[0], rb.get(), max_unique_counts[0][0]
vdata = vamp.collect(y, client.samplerate, "nnls-chroma:chordino")
idx = 0
for i in vdata['list']:
c = vdata['list'][idx]['label']
if c != 'N':
chords = c
idx = idx + 1
os.system('rm -f out.wav')
if stopThreads:
break
def audio_to_pitch_melodia(wav_data,
fs=44100,
minfqr=55.0,
maxfqr=1760.0,
voicing=0.2,
minpeaksalience=0.0):
# 用代码调用mtg-melodia时只能使用默认的block=2048，step=128
params = dict(minfqr=minfqr,
maxfqr=maxfqr,
voicing=voicing,
minpeaksalience=minpeaksalience)
melody = collect(wav_data, fs, 'mtg-melodia:melodia', parameters=params)
timestep = melody['vector'][0].to_float()
pitch = melody['vector'][1]
# 采样时间的起点是第8个timestep
starttime = timestep * 8
# 列表：(时间，频率)
result = []
for i, p in enumerate(pitch):
result.append((starttime + i * timestep, p))
return result
def get_hpcp(x, sr, n_bins=12, f_min=55, f_ref=440.0, min_magn=-100):
"""Compute HPCP features from raw audio using the HPCP Vamp plugin.
Vamp, vamp python module and plug-in must be installed.
Args:
x (1d-array): audio signal, mono
sr (int): sample rate
n_bins (int): number of chroma bins
f_min (float): minimum frequency
f_ref (float): A4 tuning frequency
min_magn (float): minimum magnitude for peak detection, in dB
Returns:
1d-array: time vector
2d-array: HPCP features
"""
plugin = 'vamp-hpcp-mtg:MTG-HPCP'
params = {'LF': f_min, 'nbins': n_bins, 'reff0': f_ref,
'peakMagThreshold': min_magn}
data = vamp.collect(x, sr, plugin, parameters=params)
vamp_hop, hpcp = data['matrix']
t = float(vamp_hop) * (8 + np.arange(len(hpcp)))
return t, hpcp
def chordBeats(infile, outfile):
print 'Loading audio file...', infile
audio = essentia.standard.MonoLoader(filename=infile)()
bt = BeatTrackerMultiFeature()
beats, _ = bt(audio)
#beats = beats[::4]
parameters = {}
stepsize, chroma = vamp.collect(audio,
44100,
"nnls-chroma:nnls-chroma",
output="chroma",
step_size=2048)["matrix"]
# to essentia convention
#chroma = np.roll(chroma, 3, 1)
chords = ChordsDetectionBeats(hopSize=2048)
syms, strengths = chords(chroma, beats)
segments = essentia_chord_utils.toMirexLab(0.0,
len(audio) / 44100.0, beats,
syms, strengths)
with open(outfile, 'w') as content_file:
for s in segments:
content_file.write(str(s) + '\n')
def extract_melody(self):
"""
Extracts melody from the audio using the melodia vamp plugin. Uses arguments kept
in self:
- `self.minimum_frequency` (default: 55 Hz)
- `self.maximum_frequency` (default: 1760 Hz)
- `self.voicing_tolerance` (default: 0.2)
- `self.minimum_peak_salience` (default: 0.0)
This function sets two class members used in other parts:
- `self.melody`: (numpy array) contains the melody in Hz for every timestep
(0 indicates no voice).
- `self.timestamps`: (numpy array) contains the timestamps for each melody note
"""
params = {
'minfqr': self.minimum_frequency,
'maxfqr': self.maximum_frequency,
'voicing': self.voicing_tolerance,
'minpeaksalience': self.minimum_peak_salience
}
data = vamp.collect(self.audio_signal.audio_data,
self.sample_rate,
"mtg-melodia:melodia",
parameters=params)
_, melody = data['vector']
hop = 128. / 44100. # hard coded hop in Melodia vamp plugin, converting it to frames.
timestamps = 8 * hop + np.arange(len(melody)) * hop
melody[melody < 0] = 0
self.melody = melody
self.timestamps = timestamps
def extract_melody(self):
params = {}
params['minfqr'] = self.minimum_frequency
params['maxfqr'] = self.maximum_frequency
params['voicing'] = self.voicing_tolerance
params['minpeaksalience'] = self.minimum_peak_salience
try:
data = vamp.collect(self.audio_signal.audio_data,
self.sample_rate,
"mtg-melodia:melodia",
parameters=params)
except Exception as e:
print(
'**~*~*~*~*~*~*~*~*~*~*~*~*~*~*~*~*~*~*~*~*~*~*~*~*~*~*~*~*~*~**\n'
'* Are Vamp and Melodia installed correctly? *\n'
'* Check https://bit.ly/2DXbrAk for installation instructions! *\n'
'**~*~*~*~*~*~*~*~*~*~*~*~*~*~*~*~*~*~*~*~*~*~*~*~*~*~*~*~*~*~**'
)
raise e
_, melody = data['vector']
hop = 128. / 44100. # hard coded hop in Melodia vamp plugin, converting it to frames.
timestamps = 8 * hop + np.arange(len(melody)) * hop
melody[melody < 0] = 0
self.melody = melody
self.timestamps = timestamps
def extract(self, signal):
"""
Extract Melody of the sound x in numpy array format.
"""
start_time = time.time()
if signal.ndim > 1:
print "INFO: Input signal has more than 1 channel; the channels will be averaged."
signal = np.mean(signal, axis=1)
# Remove percussion if separate:
if self.separate:
audio, percussion = librosa.effects.hpss(signal)
#Vamp UPF melody extractor plugin
data = vamp.collect(audio,
self.fs,
"mtg-melodia:melodia",
parameters=self.params)
hop, melody = data['vector']
timestamps = 8 * 128 / self.fs + np.arange(
len(melody)) * (128 / self.fs)
# Melodia returns unvoiced (=no melody) sections as negative values.we will put them to 0:
melody_pos = melody[:]
melody_pos[melody <= 0] = 0
# Finally, you might want to plot the pitch sequence in cents rather than in Hz.
# This especially makes sense if you are comparing two or more pitch sequences
# to each other (e.g. comparing an estimate against a reference).
melody_cents = 1200 * np.log2(melody / 55.0)
melody_cents[melody <= 0] = 0
print("--- %s seconds --- for melody extraction" %
(time.time() - start_time))
return melody_cents, timestamps
def get_hpcp(x, sr, n_bins=12, f_min=55, f_ref=440.0, min_magn=-100):
"""Compute HPCP features from raw audio using the HPCP Vamp plugin.
Vamp, vamp python module and plug-in must be installed.
Args:
x (1d-array): audio signal, mono
sr (int): sample rate
n_bins (int): number of chroma bins
f_min (float): minimum frequency
f_ref (float): A4 tuning frequency
min_magn (float): minimum magnitude for peak detection, in dB
Returns:
1d-array: time vector
2d-array: HPCP features
"""
plugin = 'vamp-hpcp-mtg:MTG-HPCP'
params = {
'LF': f_min,
'nbins': n_bins,
'reff0': f_ref,
'peakMagThreshold': min_magn
}
data = vamp.collect(x, sr, plugin, parameters=params)
vamp_hop, hpcp = data['matrix']
t = float(vamp_hop) * (8 + np.arange(len(hpcp)))
return t, hpcp
def rawChromaFromAudio(audiofile, sampleRate=44100, stepSize=2048):
mywindow = np.array([
0.001769, 0.015848, 0.043608, 0.084265, 0.136670, 0.199341, 0.270509,
0.348162, 0.430105, 0.514023, 0.597545, 0.678311, 0.754038, 0.822586,
0.882019, 0.930656, 0.967124, 0.990393, 0.999803, 0.999803, 0.999803,
0.999803, 0.999803, 0.999803, 0.999803, 0.999803, 0.999803, 0.999803,
0.999803, 0.999803, 0.999803, 0.999803, 0.999803, 0.999803, 0.999803,
0.999803, 0.999803, 0.999803, 0.999803, 0.999803, 0.999803, 0.999803,
0.999650, 0.996856, 0.991283, 0.982963, 0.971942, 0.958281, 0.942058,
0.923362, 0.902299, 0.878986, 0.853553, 0.826144, 0.796910, 0.766016,
0.733634, 0.699946, 0.665140, 0.629410, 0.592956, 0.555982, 0.518696,
0.481304, 0.444018, 0.407044, 0.370590, 0.334860, 0.300054, 0.266366,
0.233984, 0.203090, 0.173856, 0.146447, 0.121014, 0.097701, 0.076638,
0.057942, 0.041719, 0.028058, 0.017037, 0.008717, 0.003144, 0.000350
])
audio = essentia.standard.MonoLoader(filename=audiofile,
sampleRate=sampleRate)()
# estimate audio duration just for caching purposes:
audioDuration(audiofile, sampleRate=sampleRate, audioSamples=audio)
stepsize, semitones = vamp.collect(audio,
sampleRate,
"nnls-chroma:nnls-chroma",
output="semitonespectrum",
step_size=stepSize)["matrix"]
chroma = np.zeros((semitones.shape[0], 12))
for i in range(semitones.shape[0]):
tones = semitones[i] * mywindow
cc = chroma[i]
for j in range(tones.size):
cc[j % 12] = cc[j % 12] + tones[j]
# roll from 'A' based to 'C' based
chroma = np.roll(chroma, shift=-3, axis=1)
return chroma
def get_melody(x, sr, f_min=55, f_max=1760, min_salience=0.0, unvoiced=True):
"""Extract main melody from raw audio using the Melodia Vamp plugin.
Vamp, vamp python module and plug-in must be installed.
Args:
x (np.array): audio signal, mono
sr (int): sample rate
f_min (float): minimum frequency
f_max (float): maximum frequency
Return:
1d-array: time vector
1d-array: main melody (in cents)
"""
plugin = 'mtg-melodia:melodia'
params = {'minfqr': f_min, 'maxfqr': f_max,
'minpeaksalience': min_salience}
data = vamp.collect(x, sr, plugin, parameters=params)
vamp_hop, f0 = data['vector']
if unvoiced:
f0 = abs(f0)
f0[f0 == 0] = None
else:
f0[f0 <= 0] = None
hz2midi = lambda f: 69 + 12 * np.log2(abs(f) / 440)
melody = hz2midi(f0)
melody = melody[:, np.newaxis]
t = float(vamp_hop) * (8 + np.arange(len(melody)))
return t, melody
def melodia(self):
logging.debug("Generating melody")
df = self.df
melodia = vamp.collect(self.y, self.sr, "mtg-melodia:melodia")
hop, melody = melodia['vector']
timestamps = 8 * 128 / self.sr + np.arange(
len(melody)) * (128 / self.sr)
timestamps /= 2
melody_pos = melody[:]
melody_pos[melody <= 0] = None
df_melodia = pd.DataFrame(np.vstack([timestamps, melody]).T)
#df_melodia.to_csv('/dataset/test/melodia.csv', index=False)
df_melodia[1] = df_melodia[1].fillna(-1)
#df_melodia = df_melodia[df_melodia[1]>0]
#df_melodia[2] = librosa.hz_to_note(df_melodia[1], octave=False)
#df_melodia[[0, 0, 2]].to_csv('%s.melody_note_all.txt' % self.file, sep='\t', header=False, index=False)
tmp_df = pd.merge_asof(df_melodia,
self.df_downbeats,
left_on=0,
right_on='beat_time')
tmp_df = tmp_df[(tmp_df.beat_time > 0) & (tmp_df[1] > 0)]
tmp_df['melody_note'] = tmp_df[1].map(self.write_melody_note)
tmp_df['melody_note_ori'] = librosa.hz_to_note(tmp_df[1], octave=False)
tmp_df = tmp_df.groupby([
'beat_time'
]).agg(lambda x: scipy.stats.mode(x)[0]).reset_index(drop=False)
tmp_df = tmp_df[['beat_time', 'melody_note', 'melody_note_ori']]
self.df_melodia = tmp_df
def extract_notes_pYIN_vamp(x,
Fs=config.fs,
H=config.hopsize,
N=config.framesize):
# pYIN parameters
param = {'threshdistr': 2, 'outputunvoiced': 2, 'precisetime': 0}
# Options: smoothedpitchtrack, f0candidates, f0probs, voicedprob, candidatesalience, smoothedpitchtrack, notes
pYIN_note_output = vamp.collect(x,
Fs,
"pyin:pyin",
output='notes',
parameters=param,
step_size=H,
block_size=N)
# reformating
traj = np.empty((0, 3))
for j, entry in enumerate(pYIN_note_output['list']):
timestamp = float(entry['timestamp'])
duration = float(entry['duration'])
note = float(entry['values'][0])
if j == 0:
traj = np.vstack((traj, [0, timestamp, 0]))
elif timestamp != traj[-1][1]:
traj = np.vstack((traj, [traj[-1][1], timestamp, 0]))
traj = np.vstack((traj, [timestamp, timestamp + duration, note]))
return traj
def audio_to_midi_melodia(infile,
outfile,
bpm,
smooth=0.25,
minduration=0.1,
savejams=False):
fs = 44100
hop = 128
data, sr = librosa.load(infile, sr=fs, mono=True)
melody = vamp.collect(data,
sr,
"mtg-melodia:melodia",
parameters={"voicing": 0.2})
pitch = melody['vector'][1]
pitch = np.insert(pitch, 0, [0] * 8)
midi_pitch = hz2midi(pitch)
notes = midi_to_notes(midi_pitch, fs, hop, smooth, minduration)
save_midi(outfile, notes, bpm)
if savejams:
jamsfile = outfile.replace(".mid", ".jams")
track_duration = len(data) / float(fs)
save_jams(jamsfile, notes, track_duration, os.path.basename(infile))
print("Conversion complete.")
def calculateBeatsAndSemitones(infile):
print 'Loading audio file...', infile
proc = BeatTrackingProcessor(fps=100,
method='comb',
min_bpm=40,
max_bpm=240,
act_smooth=0.09,
hist_smooth=7,
alpha=0.79)
act = RNNBeatProcessor()(infile)
beats = proc(act).astype('float32')
audio = essentia.standard.MonoLoader(filename=infile)()
#bt = BeatTrackerMultiFeature()
#beats, _ = bt(audio)
# TODO: best partameters.
parameters = {}
#stepsize, chroma = vamp.collect(
# audio, 44100, "nnls-chroma:nnls-chroma", output = "chroma", step_size=2048)["matrix"]
stepsize, semitones = vamp.collect(audio,
44100,
"nnls-chroma:nnls-chroma",
output="semitonespectrum",
step_size=2048)["matrix"]
return len(audio), beats, semitones
def extract_melody(filename_wav):
data, rate = librosa.load(
os.path.join(utils.get_dirname_audio_warped(), filename_wav))
data_melody = vamp.collect(data, rate, "mtg-melodia:melodia")
return data_melody
def extractRawMelody(filename, position, around):
data, rate = librosa.load(filename, offset = position - around,\
duration = 2 * around)
result = vamp.collect(data, rate, "mtg-melodia:melodia");
frameDur, melody = result['vector']
#frameDur = frameDur.to_frame(rate)
return melody, rate
def add_recommend_database(self, url, json_path, audio_path):
"""
' 목적 : 음원의 유튜브 url으로부터 추출한, 해당 음원의 코드 정보와 url을 입력받은 json_path에 저장하는 함수 (코드 정보는 vamp 플러그인으로부터 추출)
' 리턴값 : 없음
"""
# 사용자가 앱을 수행하다 중단했을시, 제거되지 않고 남아있는 파일을 제거하여 youtube_dl 오류방지
files = glob.glob(audio_path + "/*.mp4")
for x in files:
if not os.path.isdir(x):
os.remove(x)
ydl_opts = {
'format': 'best',
'outtmpl': audio_path + '/%(title)s.%(ext)s',
}
with youtube_dl.YoutubeDL(ydl_opts) as ydl:
ydl.download([url])
files = glob.glob(audio_path + "/*.mp4")
audio_file_name = ''
signal, sr = 0, 0
for x in files:
if not os.path.isdir(x):
filename = os.path.splitext(x)
audio_file_name = filename[0].split(audio_path + '/')[1]
signal, sr = librosa.load(x)
os.remove(x)
# load chord data from vamp plugin
predicted_chord_list = vamp.collect(signal, sr, "nnls-chroma:chordino")
# -----------------------추출한 데이터 형식을 다루기 쉽게 가공----------------------- #
chord_list_with_timestamp = predicted_chord_list["list"]
chord_list = []
for i in chord_list_with_timestamp:
chord_list.append(i['label'])
double_chord_list_with_url = []
for idx in range(len(chord_list) - 1):
if chord_list[idx] != 'N' and chord_list[idx + 1] != 'N':
double_chord = chord_list[idx] + '-' + chord_list[idx + 1]
double_chord_list_with_url.append(double_chord)
double_chord_list_with_url.append(
double_chord_list_with_url[-1].split('-')[-1] + '-' + url)
# ---------------------------------------------------------------------------- #
with open(json_path, 'r', encoding='utf-8') as f:
chord_list_dict = json.load(f)
dict_key_music_name = audio_file_name
if dict_key_music_name not in chord_list_dict:
chord_list_dict[dict_key_music_name] = double_chord_list_with_url
with open(json_path, "w", encoding='utf-8') as fp:
json.dump(chord_list_dict, fp, indent=4, ensure_ascii=False)
def extract_melody_melodia(audio_path):
voicing = .6
# Comments in this function are given by the creator of melodia
# This is how we load audio using Librosa
audio_1, sr_1 = librosa.load(audio_path, sr=44100, mono=True)
# data_1 = vamp.collect(audio_1, sr_1, "mtg-melodia:melodia")
# vector is a tuple of two values: the hop size used for analysis and the array of pitch values
# Note that the hop size is *always* equal to 128/44100.0 = 2.9 ms
# hop_1, melody_1 = data_1['vector']
# parameter values are specified by providing a dicionary to the optional "parameters" parameter:
params = {
"minfqr": 100.0,
"maxfqr": 1760.0,
"voicing": voicing,
"minpeaksalience": 0.0
}
data_1 = vamp.collect(audio_1,
sr_1,
"mtg-melodia:melodia",
parameters=params)
hop_1, melody_1 = data_1['vector']
# <h3>\*\*\* SUPER IMPORTANT SUPER IMPORTANT \*\*\*</h3>
# For reasons internal to the vamp architecture, THE TIMESTAMP OF THE FIRST VALUE IN THE MELODY ARRAY IS ALWAYS:
#
# ```
# first_timestamp = 8 * hop = 8 * 128/44100.0 = 0.023219954648526078
# ```
#
# This means that the timestamp of the pitch value at index i (starting with i=0) is given by:
#
# ```
# timestamp[i] = 8 * 128/44100.0 + i * (128/44100.0)
# ```
#
# So, if you want to generate a timestamp array to match the pitch values, you do it like this:
timestamps_1 = 8 * 128 / 44100.0 + np.arange(
len(melody_1)) * (128 / 44100.0)
# Melodia has 4 parameters:
# * **minfqr**: minimum frequency in Hertz (default 55.0)
# * **maxfqr**: maximum frequency in Hertz (default 1760.0)
# * **voicing**: voicing tolerance. Greater values will result in more pitch contours included in the final melody.
# Smaller values will result in less pitch contours included in the final melody (default 0.2).
# * **minpeaksalience**: (in Sonic Visualiser "Monophonic Noise Filter") is a hack to avoid silence turning into
# junk contours when analyzing monophonic recordings (e.g. solo voice with no accompaniment).
# Generally you want to leave this untouched (default 0.0).
melody = melody_1.tolist()
output = {'data': [{'value': melody, 'time': timestamps_1.tolist()}]}
return output
def generate_melody(audio_file_path):
# This is the audio file we'll be analyzing.
# You can download it here: http://labrosa.ee.columbia.edu/projects/melody/mirex05TrainFiles.zip
audio_file = audio_file_path
# This is how we load audio using Librosa
audio, sr = librosa.load(audio_file,
offset=30.0,
duration=50.0,
sr=44100,
mono=True)
# Exracting the melody using Melodia with default parameter values
data = vamp.collect(audio, sr, "mtg-melodia:melodia")
# print(data)
# vector is a tuple of two values: the hop size used for analysis and the array of pitch values
# Note that the hop size is *always* equal to 128/44100.0 = 2.9 ms
hop, melody = data['vector']
# print(hop)
# print(melody)
# timestamps = 8 * 128/44100.0 + np.arange(len(melody)) * (128/44100.0)
# Extracting the melody using Melodia with custom parameter values
# parameter values are specified by providing a dicionary to the optional "parameters" parameter:
# params = {"minfqr": 100.0, "maxfqr": 800.0, "voicing": 0.2, "minpeaksalience": 0.0}
# data = vamp.collect(audio, sr, "mtg-melodia:melodia", parameters=params)
# hop, melody = data['vector']
# Melodia returns unvoiced (=no melody) sections as negative values. So by default, we get:
# plt.figure(figsize=(18,6))
# plt.plot(timestamps, melody)
# plt.xlabel('Time (s)')
# plt.ylabel('Frequency (Hz)')
# plt.show()
# A clearer option is to get rid of the negative values before plotting
# melody_pos = melody[:]
# melody_pos[melody<=0] = None
# plt.figure(figsize=(18,6))
# plt.plot(timestamps, melody_pos)
# plt.xlabel('Time (s)')
# plt.ylabel('Frequency (Hz)')
# plt.show()
# Finally, you might want to plot the pitch sequence in cents rather than in Hz.
# This especially makes sense if you are comparing two or more pitch sequences
# to each other (e.g. comparing an estimate against a reference).
# melody_cents = 1200*np.log2(melody/55.0)
# melody_cents[melody<=0] = None
# plt.figure(figsize=(18,6))
# plt.plot(timestamps, melody_cents)
# plt.xlabel('Time (s)')
# plt.ylabel('Frequency (cents relative to 55 Hz)')
# plt.show()
return melody
def audio_to_midi_melodia(infile,
outfile,
bpm,
smooth=0.25,
minduration=0.1,
savejams=False):
# define analysis parameters
fs = 44100
hop = 128
# load audio using librosa
print("Loading audio...")
data, sr = librosa.load(infile, sr=fs, mono=True)
# extract melody using melodia vamp plugin
print("Extracting melody f0 with MELODIA...")
melody = vamp.collect(data,
sr,
"mtg-melodia:melodia",
parameters={"voicing": 0.2})
# hop = melody['vector'][0]
pitch = melody['vector'][1]
# impute missing 0's to compensate for starting timestamp
pitch = np.insert(pitch, 0, [0] * 8)
# debug
# np.asarray(pitch).dump('f0.npy')
# print(len(pitch))
# convert f0 to midi notes
print("Converting Hz to MIDI notes...")
midi_pitch = hz2midi(pitch)
# segment sequence into individual midi notes
notes = midi_to_notes(midi_pitch, fs, hop, smooth, minduration)
# save note sequence to a midi file
print("Saving MIDI to disk...")
save_midi(outfile, notes, bpm)
'''import matplotlib.pyplot as plt
notes = np.asarray(notes)
x = notes[:,0];
y = np.fmod(notes[:,2],12);
plt.plot(x, y)
#plt.xticks(np.arange(0,lim,10))
#plt.yticks(np.arange(0,12))
#plt.grid()
plt.show()'''
if savejams:
print("Saving JAMS to disk...")
jamsfile = outfile.replace(".mid", ".jams")
track_duration = len(data) / float(fs)
save_jams(jamsfile, notes, track_duration, os.path.basename(infile))
print("Conversion complete.")
def test_collect_fixed_sample_rate_2():
buf = input_data(blocksize * 10)
rdict = vamp.collect(buf, rate, plugin_key, "curve-fsr-timed")
step, results = rdict["vector"]
assert abs(float(step) - 0.4) < eps
assert len(results) == 10
for i in range(len(results)):
assert abs(results[i] - i * 0.1) < eps
def test_collect_fixed_sample_rate_2():
buf = input_data(blocksize * 10)
rdict = vamp.collect(buf, rate, plugin_key, "curve-fsr-timed")
step, results = rdict["vector"]
assert abs(float(step) - 0.4) < eps
assert len(results) == 10
for i in range(len(results)):
assert abs(results[i] - i * 0.1) < eps
def get_bars_and_beats_qm(self): import vamp if self.metre_qm is None: barbeattrack_output = vamp.collect(self.signal_mono, self.fs, 'qm-vamp-plugins:qm-barbeattracker') beat_times = np.array([float(item['timestamp']) for item in barbeattrack_output['list']]) beat_metre = np.array([int(item['label']) for item in barbeattrack_output['list']]) standard_beat_times_metre = np.array(zip(*(beat_times,beat_metre))) self.metre_qm = standard_beat_times_metre
def test_process_summary_param_kwargs_1():
buf = input_data(blocksize * 10)
rdict = vamp.collect(
plugin_key=plugin_key, output="input-summary", parameters={"produce_output": False}, data=buf, sample_rate=rate
)
assert "vector" in rdict
step, results = rdict["vector"]
assert len(results) == 0
def convertToMIDI(infile, outfile, bpm=120, smooth=0, minduration=0.1,
savejams=False):
# define analysis parameters
fs = 44100
hop = 128
# load audio using librosa
print("Loading audio...")
data, sr = soundfile.read(infile)
# mixdown to mono if needed
if len(data.shape) > 1 and data.shape[1] > 1:
data = data.mean(axis=1)
# resample to 44100 if needed
if sr != fs:
data = resampy.resample(data, sr, fs)
sr = fs
# extract melody using melodia vamp plugin
print("Extracting melody f0 with MELODIA...")
melody = vamp.collect(data, sr, "mtg-melodia:melodia",
parameters={"voicing": 10})
# hop = melody['vector'][0]
pitch = melody['vector'][1]
# impute missing 0's to compensate for starting timestamp
pitch = np.insert(pitch, 0, [0]*8)
'''print(pitch)
stored = 0.0
for i in pitch:
if i != stored:
print(i, end = ' ')
stored = i'''
# debug
# np.asarray(pitch).dump('f0.npy')
# print(len(pitch))
# convert f0 to midi notes
print("Converting Hz to MIDI notes...")
midi_pitch = hz2midi(pitch)
stored = 0.0
for i in midi_pitch:
if i != stored:
stored = i
# segment sequence into individual midi notes
notes = midi_to_notes(midi_pitch, fs, hop, smooth, minduration)
# save note sequence to a midi file
print("Saving MIDI to disk...")
save_midi(outfile, notes, bpm)
print("Conversion complete.")
#convertToMIDI('../testSongs/ksg.wav', '../testSongs/ksg.mid', 75, smooth=0, minduration=0.06)
def test_process_summary_param_kwargs_2():
buf = input_data(blocksize * 10)
rdict = vamp.collect(plugin_key=plugin_key,
output="input-summary",
data=buf,
sample_rate=rate)
assert ("vector" in rdict)
step, results = rdict["vector"]
assert len(results) > 0
def rock(audio):
jojo = vamp.collect(audio,
config.fs,
"pyin:pyin",
step_size=config.hopsize,
output="notes")
import pdb
pdb.set_trace()
def test_collect_grid_one_sample_per_step():
buf = input_data(blocksize * 10)
rdict = vamp.collect(buf, rate, plugin_key, "grid-oss")
step, results = rdict["matrix"]
assert abs(float(step) - (1024.0 / rate)) < eps
assert len(results) == 10
for i in range(len(results)):
expected = np.array([(j + i + 2.0) / 30.0 for j in range(0, 10)])
assert (abs(results[i] - expected) < eps).all()
def test_collect_grid_one_sample_per_step():
buf = input_data(blocksize * 10)
rdict = vamp.collect(buf, rate, plugin_key, "grid-oss")
step, results = rdict["matrix"]
assert abs(float(step) - (1024.0 / rate)) < eps
assert len(results) == 10
for i in range(len(results)):
expected = np.array([(j + i + 2.0) / 30.0 for j in range(0, 10)])
assert (abs(results[i] - expected) < eps).all()
def audio_to_midi_melodia(infile,
outfile,
bpm,
smooth=0.25,
minduration=0.1,
savejams=False):
# define analysis parameters
fs = 44100
hop = 128
# load audio using librosa
print("Loading audio...")
data, sr = soundfile.read(infile)
# mixdown to mono if needed
if len(data.shape) > 1 and data.shape[1] > 1:
data = data.mean(axis=1)
# resample to 44100 if needed
if sr != fs:
data = resampy.resample(data, sr, fs)
sr = fs
# extract melody using melodia vamp plugin
print("Extracting melody f0 with MELODIA...")
melody = vamp.collect(data,
sr,
"mtg-melodia:melodia",
parameters={"voicing": 0.2})
# hop = melody['vector'][0]
pitch = melody['vector'][1]
# impute missing 0's to compensate for starting timestamp
pitch = np.insert(pitch, 0, [0] * 8)
# debug
# np.asarray(pitch).dump('f0.npy')
# print(len(pitch))
# convert f0 to midi notes
print("Converting Hz to MIDI notes...")
midi_pitch = hz2midi(pitch)
# segment sequence into individual midi notes
notes = midi_to_notes(midi_pitch, fs, hop, smooth, minduration)
# save note sequence to a midi file
print("Saving MIDI to disk...")
save_midi(outfile, notes, bpm)
if savejams:
print("Saving JAMS to disk...")
jamsfile = os.path.splitext(outfile)[0] + ".jams"
track_duration = len(data) / float(fs)
save_jams(jamsfile, notes, track_duration, os.path.basename(infile))
print("Conversion complete.")
def test_collect_variable_sample_rate():
buf = input_data(blocksize * 10)
rdict = vamp.collect(buf, rate, plugin_key, "curve-vsr")
results = rdict["list"]
assert len(results) == 10
i = 0
for r in results:
assert r["timestamp"] == vamp.vampyhost.RealTime("seconds", i * 0.75)
assert abs(r["values"][0] - i * 0.1) < eps
i = i + 1
def main():
fpath = raw_input('Enter full path to audio file: ')
data, rate = librosa.load(fpath)
plugin_choice = choose_vamp_plugin()
# this script is assuming that university of alicante polyphonic transcription is being used
num_voices = int(raw_input('Enter desired number of voices: '))
plugin_output = vamp.collect(data, rate, plugin_choice, parameters={'maxpolyphony':num_voices})
note_start_lst, dur_lst, note_lst = parse_note_transcription_output(plugin_output)
print 'note_start_lst length: %d, dur_lst length: %d, note_lst length: %d.' % (len(note_start_lst), len(dur_lst), len(note_lst))
note_df = make_note_dataframe(note_start_lst, dur_lst, note_lst)
print note_df
make_reordered_wav_file(data, rate, note_df)
def test_process_summary_param_kwargs_3():
buf = input_data(blocksize * 10)
rdict = vamp.collect(
plugin_key=plugin_key,
output="input-summary",
data=buf,
sample_rate=rate,
process_timestamp_method=vamp.vampyhost.SHIFT_DATA,
)
assert "vector" in rdict
step, results = rdict["vector"]
assert len(results) > 0
def test_collect_one_sample_per_step():
buf = input_data(blocksize * 10)
rdict = vamp.collect(buf, rate, plugin_key, "input-timestamp")
step, results = rdict["vector"]
assert abs(float(step) - (1024.0 / rate)) < eps
assert len(results) == 10
for i in range(len(results)):
# The timestamp should be the frame number of the first frame in the
# input buffer
expected = i * blocksize
actual = results[i]
assert actual == expected
def test_process_summary_param_kwargs_fail():
buf = input_data(blocksize * 10)
try:
rdict = vamp.collect(
plugin_key=plugin_key,
output="input-summary",
data=buf,
sample_rate=rate,
process_timestamp_method=vamp.vampyhost.SHIFT_DATA,
unknown_argument=1,
)
except Exception: # unknown kwarg
pass
def transcribe(self, audio_path):
if not os.path.exists(audio_path):
raise ValueError('Invalid audio path')
x, fs = load(audio_path, mono=True)
notes = vamp.collect(x, fs, "qm-vamp-plugins:qm-transcription", output="transcription")['list']
# access attributes of a note event by:
# ts: f.timestamp
# duration: f.duration
# MIDI notes: f.values
return notes
def audio_to_midi_melodia(infile, outfile, bpm, smooth=0.25, minduration=0.1,
savejams=False):
# define analysis parameters
fs = 44100
hop = 128
# load audio using librosa
print("Loading audio...")
data, sr = librosa.load(infile, sr=fs, mono=True)
# extract melody using melodia vamp plugin
print("Extracting melody f0 with MELODIA...")
melody = vamp.collect(data, sr, "mtg-melodia:melodia",
parameters={"voicing": 0.2})
# hop = melody['vector'][0]
pitch = melody['vector'][1]
# impute missing 0's to compensate for starting timestamp
pitch = np.insert(pitch, 0, [0]*8)
# debug
# np.asarray(pitch).dump('f0.npy')
# print(len(pitch))
# convert f0 to midi notes
print("Converting Hz to MIDI notes...")
midi_pitch = hz2midi(pitch)
# segment sequence into individual midi notes
notes = midi_to_notes(midi_pitch, fs, hop, smooth, minduration)
# save note sequence to a midi file
print("Saving MIDI to disk...")
save_midi(outfile, notes, bpm)
if savejams:
print("Saving JAMS to disk...")
jamsfile = outfile.replace(".mid", ".jams")
track_duration = len(data) / float(fs)
save_jams(jamsfile, notes, track_duration, os.path.basename(infile))
print("Conversion complete.")
def transcription_cal(audio_filename, monoNoteOut_filename):
sr = 44100
loader = es.MonoLoader(filename=audio_filename, downmix = 'mix', sampleRate = sr)
audio = loader()
# original pyin
# data = vamp.collect(audio, sr, "pyin:pyin", output = 'notes')
# modified pyin
data = vamp.collect(audio, sr, "pyinbobigram:pyin-jingju", output = 'notes')
# for note in data['list']:
# print note['duration'], note['timestamp'], note['values'][0]
with open(monoNoteOut_filename, 'w') as csvfile:
csv_writer = csv.writer(csvfile, delimiter=',')
for note in data['list']:
csv_writer.writerow([note['timestamp'], note['values'][0], note['duration']])
def get_pc_data_aubio(y,fs):
""" Use SLIVET note transcription method to get
melodic sequence.
"""
data = vamp.collect(y,fs,'vamp-aubio:aubionotes')
freqs = [d['values'][0] for d in data['list']]
midi = librosa.core.hz_to_midi(freqs)
midi = [int(m) for m in np.round(midi)]
start_t = [d['timestamp'].to_float() for d in data['list']]
end_t = start_t[1:]
end_t.append(y.shape[0]/float(fs))
# to pitch class representation
labels = ['pc' + str(coreutils.midi_note_to_pc(n)) for n in midi]
melody_sequence = coredata.Sequence(labels=labels,\
start_times=start_t,end_times=end_t)
return melody_sequence
def get_midi_data_slivet(y,fs):
""" Use SLIVET note transcription method to get
melodic sequence.
"""
data = vamp.collect(y,fs,'silvet:silvet')
labels = []
start_t = []
end_t = []
for d in data['list']:
n = librosa.note_to_midi(d['label'])
# pc = coreutils.midi_note_to_pc(n)
# label = 'pc' + str(pc)
st = d['timestamp'].to_float()
# et = st + d['duration'].to_float()
# labels.append(label)
labels.append(n)
start_t.append(st)
# end_t.append(et)
end_t = start_t[1:]
# for some reason the last duration gets
# screwed up
# st = start_t[-1]
# et = st + (float(y.shape[0])/float(fs))
delta_t = start_t[-1] - start_t[-2]
et = end_t[-1] + delta_t
end_t.append(et)
# to pitch class representation
labels = ['pc' + str(coreutils.midi_note_to_pc(n)) for n in labels]
melody_sequence = coredata.Sequence(labels=labels,\
start_times=start_t,end_times=end_t)
return melody_sequence
def test_collect_runs_at_all():
buf = input_data(blocksize * 10)
rdict = vamp.collect(buf, rate, plugin_key, "input-timestamp")
step, results = rdict["vector"]
assert results != []
#fragment = '../traditional_dataset/allemande/fragments/allemande_first_fragment_nicolet'
#fragment = '../traditional_dataset/allemande/fragments/allemande_third_fragment_rampal'
#fragment = '../traditional_dataset/allemande/fragments/allemande_fourth_fragment_larrieu'
#fragment = '../traditional_dataset/allemande/fragments/allemande_fifth_fragment_preston'
fragment = '../traditional_dataset/sequenza/fragments/sequenza_first_fragment_robison'
audio_file = fragment + '_mono.wav'
gt_file = fragment + '.csv'
audio, sr = librosa.load(audio_file, sr=44100, mono=True)
#%%
# parameter values are specified by providing a dicionary:
params = {"minfqr": 100.0, "maxfqr": 2350.0, "voicing": 0.9, "minpeaksalience": 0.0}
data = vamp.collect(audio, sr, "mtg-melodia:melodia", parameters=params)
hop, melody_melodia = data['vector']
#melody_librosa, magnitudes = librosa.piptrack(audio, sr=sr, hop_length=128)
#print(hop)
#print(melody)
import numpy as np
timestamps = 8 * 128/44100.0 + np.arange(len(melody_melodia)) * (128/44100.0)
melody_hz = copy.deepcopy(melody_melodia)
melody_hz[melody_melodia<=0] = None
#%%
import melosynth as ms
ms.melosynth_pitch(melody_melodia, 'melosynth.wav', fs=44100, nHarmonics=1, square=True, useneg=False)
