def __test(y, sr, S, n_fft, hop_length, fmin, fmax, threshold):
pitches, mags = librosa.piptrack(y=y,
sr=sr,
S=S,
n_fft=n_fft,
hop_length=hop_length,
fmin=fmin,
fmax=fmax,
threshold=threshold)
def get_librosa_pitch(signal, fs, window):
pitches, magnitudes = librosa.piptrack(y=signal.astype('float'), sr=fs, n_fft=int(window),
hop_length=int(window/10))
pitch_pos = np.argmax(magnitudes, axis=0)
pitches_final = []
for i in range(len(pitch_pos)):
pitches_final.append(pitches[pitch_pos[i], i])
pitches_final = np.array(pitches_final)
pitches_final[pitches_final > 2000] = 0 # cut high pitches
return mf(pitches_final, 3) # use medfilt for smoothing
def get_pitch(filename= read_and_write_audio()):
x, sr_ = librosa.load(filename)
pitches, mags = librosa.piptrack(y=x,sr=sr_)
index = mags[:,128].argmax()
pitch = pitches[index, 128]
note_name = librosa.hz_to_note(pitch)
if len(note_name) > 2:
return note_name[0:2]
return note_name[0]
def get_pitch(filename):
x, sr_ = librosa.load(filename)
pitches, mags = librosa.piptrack(y=x, sr=sr_)
index = mags[:, 128].argmax()
pitch = pitches[index, 128]
note_name = librosa.hz_to_note(pitch)
if len(note_name) > 2:
return f'{note_name[0]}#'
return note_name[0]
def __test(S, freq):
pitches, mags = librosa.piptrack(S=S, fmin=100)
idx = (mags > 0)
assert len(idx) > 0
recovered_pitches = pitches[idx]
# We should be within one cent of the target
assert np.all(np.abs(np.log2(recovered_pitches) - np.log2(freq)) <= 1e-2)
def baseline_tracking(audio_file, result_file=None):
start = time.clock()
y, sr = librosa.load(audio_file, sr=database_sr)
print('Audio file loaded: ' + audio_file)
print('{:f}s for loading the audio file.'.format(time.clock() - start))
start = time.clock()
pitches, magnitudes = librosa.piptrack(y=y, sr=database_sr)
mag_thresh = 2 * np.mean(magnitudes) / 3
d_range, time_range = pitches.shape
pitches_max, magnitudes_max = \
extract_pitch_max(pitches, magnitudes, time_range)
ret_pitch = []
if result_file:
file = open(result_file, 'w+')
# print out tracked notes over time
for t in range(time_range):
# filter out frequencies with zero value
pitch_t = pitches_max[t]
# filter out notes that are too weak
mag_t = magnitudes_max[t]
if debug:
print("pitch_idx shape:")
print(pitch_idx.shape)
print("mag_idx shape:")
print(mag_idx.shape)
print("idx shape:")
print(merged_idx.shape)
# only print at a time t if there're notes present
# if pitch_t != 0 and mag_t > mag_thresh:
if True:
ret_pitch.append(pitch_t)
if result_file:
file.write(str(t) + ',' + str(pitch_t) + '\n')
elif debug:
print(t)
print(pitch_t)
print()
if result_file:
print("Saved in " + result_file)
file.close()
print("Stat:")
print('{:f}s for piptrack'.format(time.clock() - start))
print("len(pitches): {:d}".format(len(pitches)))
print("pitch shape:")
print(pitches.shape)
print("magnitudes shape:")
print(magnitudes.shape)
print("sampling rates:")
print(sr)
return pitches, ret_pitch
def __test(S, freq):
pitches, mags = librosa.piptrack(S=S, fmin=100)
idx = (mags > 0)
assert len(idx) > 0
recovered_pitches = pitches[idx]
# We should be within one cent of the target
assert np.all(np.abs(np.log2(recovered_pitches) - np.log2(freq)) <= 1e-2)
def onlyPitch(X, sample_rate):
stft = np.abs(librosa.stft(X))
pitches, magnitudes = librosa.piptrack(X,
sr=sample_rate,
S=stft,
fmin=70,
fmax=400)
pitch = []
for i in range(magnitudes.shape[1]):
index = magnitudes[:, i].argmax()
pitch.append(pitches[index, i])
return np.asarray(pitch)
def get_pitches_by_beats(y, starts, ends):
pitches = []
for b in range(len(starts)):
seg1 = y[np.int(starts[b] * global_sr):np.int(ends[b] * global_sr)]
pitches1, magnitudes1 = librosa.piptrack(y=seg1,
sr=global_sr,
fmin=20,
fmax=8000,
hop_length=2048)
max_p = np.median((extract_max(pitches1)))
pitches.append(max_p)
return np.array(pitches)
def load_music(filepath):
print("=========loac music %s=========" % (filepath))
# 节拍点的音高
beat_pitches = []
# 节拍点的八音12度的强度
beat_chroma = []
# 节拍点的八音12度最强度的索引
beat_chroma_max_index = []
y, sr = librosa.load(filepath, sr=None)
S = np.abs(librosa.stft(y))
# 起点强度（音符按键起始点）
onset_env = librosa.onset.onset_strength(y=y, sr=sr)
# 节拍点（帧索引）
tempo, beats = librosa.beat.beat_track(onset_envelope=onset_env, sr=sr)
# 节拍点 （时间点，单位秒）
beat_times = librosa.frames_to_time(beats, sr=sr)
# 频率、振幅强度，横坐标是采样点，纵坐标是帧
pitches, magnitudes = librosa.piptrack(S=S, sr=sr)
# 八音12度的强度
chroma_stft = librosa.feature.chroma_stft(S=S, sr=sr)
for beat in beats:
index = magnitudes[:, beat].argmax()
pitch = pitches[index, beat]
beat_pitches.append(float(pitch))
chromas = []
max_chroma = 0
max_chroma_index = 0
chr_st_index = 0
for chr_st in chroma_stft:
chromas.append(chr_st[beat])
if chr_st[beat] > max_chroma:
max_chroma = chr_st[beat]
max_chroma_index = chr_st_index
chr_st_index += 1
beat_chroma.append(chromas)
beat_chroma_max_index.append(max_chroma_index)
return {
"beat_times": beat_times.tolist(),
"beat_pitches": beat_pitches,
"beat_chroma_max_index": beat_chroma_max_index
}
def pitch_wave(path):
"""
sr, audio = wavfile.read(path)
time, frequency, confidence, activation = crepe.predict(audio, sr, viterbi=True)
print(time, frequency, confidence, activation)
return(time, frequency, confidence, activation)
"""
print("-*-*-*-*-pitch : wave.ver-*-*-*-*-")
y, sr = librosa.load(path)
pitches, magnitudes = librosa.piptrack(y=y, sr=sr)
plt.title("pitches")
plt.plot(pitches)
plt.show()
def bank(indata):
# segment cycle into quarter-notes.
# llen / 4 sized segments.
# try to find the pitch of each quarter note.
# save the audio to bank - marked with its pitch
l, _ = indata.shape
#q_size = int(l/4)
q_size = int(l / 8)
q = np.zeros((q_size, 2), dtype='float32')
count = 0
for x in range(1, 8):
q[:] = indata[count:count + q_size]
pitches, magnetude = librosa.piptrack(y=q[:, 1],
sr=SAMPLE_RATE,
fmin=250.0,
fmax=1050.0)
_, ts = magnetude.shape
print(ts)
print(pitches.shape)
index = max(magnetude[:, i].argmax() for i in range(ts))
prev = magnetude[:, 0].argmax()
best_ts = 0
for i in range(ts):
if magnetude[:, i].argmax() > prev:
best_ts = i
prev = magnetude[:, i].argmax()
pitch = pitches[index, best_ts]
if pitch != 0.0:
note_info = librosa.hz_to_note(pitch, cents=True)
# if your pretty close - >< 40cent its added to the notebank
print(note_info)
cents = int(note_info[0].replace('+', '-').split('-', 1)[-1])
print("Cents: ", cents)
if cents < 25:
note = librosa.hz_to_note(pitch, octave=False)
print("Bank: ", note)
keyId = note_names.index(note[0])
#if keyId in noteBank:
#noteBank[keyId].append(q)
#else:
#noteBank[keyId] = list(q)
noteBank[keyId] = q
count += q_size
def cal_pitch(audio_path='/data/pytong/wav/itg_0603/EE1811306983070-interview-iOS-2020-05-29-09-06-23.wav',
sample_rate=16000, hop=160):
signal, rate = librosa.load(audio_path, sr=sample_rate, mono=True)
pitches, magnitudes = librosa.piptrack(y=signal, sr=rate,
hop_length=hop)
pitch_result = []
pitches_reverted = list(zip(*pitches))
magnitudes_reverted = list(zip(*magnitudes))
for cur_pitch_list, cur_magnitude_list in zip(pitches_reverted, magnitudes_reverted):
index = cur_magnitude_list.index(max(cur_magnitude_list))
pitch = cur_pitch_list[index]
pitch_result.append(pitch)
print(len(pitch_result), pitch_result[0:2])
return pitch_result
def extract_features(self, filepath, filename, offset):
full_path = filepath + '/' + filename
if not os.path.isfile(full_path):
raise ExtractionError
samples, sample_rate = librosa.load(full_path,
offset=offset,
sr=None,
duration=1)
tempogram = librosa.feature.tempogram(y=samples, sr=sample_rate)
mfcc = librosa.feature.mfcc(y=samples, sr=sample_rate)
pitches, magnitudes = librosa.piptrack(y=samples, sr=sample_rate)
return np.concatenate([
tempogram.flatten()[:MAX_TEMPOGRAM_WEIGHT],
mfcc.flatten()[:MAX_MFCC_WEIGHT],
magnitudes.flatten()[:MAX_PITCH_WEIGHT]
])
def findpitch(file):
y, sr = librosa.load(file)
stft = np.abs(librosa.stft(y))
pitches, magnitudes = librosa.piptrack(y, sr=sr, S=stft, fmin=70, fmax=400)
pitch = []
for i in range(magnitudes.shape[1]):
index = magnitudes[:, 1].argmax()
pitch.append(pitches[index, i])
# pitch_tuning_offset = librosa.pitch_tuning(pitches)
pitch = [p for p in pitch if p > 0]
# pitchmean = np.mean(pitch)
# pitchstd = np.std(pitch)
# pitchmax = np.max(pitch)
pitchmin = np.min(pitch)
return pitchmin
def find_songKey(indata):
# find key of recent layer
global song_key
pitches, magnetude = librosa.piptrack(y=indata[:,1], sr=SAMPLE_RATE, fmin=250.0,fmax=1050.0)
time_slice = 1
index = magnetude[:, time_slice].argmax()
pitch = pitches[index, time_slice]
while pitch == 0.0:
time_slice = (time_slice + 10)%len(pitches)
index = magnetude[:, time_slice].argmax()
pitch = pitches[index, time_slice]
key = librosa.hz_to_note(pitch, octave=False)
print("findKey: ",key)
keyId = note_names.index(key[0])
song_key = keyId
def extract_features(filename):
# 파일 로드 부분
y, sr = librosa.load(filename)
# mel spectrogram을 fmax = 8000으로 계산 후 차원 고려하여 transpose
S = librosa.feature.melspectrogram(y=y, sr=sr, n_mels=128, fmax=8000)
S = np.transpose(S)
# piptrack 알고리즘
pitches, magnitudes = librosa.piptrack(y=y, sr=sr)
pitches = np.transpose(pitches)
magnitudes = np.transpose(magnitudes)
# 가장 유효한 값들을 나열한 일차원 sequence로 변환
pit_seq = pip_to_pitch(pitches)
mag_seq = pip_to_pitch(magnitudes)
# 발화 전체 중 silence의 비율 (delay_p)
delay_p, mag_index = find_delay(mag_seq, np.std(mag_seq))
# 유효 frame에 대해 pitch 의 mean 및 stdev
mean, error = effective_analysis(pit_seq, mag_index)
return delay_p, mean, error
def f0_estimate(f, S, fs, fmin=50, fmax=500, threshold=0.2):
'''
:param f: stft的频率值
:param S: stft频谱
:param fs: 采样率
:param fmin:
:param fmax:
:param threshold:
:return: f0 shape=(,t)
'''
y = np.abs(S)
pitches, mag = lib.piptrack(S=y,
sr=fs,
fmin=fmin,
fmax=fmax,
threshold=threshold)
f0 = f[np.argmax(mag, axis=0)]
return f0
def aaa():
p = pyaudio.PyAudio()
stream = p.open(format=FORMAT,
channels=CHANNELS,
rate=RATE,
input=True,
frames_per_buffer=CHUNK)
print("* recording")
frames = []
for i in range(0, int(RATE / CHUNK * RECORD_SECONDS)):
data = stream.read(CHUNK)
frames.append(data)
print("* done recording")
stream.stop_stream()
stream.close()
p.terminate()
wf = wave.open(WAVE_OUTPUT_FILENAME, 'wb')
wf.setnchannels(CHANNELS)
wf.setsampwidth(p.get_sample_size(FORMAT))
wf.setframerate(RATE)
wf.writeframes(b''.join(frames))
wf.close()
y , sr =librosa.load(r'C:\Users\PIYUSH\Desktop\output.wav',sr=32050)
#import librosa
s = np.abs(librosa.stft(y))
pitches, magnitudes = librosa.piptrack(y=y, sr=sr)
s_mean=(np.mean(s))
s_pitch=(np.mean(magnitudes))
result=(s_mean+s_pitch)/2
return result
def get_audio_analysis(song_url):
if(song_url is None):
return None, None, None, None
response = urlopen(song_url)
urllib.request.urlretrieve(song_url, "current.mp3")
y, sr = librosa.load("./current.mp3")
tempo, beats = librosa.beat.beat_track(y=y, sr=sr)
#Tempo = beats/minute
pitches, magnitudes = librosa.piptrack(y=y, sr=sr)
#pitch = Frequency
pitch_ave = np.average(pitches)
harm = np.sum(librosa.effects.harmonic(y))
perc = np.sum(librosa.effects.percussive(y))
return tempo, pitch_ave, harm, perc
def get_freq_mag(y,sr,window_size=None, window_shift=None):
'''
valores por defecto:
- ventanas de 25ms
- desplazamientos de la ventana de 10ms
'''
if window_size is None:
n_fft = int(0.025*sr)
else:
n_fft = int(window_size*0.001*sr)
if window_shift is None:
hop_length = int(0.010*sr)
else:
hop_length = int(window_shift*0.001*sr)
#Recoger las frecuencias presentes y sus magnitudes
frequencies,magnitudes = librosa.piptrack(y,sr,hop_length=hop_length,n_fft=n_fft)
frequencies = np.transpose(frequencies)
magnitudes = np.transpose(magnitudes)
return frequencies, magnitudes
def long_term_info(y, sr):
interval = 0.01
window = 0.256
start_stft = time.time()
stft = librosa.stft(y,
hop_length=int(interval * sr),
n_fft=int(window * sr))
end_stft = time.time()
stft = np.transpose(stft)
power = np.abs(stft)**2
start_pitch = time.time()
pitch, mag = librosa.piptrack(y=y,
sr=sr,
hop_length=int(interval * sr),
n_fft=int(window * sr))
end_pitch = time.time()
duration_stft = end_stft - start_stft
duration_pitch = end_pitch - start_pitch
pitch = np.transpose(pitch)
return stft, power, pitch
def get_freq_mag(y,sr,window_size=None, window_shift=None):
'''
default values:
- windows of 25ms
- window shifts of 10ms
'''
if window_size is None:
n_fft = int(0.025*sr)
else:
n_fft = int(window_size*0.001*sr)
if window_shift is None:
hop_length = int(0.010*sr)
else:
hop_length = int(window_shift*0.001*sr)
#collect frequencies present and their magnitudes
frequencies,magnitudes = librosa.piptrack(y,sr,hop_length=hop_length,n_fft=n_fft)
frequencies = np.transpose(frequencies)
magnitudes = np.transpose(magnitudes)
return frequencies, magnitudes
def getVar(path):
# pitches, magnitudes = lib.piptrack(y=wf, sr=sr, fmin=50, fmax=500)
# pitches = pitches[np.nonzero(pitches)]
# # pitches = np.asarray(pitches)
# # duration = np.array(list(range(0, duration)))
# plt.subplot(211)
# plt.plot(wf)
# plt.title('wave')
# plt.subplot(212)
# plt.imshow(pitches, aspect="auto", interpolation="nearest", origin="bottom")
# plt.plot(pitches)
# plt.title("pitches")
# plt.show()
y, sr = librosa.load(path)
# D = librosa.amplitude_to_db(np.abs(librosa.stft(y)), ref=np.max, top_db=10)
duration = float(lib.get_duration(y=y, sr=sr))
drf = lib.time_to_frames(duration, sr=sr)
pits, mags = lib.piptrack(y=y, sr=sr, fmin=50, fmax=500)
print("pits type === ", type(pits), end='\n')
print("megs len === ", len(mags[0]), end='\n')
print(np.shape(mags))
plt.subplot(211)
librosa.display.specshow(
pits, x_axis='time', y_axis='log, frames', fmax=500)
plt.title('pits')
plt.subplot(212)
librosa.display.specshow(mags, x_axis='time', y_axis='log')
plt.title('mags')
# plt.plot(duration)
# plt.title("check duration")
plt.show()
def findKey_arbitrary(indata):
# find key of recent layer
# find the notes for the appropriate chord in the note bank.
pitches, magnetude = librosa.piptrack(y=indata[:, 1],
sr=SAMPLE_RATE,
fmin=250.0,
fmax=1050.0)
time_slice = 1
index = magnetude[:, time_slice].argmax()
pitch = pitches[index, time_slice]
while pitch == 0.0:
time_slice = (time_slice + 10) % len(pitches)
index = magnetude[:, time_slice].argmax()
pitch = pitches[index, time_slice]
key = librosa.hz_to_note(pitch, octave=False)
print("findKey: ", key)
keyId = note_names.index(key[0])
seq = [keyId % 12, (keyId + 4) % 12,
(keyId + 7) % 12] # Major chord always
construct_chord(seq)
def calc_pitches(audio_file):
with tempfile.TemporaryDirectory() as tmpdirname:
#conversion to wav 44100 s16 mono
in_path = tmpdirname + '/in.mp3'
out_path = tmpdirname + '/out.wav'
in_file = open(in_path, 'wb')
in_file.write(audio_file.read())
audio_file.seek(0)
subprocess.call([
'ffmpeg', '-i', in_path, '-acodec', 'pcm_s16le', '-ac', '1', '-ar',
'44100', out_path
])
#pitch detection
y, sr = librosa.load(out_path, sr=44100)
duration = librosa.get_duration(y=y, sr=sr)
pitches, magnitudes = librosa.piptrack(y=y,
sr=sr,
fmin=min_freq,
fmax=max_freq,
hop_length=int(duration * sr /
10))
#for i in range(len(pitches)):
# print(pitches[i])
# print(magnitudes[i])
# print('\n')
#get unique freq bins
max_mag = np.amax(magnitudes, axis=1)
freqs = list(
set([
min_freq + freq_bin_width * round(
(f[1] - min_freq) / freq_bin_width)
for f in list(enumerate(np.amax(pitches, axis=1)))
if f[1] >= min_freq and max_mag[f[0]] > 7
]))
return freqs
return []
def run(self):
logging.info("Starting Pitch detector")
# This loop condition have to be checked frequently, so the code inside may not be blocking
while not self.terminated:
new_frame = self.audio_frames.get() # Get new frame (blocking)
if self.counter == 0:
self.frames = new_frame
self.counter += 1
elif self.counter >= BUFFER_SIZE:
self.frames = np.append(self.frames, new_frame)
pitches, magnitudes = librosa.piptrack(self.frames,
SAMPLE_RATE)
# Select out pitches with high energy
pitches = pitches[magnitudes > np.median(magnitudes)]
new_tuning = int(50 + 100 * librosa.pitch_tuning(pitches))
if np.abs(self.last_pitch -
new_tuning) > PITCH_CHANGE_THRESHOLD:
self.last_pitch = new_tuning
self.manager.new_tuning(new_tuning)
self.counter = 0
else:
self.frames = np.append(self.frames, new_frame)
self.counter += 1
def audioAnalyze(y, sr):
# librosa.piptrack to extract pitch
pitches, magnitudes = librosa.piptrack(y=y, sr=sr)
# the first ~10 columns are often meaningless
pitches_candidate = []
for time in range(10, len(pitches[0])):
pitches_candidate.append(detect_pitch(pitches, time))
# onset_detection: assumption: chord usually happen just after onset
onset_frames = librosa.onset.onset_detect(y, sr)
chord_note_candidate = []
for x in onset_frames:
if x <= len(pitches_candidate):
# so that the frequencies can be conveniently rounded to exact frequencies for excluding the harmonics
chord_note_candidate.append(
librosa.note_to_hz(librosa.hz_to_note(pitches_candidate[x])))
chord_note_no_harmonics = []
for x in chord_note_candidate:
# x is a list of list (...), so x[0] corresponding to freq in an onset
temp = []
for i in range(0, len(x[0])):
if i == 0:
temp.append(x[0][i])
for j in range(0, i):
if j == i - 1:
temp.append(x[0][i])
break
elif checkHarmonics(x[0][j], x[0][i]):
break
chord_note_no_harmonics.append(temp)
# https://stackoverflow.com/questions/3724551/python-uniqueness-for-list-of-lists
chord_note_no_harmonics = [
list(x) for x in set(tuple(x) for x in chord_note_no_harmonics)
]
return chord_note_no_harmonics
def __test(y, sr, S, n_fft, hop_length, fmin, fmax, threshold):
pitches, mags = librosa.piptrack(
y=y, sr=sr, S=S, n_fft=n_fft, hop_length=hop_length, fmin=fmin,
fmax=fmax, threshold=threshold)
import librosa
import matplotlib.pyplot as plt
filename = 'temp.wav'
y ,sr = librosa.load(filename)
print(sr)
pitches, magnitudes= librosa.piptrack(y=y, sr=sr)
mfccs = librosa.feature.mfcc(y=y, sr=sr)
print('mfcc calc done')
#visualize mfccs
librosa.display.specshow(mfccs, x_axis='time')
#plt.imshow(mfccs)
plt.colorbar()
plt.title('MFCC')
plt.tight_layout()
plt.show()
import librosa
import librosa.display
print("Now Loading file...")
s_n = "WAV_CUT/BornToBeNN_song3.wav"
y, sr = librosa.load(s_n)
print("Now Dividing file...")
y_h, y_p = librosa.effects.hpss(y)
print("Now Pitching file...")
pitches, magnitudes = librosa.piptrack(y=y, sr=sr)
s_h = "WAV_CUT/BornToBeNN_song3_har.wav"
s_p = "WAV_CUT/BornToBeNN_song3_per.wav"
print("Now Saving file...")
librosa.output.write_wav(s_h, y_h, sr)
librosa.output.write_wav(s_p, y_p, sr)
print(len(pitches))
for j in range(0, len(pitches[0]) - 1):
for i in range(0, len(pitches) - 1):
if magnitudes[i, j] != 0:
print("At " + str(0.3 * j) + " ~ " + str(0.3 * (j + 1)) +
" Seconds.")
print("Hz : " + str(pitches[i, j]) + "\tdB : " +
str(magnitudes[i, j]))
#pitches[f, t] magnitudes[f, t]
#f : Frequency num Frequency large
#t : at time t at time t
def features(X, sample_rate):
stft = np.abs(librosa.stft(X))
# fmin 和 fmax 对应于人类语音的最小最大基本频率
pitches, magnitudes = librosa.piptrack(X, sr=sample_rate, S=stft, fmin=70, fmax=400)
pitch = []
for i in range(magnitudes.shape[1]):
index = magnitudes[:, 1].argmax()
pitch.append(pitches[index, i])
pitch_tuning_offset = librosa.pitch_tuning(pitches)
pitchmean = np.mean(pitch)
pitchstd = np.std(pitch)
pitchmax = np.max(pitch)
pitchmin = np.min(pitch)
# 频谱质心
cent = librosa.feature.spectral_centroid(y=X, sr=sample_rate)
cent = cent / np.sum(cent)
meancent = np.mean(cent)
stdcent = np.std(cent)
maxcent = np.max(cent)
# 谱平面
flatness = np.mean(librosa.feature.spectral_flatness(y=X))
# 使用系数为50的MFCC特征
mfccs = np.mean(librosa.feature.mfcc(y=X, sr=sample_rate, n_mfcc=50).T, axis=0)
mfccsstd = np.std(librosa.feature.mfcc(y=X, sr=sample_rate, n_mfcc=50).T, axis=0)
mfccmax = np.max(librosa.feature.mfcc(y=X, sr=sample_rate, n_mfcc=50).T, axis=0)
# 色谱图
chroma = np.mean(librosa.feature.chroma_stft(S=stft, sr=sample_rate).T, axis=0)
# 梅尔频率
mel = np.mean(librosa.feature.melspectrogram(X, sr=sample_rate).T, axis=0)
# ottava对比
contrast = np.mean(librosa.feature.spectral_contrast(S=stft, sr=sample_rate).T, axis=0)
# 过零率
zerocr = np.mean(librosa.feature.zero_crossing_rate(X))
S, phase = librosa.magphase(stft)
meanMagnitude = np.mean(S)
stdMagnitude = np.std(S)
maxMagnitude = np.max(S)
# 均方根能量
rmse = librosa.feature.rmse(S=S)[0]
meanrms = np.mean(rmse)
stdrms = np.std(rmse)
maxrms = np.max(rmse)
ext_features = np.array([
flatness, zerocr, meanMagnitude, maxMagnitude, meancent, stdcent,
maxcent, stdMagnitude, pitchmean, pitchmax, pitchstd,
pitch_tuning_offset, meanrms, maxrms, stdrms
])
ext_features = np.concatenate((ext_features, mfccs, mfccsstd, mfccmax, chroma, mel, contrast))
return ext_features
win_length=win_length)
times = librosa.frames_to_time(beats, sr=sr)
frames = librosa.time_to_frames(
times,
sr=sr,
hop_length=hop_length,
n_fft=n_fft,
win_length=win_length,
)
# file=open((os.path.join(args.outdir, name))+'_beats.txt', "w+")
# for beat in beats:
# file.write(str(beat)+' ')
# file.close()
np.save((os.path.join(args.outdir, name)) + "_beats",
np.array(beats))
"""extract pitches"""
pitches, magnitudes = librosa.piptrack(
y=y,
sr=sr,
n_fft=n_fft,
hop_length=hop_length,
win_length=win_length,
)
pitches = pitches.T
# file=open((os.path.join(args.outdir, name))+'_pitches.txt',"w+")
pitch = np.zeros((pitches.shape[0]))
for i in range(pitches.shape[0]):
pitch[i] = max(pitches[i])
# file.close()
np.save((os.path.join(args.outdir, name)) + "_pitch", pitch)
def get_pitch(wavefile):
y, sr = librosa.load(wavefile)
if len(y) % 2 != 0:
y = y[:-1]
pitches, mag = librosa.piptrack(y=y, sr=sr)
return pitches, mag