def test_yin_multi(y_multi):
y, sr = y_multi
Pall = librosa.yin(y, fmin=30, fmax=300)
P0 = librosa.yin(y[0], fmin=30, fmax=300)
P1 = librosa.yin(y[1], fmin=30, fmax=300)
assert np.allclose(Pall[0], P0)
assert np.allclose(Pall[1], P1)
assert not np.allclose(P0, P1)
def compute():
self.F0 = librosa.yin(y=samples.mean(axis=1),
sr=self.samplerate,
frame_length=2048,
fmin=librosa.note_to_hz('C2'),
fmax=librosa.note_to_hz('C5'),
center=False)
def analyze(cls, events: List[CorpusEvent],
metadata: Metadata) -> List[CorpusEvent]:
if not FeatureUtils.is_valid_audio(events, metadata):
raise FeatureError(
f"Feature '{cls.__name__}' does not support content of "
f"type {metadata.content_type.__class__.__name__}")
metadata: AudioMetadata = typing.cast(AudioMetadata, metadata)
# TODO: Pass rather than hard-code
yin_frames: np.ndarray = librosa.yin(metadata.foreground_data,
fmin=50,
fmax=4186,
sr=metadata.sr,
frame_length=2048,
hop_length=metadata.hop_length)
yin_midipitches: np.ndarray = np.round(
12 * np.log2(yin_frames / 8.175798915643707))
for event in events:
onset_frame: int = librosa.time_to_frames(
event.onset, sr=metadata.sr, hop_length=metadata.hop_length)
end_frame: int = librosa.time_to_frames(
event.onset + event.duration,
sr=metadata.sr,
hop_length=metadata.hop_length)
hist, _ = np.histogram(yin_midipitches[onset_frame:end_frame],
bins=128,
range=(0, 128))
pitch: int = int(np.argmax(hist))
event.set_feature(cls(value=pitch))
return events
def calculate_f0(file_path, fmin=65, fmax=2093, sample_rate=44100, frame_length=2048, win_length=735 * 2,
hop_length=735, threshold=0.1, seuil_voice=250):
"""
We calculated the fundamental frequency by using the algorithm yin and all the f0 higher than seuil_voice will be
considered as unvoiced and be reset to 0
:param seuil_voice:
:param threshold:
:param file_path: the path of the wav file .wav
:param fmin:
:param fmax:
:param sample_rate: the sample rate of the wav file
:param frame_length:
:param win_length:
:param hop_length:
:return: a numpy array which contains the f0 at every moment
"""
wav_file, _ = librosa.load(file_path, sr=sample_rate)
result_f0 = librosa.yin(wav_file, fmin=fmin, fmax=fmax, sr=sample_rate, frame_length=frame_length,
win_length=win_length, hop_length=hop_length, trough_threshold=threshold)
# choose a threshold to distinguish voiced and unvoiced part of the waveform
# for the unvoiced parts, we consider the F0 = 0
result_f0[result_f0 >= seuil_voice] = 0
# initialize a vector to save the voiced unvoiced flags
uv_flags = np.zeros(result_f0.shape)
uv_flags[result_f0 > 0] = 1
return result_f0, uv_flags
def __call__(self, data):
signal = data['signal']
sr = data['sample_rate']
self.n_fft = int(np.ceil(0.025 * sr))
self.win_length = int(np.ceil(0.025 * sr))
self.hop_length = int(np.ceil(0.01 * sr))
audio_mfcc = librosa.feature.mfcc(y=signal.numpy().reshape(-1),
sr=sr,
n_mfcc=13,
n_fft=self.n_fft,
hop_length=self.hop_length)
#f0, voiced_flag, voiced_probs = librosa.pyin(y=signal.numpy().reshape(-1),
# sr=sr,
# frame_length = self.n_fft,
# hop_length = self.hop_length,
# fmin=librosa.note_to_hz('C2'),
# fmax=librosa.note_to_hz('C7'))
f0 = librosa.yin(y=signal.numpy().reshape(-1),
sr=sr,
frame_length=self.n_fft,
hop_length=self.hop_length,
fmin=librosa.note_to_hz('C2'),
fmax=librosa.note_to_hz('C7'))
f0[f0 > 500] = 0
f0[np.isnan(f0)] = 0
delta1 = librosa.feature.delta(audio_mfcc)
#delta1 = delta1 / np.linalg.norm(delta1)
delta2 = librosa.feature.delta(audio_mfcc, order=2)
#delta2 = delta2 / np.linalg.norm(delta2)
mfcc_result = np.concatenate(
(audio_mfcc.transpose(), delta1.transpose(), delta2.transpose(),
f0.transpose().reshape((-1, 1))),
axis=1)
mfcc_result = torch.from_numpy(mfcc_result)
mfcc_result = mfcc_pad2(mfcc_result)
#f0 = f0 / np.linalg.norm(f0)
data['MFCC'] = mfcc_result
data['input'] = mfcc_result
return data
def __call__(self, data):
signal = data['signal']
sr = data['sample_rate']
self.n_fft = int(np.ceil(0.025 * sr))
self.win_length = int(np.ceil(0.025 * sr))
self.hop_length = int(np.ceil(0.01 * sr))
audio_mfcc = torch.FloatTensor(
librosa.feature.mfcc(y=signal.numpy().reshape(-1),
sr=sr,
n_mfcc=13,
n_fft=self.n_fft,
hop_length=self.hop_length))
#f0, voiced_flag, voiced_probs = librosa.pyin(y=signal.numpy().reshape(-1),
# sr=sr,
# frame_length=n_fft,
# hop_length=hop_length,
# fmin=librosa.note_to_hz('C2'),
# fmax=librosa.note_to_hz('C7'))
f0 = torch.FloatTensor(
librosa.yin(y=signal.numpy().reshape(-1),
sr=sr,
frame_length=self.n_fft,
hop_length=self.hop_length,
fmin=librosa.note_to_hz('C2'),
fmax=librosa.note_to_hz('C7')))
f0[f0 > 500] = 0
f0[torch.isnan(f0)] = 0
delta1 = torch.FloatTensor(librosa.feature.delta(audio_mfcc))
delta2 = torch.FloatTensor(librosa.feature.delta(audio_mfcc, order=2))
# mfcc_result = np.concatenate((audio_mfcc, delta1, delta2), axis=0)
f0_delta1 = torch.FloatTensor(librosa.feature.delta(f0)).view(1, -1)
f0_delta2 = torch.FloatTensor(librosa.feature.delta(f0, order=2)).view(
1, -1)
audio_mfcc = np.concatenate((audio_mfcc, f0.reshape((1, -1))))
delta1 = torch.cat((delta1, f0_delta1))
delta2 = torch.cat((delta2, f0_delta2))
audio_mfcc = torch.from_numpy(audio_mfcc)
mfcc_result = torch.stack([audio_mfcc, delta1, delta2])
mfcc_result = mfcc_pad1(mfcc_result)
data['MFCC'] = mfcc_result
data['input'] = mfcc_result
return data
def calculate_f0_without_threshold(file_path, fmin=65, fmax=2093, sample_rate=44100, frame_length=2048,
win_length=735 * 2,
hop_length=735, threshold=0.1):
"""
calculation of the fundamental frequency without the threshold
:param file_path:
:param fmin:
:param fmax:
:param sample_rate:
:param frame_length:
:param win_length:
:param hop_length:
:param threshold:
:return:
"""
wav_file, _ = librosa.load(file_path, sr=sample_rate)
result_f0 = librosa.yin(wav_file, fmin=fmin, fmax=fmax, sr=sample_rate, frame_length=frame_length,
win_length=win_length, hop_length=hop_length, trough_threshold=threshold)
return result_f0
import sounddevice as sd
from scipy.io.wavfile import write
import librosa
fs = 44100 #Sample rate
seconds = 4 #duration of recording
print('start recording')
myrecording = sd.rec(int(seconds * fs), samplerate=fs, channels=1)
sd.wait() #wait until recording is finished
print('finished recording')
write('output.wav', fs, myrecording) #save as wav file
y, sr = librosa.load('output.wav')
librosa.yin(y, fmin=310, fmax=400)
# Guitar strings are E2=82.41Hz, A2=110Hz, D3=146.8Hz, G3=196Hz, B3=246.9Hz, E4=329.6Hz
y = normalize_signal(y)
f_min = 80 #Hz
f_max = 1300 #Hz
taus = np.arange(start=sr // f_max, stop=sr // f_min)
coinv_window = int(taus[-1] + 1)
n_windows = len(y) // coinv_window - 1
pitches1 = step1(y)
pitches2 = step2(y)
pitches3 = step3(y)
pitches4 = step4(y)
pitches5 = step5(y)
pitches6 = step6(y)
librosa_pitches = librosa.yin(y, 100, 1300, win_length=coinv_window)
fig, axs = plt.subplots(3, 2)
custom_xlim = 0, len(y) / sr
custom_ylim = 0, max(pitches5) * 1.25
axs[0, 0].plot(np.linspace(0, len(y) / sr, num=n_windows), pitches1)
axs[0, 0].hlines(440,
-0.1 * len(y) / sr,
len(y) / sr * 1.1,
colors='r',
linestyles='--')
axs[0, 0].set_title('Step 1')
axs[0, 1].plot(np.linspace(0, len(y) / sr, num=n_windows), pitches2)
axs[0, 1].hlines(440,
-0.1 * len(y) / sr,
len(y) / sr * 1.1,
colors='r',
def __init__(self, *args, **kwargs):
super(self.__class__, self).__init__(*args, **kwargs)
args, _ = parser.parse_known_args()
if args.samplerate == None:
self.samplerate = \
int(sd.query_devices(args.input_device)['default_samplerate'])
else:
self.samplerate = int(args.samplerate)
print(f"INFO -- Sampling rate at {self.samplerate} Hz")
self.threadpool = QtCore.QThreadPool()
self.q = queue.Queue()
self.setFixedSize(args.width, args.height)
self.mainbox = QtWidgets.QWidget()
self.setCentralWidget(self.mainbox)
self.layout = QtWidgets.QGridLayout()
self.mainbox.setLayout(self.layout)
# Widgets
self.spec_plot = SpectrogramWidget()
self.wave_plot = WaveFormWidget()
for i, widget in enumerate([self.spec_plot, self.wave_plot]):
self.layout.addWidget(widget, i, 0)
# Initialize x and y
self.length = self.samplerate * args.duration
self.y = np.random.rand(self.length, len(args.channels))
self.x = np.linspace(0, args.duration, num=self.length)
self.zcr = librosa.feature.zero_crossing_rate(self.y.mean(axis=1))[0]
# Wave Plot
self.waveline_1 = self.wave_plot.plot(x=self.x,
y=self.y[:, 0],
pen=pg.mkPen('g', width=0.5),
name='channel_1')
self.waveline_2 = self.wave_plot.plot(x=self.x,
y=self.y[:, 1],
pen=pg.mkPen('y', width=0.5),
name='channel_2')
self.waveline_3 = self.wave_plot.plot(x=np.linspace(
0, args.duration, self.zcr.shape[0]),
y=self.zcr,
pen=pg.mkPen('r', width=2),
name='zcr')
# Spectrogram
self.fmax = int(
librosa.core.fft_frequencies(sr=self.samplerate,
n_fft=args.n_fft)[-1])
D = librosa.stft(y=self.y.mean(axis=1), n_fft=args.n_fft, center=False)
self.specdata = librosa.amplitude_to_db(np.abs(D), ref=np.max)
# M = librosa.feature.melspectrogram(
# y=self.y.mean(axis=1),
# sr=self.samplerate,
# n_fft=args.n_fft,
# n_mels=args.n_mels)
# self.specdata = librosa.power_to_db(S=M, ref=np.max)
self.F0 = librosa.yin(y=self.y.mean(axis=1),
sr=self.samplerate,
frame_length=2048,
fmin=librosa.note_to_hz('C2'),
fmax=librosa.note_to_hz('C5'),
center=False)
self.spec_image = pg.ImageItem(item=self.specdata.T)
self.spec_plot.addItem(item=self.spec_image)
self.f0_line = self.spec_plot.plot(x=np.linspace(
0, args.duration, self.F0.shape[0]),
y=self.F0,
pen=pg.mkPen('r', width=2),
name='f0')
self.bar = pg.ColorBarItem(values=(librosa.note_to_hz('C2'),
librosa.note_to_hz('C5')),
cmap=pg.colormap.get('CET-L9'))
self.bar.setImageItem(self.spec_image)
# Start audio stream and animations
self.start_stream()
if args.input_device == 'Virtual Input (VB-Audio Virtual Cable), Windows DirectSound':
self.play_media(media_url=args.media_url,
type='stream',
volume=100)
self.animate()
self.show()
'audio_files/oh/oh_0.wav',
'audio_files/eeh/eeh_0.wav',
'audio_files/ah/ah_1.wav',
'audio_files/oh/oh_1.wav',
'audio_files/eeh/eeh_1.wav',
'audio_files/ah/ah_2.wav',
'audio_files/oh/oh_2.wav',
'audio_files/eeh/eeh_2.wav']
sythesized = np.zeros(0)
recorded = np.zeros(0)
for i, file_name in enumerate(file_names):
y, sr = librosa.load(file_name, sr=44100)
pitch = librosa.yin(y, 50, 1000)
fr = np.mean(pitch)
n_fft = 8096
hop_size = 256
p = librosa.stft(y, n_fft=n_fft, hop_length=hop_size)
d = librosa.amplitude_to_db( np.abs(p), ref=np.max )
freqs = librosa.fft_frequencies(sr=sr, n_fft=n_fft)
m = np.mean( d , axis=1 )
m -= np.min( m )
def moving_average(x, w):
x = np.hstack( ( x[0]*np.ones(w//2) , x , x[-1]*np.ones(w//2-1) ) )
ma = np.convolve(x, np.ones(w), 'valid') / w
def f0(wave_form, sample_rate, hop_length, fmin=80, fmax=10000):
return yin(wave_form, fmin=fmin, fmax=fmax,
hop_length=hop_length).reshape(-1, 1)
def pitch(samples: Samples) -> float:
freqs = librosa.yin(samples.data, 65, 2000, sr=samples.rate)
return np.mean(freqs)
def get_f0(self):
f0 = lr.yin(self.samples, 65, 2093, SAMPLE_RATE)
return np.mean(f0)