def freq_correct(src,
dst,
fs=44100,
fc=3000,
mode='normal',
win_length=1024,
alpha=1):
out = np.zeros_like(src)
try:
if mode == 'normal':
src_oct = librosa.hz_to_octs(basefreq(src, fs, fc))
dst_oct = librosa.hz_to_octs(basefreq(dst, fs, fc))
offset = (dst_oct - src_oct) * 12 * alpha
out = librosa.effects.pitch_shift(src, 44100, n_steps=offset)
elif mode == 'track':
length = min([len(src), len(dst)])
for i in range(length // win_length):
src_oct = librosa.hz_to_octs(
basefreq(src[i * win_length:(i + 1) * win_length], fs, fc))
dst_oct = librosa.hz_to_octs(
basefreq(dst[i * win_length:(i + 1) * win_length], fs, fc))
offset = (dst_oct - src_oct) * 12 * alpha
out[i * win_length:(i + 1) *
win_length] = librosa.effects.pitch_shift(
src[i * win_length:(i + 1) * win_length],
44100,
n_steps=offset)
return out
except Exception as e:
return src
def __init__(self,
frame_size,
fmax,
fps,
oct_width,
center_note,
log_eta,
sample_rate=44100,
fold=None):
self.fps = fps
self.fmax = fmax
self.sample_rate = sample_rate
self.oct_width = oct_width
self.center_note = center_note
self.frame_size = frame_size
self.log_eta = log_eta
# parameters are based on Cho and Bello, 2014.
import librosa
ctroct = (librosa.hz_to_octs(librosa.note_to_hz(center_note))
if center_note is not None else None)
self.filterbank = librosa.filters.chroma(sr=sample_rate,
n_fft=frame_size,
octwidth=oct_width,
ctroct=ctroct).T[:-1]
# mask out everything above fmax
from bottleneck import move_mean
m = np.fft.fftfreq(frame_size,
1. / sample_rate)[:frame_size / 2] < fmax
mask_smooth = move_mean(m, window=10, min_count=1)
self.filterbank *= mask_smooth[:, np.newaxis]
def test_hz_to_octs(tuning, bins_per_octave):
freq = np.asarray([55, 110, 220, 440]) * (2.0**(tuning / bins_per_octave))
octs = [1, 2, 3, 4]
oct_out = librosa.hz_to_octs(freq,
tuning=tuning,
bins_per_octave=bins_per_octave)
assert np.allclose(octs, oct_out)
def test_hz_to_octs_dep(a440):
with warnings.catch_warnings(record=True) as out:
freq = np.asarray([55, 110, 220, 440]) * (float(a440) / 440.0)
octs = [1, 2, 3, 4]
oct_out = librosa.hz_to_octs(freq, A440=a440)
assert np.allclose(octs, oct_out)
# And that it says the right thing (roughly)
assert 'deprecated' in str(out[0].message).lower()
def main():
util.clean_tempfiles()
#_,Aaudio = sound.load('./music/大.wav')
Aaudio = dsp.sin(1000, 44100, 0.1)*10000
# print(sound.basefreq(Aaudio, 44100, fc=3000))
# xk,fft = dsp.showfreq(Aaudio,fs=44100,fc=5000)
# plt.plot(xk,fft)
# plt.show()
# Aaudio = Aaudio[1500:]
# print(Aaudio)
# sound.playtest(Aaudio)
Baudio,Bsyllables,Bfeatures,Bpeakindexs,Bbias = process_audio('./music/小星星.wav', './tmp/audio1', 0.07, 'time',crop_time=0.1,hpss='')
#sound.playtest(dsp.fft_filter(Baudio,fs=44100,fc=[800,8000]))
for syllable in Bsyllables:
base = dsp.basefreq(syllable, fs=44100, fc=3000)
print(base,librosa.hz_to_octs(base))
xk,fft = dsp.showfreq(syllable,fs=44100,fc=3000)
plt.plot(xk,fft)
plt.show()
#sound.playtest(Baudio)
make_B_by_A(Aaudio,[Baudio,Bsyllables,Bfeatures,Bpeakindexs,Bbias],fc=[400,8000])
def __test_to_octs(infile):
DATA = load(infile)
z = librosa.hz_to_octs(DATA['f'])
assert np.allclose(z, DATA['result'])
def chroma(sr, n_fft, n_chroma=12, A440=440.0, ctroct=5.0, octwidth=2):
"""Create a Filterbank matrix to convert STFT to chroma
:usage:
>>> # Build a simple chroma filter bank
>>> chroma_fb = librosa.filters.chroma(22050, 4096)
>>> # Use quarter-tones instead of semitones
>>> chroma_fbq = librosa.filters.chroma(22050, 4096, n_chroma=24)
>>> # Equally weight all octaves
>>> chroma_fb = librosa.filters.chroma(22050, 4096, octwidth=None)
:parameters:
- sr : int > 0 [scalar]
audio sampling rate
- n_fft : int > 0 [scalar]
number of FFT bins
- n_chroma : int > 0 [scalar]
number of chroma bins
- A440 : float > 0 [scalar]
Reference frequency for A440
- ctroct : float > 0 [scalar]
- octwidth : float > 0 or None [scalar]
``ctroct`` and ``octwidth`` specify a dominance window -
a Gaussian weighting centered on ``ctroct`` (in octs, A0 = 27.5Hz)
and with a gaussian half-width of ``octwidth``.
Set ``octwidth`` to ``None`` to use a flat weighting.
:returns:
- wts : ndarray [shape=(n_chroma, 1 + n_fft / 2)]
Chroma filter matrix
"""
wts = np.zeros((n_chroma, n_fft))
# Get the FFT bins, not counting the DC component
frequencies = np.linspace(0, sr, n_fft, endpoint=False)[1:]
frqbins = n_chroma * librosa.hz_to_octs(frequencies, A440)
# make up a value for the 0 Hz bin = 1.5 octaves below bin 1
# (so chroma is 50% rotated from bin 1, and bin width is broad)
frqbins = np.concatenate(([frqbins[0] - 1.5 * n_chroma], frqbins))
binwidthbins = np.concatenate((np.maximum(frqbins[1:] - frqbins[:-1],
1.0), [1]))
D = np.subtract.outer(frqbins, np.arange(0, n_chroma, dtype='d')).T
n_chroma2 = np.round(float(n_chroma) / 2)
# Project into range -n_chroma/2 .. n_chroma/2
# add on fixed offset of 10*n_chroma to ensure all values passed to
# rem are +ve
D = np.remainder(D + n_chroma2 + 10 * n_chroma, n_chroma) - n_chroma2
# Gaussian bumps - 2*D to make them narrower
wts = np.exp(-0.5 * (2 * D / np.tile(binwidthbins, (n_chroma, 1)))**2)
# normalize each column
wts = librosa.util.normalize(wts, norm=2, axis=0)
# Maybe apply scaling for fft bins
if octwidth is not None:
wts *= np.tile(
np.exp(-0.5 * (((frqbins / n_chroma - ctroct) / octwidth)**2)),
(n_chroma, 1))
# remove aliasing columns, copy to ensure row-contiguity
return np.ascontiguousarray(wts[:, :int(1 + n_fft / 2)])
def chroma(sr, n_fft, n_chroma=12, A440=440.0, ctroct=5.0, octwidth=2):
"""Create a Filterbank matrix to convert STFT to chroma
:usage:
>>> # Build a simple chroma filter bank
>>> chroma_fb = librosa.filters.chroma(22050, 4096)
>>> # Use quarter-tones instead of semitones
>>> chroma_fbq = librosa.filters.chroma(22050, 4096, n_chroma=24)
>>> # Equally weight all octaves
>>> chroma_fb = librosa.filters.chroma(22050, 4096, octwidth=None)
:parameters:
- sr : int > 0
audio sampling rate
- n_fft : int > 0
number of FFT bins
- n_chroma : int > 0
number of chroma bins
- A440 : float
Reference frequency for A440
- ctroct : float > 0
- octwidth : float or None
``ctroct`` and ``octwidth`` specify a dominance window -
a Gaussian weighting centered on ``ctroct`` (in octs, A0 = 27.5Hz)
and with a gaussian half-width of ``octwidth``.
Set ``octwidth`` to ``None`` to use a flat weighting.
:returns:
- wts : ndarray, shape=(n_chroma, 1 + n_fft / 2)
Chroma filter matrix
"""
wts = np.zeros((n_chroma, n_fft))
# Get the FFT bins, not counting the DC component
frequencies = np.linspace(0, sr, n_fft, endpoint=False)[1:]
frqbins = n_chroma * librosa.hz_to_octs(frequencies, A440)
# make up a value for the 0 Hz bin = 1.5 octaves below bin 1
# (so chroma is 50% rotated from bin 1, and bin width is broad)
frqbins = np.concatenate(([frqbins[0] - 1.5 * n_chroma], frqbins))
binwidthbins = np.concatenate((np.maximum(frqbins[1:] - frqbins[:-1],
1.0), [1]))
D = np.subtract.outer(frqbins, np.arange(0, n_chroma, dtype='d')).T
n_chroma2 = np.round(n_chroma / 2.0)
# Project into range -n_chroma/2 .. n_chroma/2
# add on fixed offset of 10*n_chroma to ensure all values passed to
# rem are +ve
D = np.remainder(D + n_chroma2 + 10*n_chroma, n_chroma) - n_chroma2
# Gaussian bumps - 2*D to make them narrower
wts = np.exp(-0.5 * (2*D / np.tile(binwidthbins, (n_chroma, 1)))**2)
# normalize each column
wts = librosa.util.normalize(wts, norm=2, axis=0)
# Maybe apply scaling for fft bins
if octwidth is not None:
wts *= np.tile(
np.exp(-0.5 * (((frqbins/n_chroma - ctroct)/octwidth)**2)),
(n_chroma, 1))
# remove aliasing columns, copy to ensure row-contiguity
return np.ascontiguousarray(wts[:, :(1 + n_fft/2)])
def __test(a440):
freq = np.asarray([55, 110, 220, 440]) * (float(a440) / 440.0)
octs = [1, 2, 3, 4]
oct_out = librosa.hz_to_octs(freq, A440=a440)
assert np.allclose(octs, oct_out)
def test_hz_to_octs(infile):
DATA = load(infile)
z = librosa.hz_to_octs(DATA["f"])
assert np.allclose(z, DATA["result"])
def __test(a440):
freq = np.asarray([55, 110, 220, 440]) * (float(a440) / 440.0)
octs = [1, 2, 3, 4]
oct_out = librosa.hz_to_octs(freq, A440=a440)
assert np.allclose(octs, oct_out)
fps, endtime, height, width = ffmpeg.get_video_infos(
os.path.join(dataset, video_names[i]))
util.makedirs(os.path.join('./tmp/video2image', '%03d' % i))
ffmpeg.video2image(
os.path.join(dataset, video_names[i]),
os.path.join('./tmp/video2image', '%03d' % i, '%05d.jpg'))
print('Generating voice...')
sinmusic, musicinfos = notation.notations2music(notations, mode='sin')
music = np.zeros_like(sinmusic)
for i in range(len(musicinfos['time'])):
for j in range(len(musicinfos['freq'][i])):
if musicinfos['freq'][i][j] != 0:
diff = np.abs(
librosa.hz_to_octs(seed_freqs) -
librosa.hz_to_octs(musicinfos['freq'][i][j]))
index = np.argwhere(diff == np.min(diff))[0][0]
_tone = seed_voices[index]
_tone = sound.freq_correct(_tone,
srcfreq=seed_freqs[index],
dstfreq=musicinfos['freq'][i][j],
alpha=1.0)
_tone = sound.highlight_bass(_tone, musicinfos['freq'][i][j],
seed_freqs[index])
music[int(musicinfos['time'][i] *
44100):int(musicinfos['time'][i] * 44100) +
len(_tone)] += _tone
# music = music+sinmusic*0.1
music = dsp.bpf(music, 44100, 20, 5000)
music = (arrop.sigmoid(music / 32768) - 0.5) * 65536
def chroma(sr, n_fft, n_chroma=12, A440=440.0, ctroct=5.0, octwidth=None):
"""Create a Filterbank matrix to convert STFT to chroma
:usage:
>>> # Build a simple chroma filter bank
>>> chroma_fb = librosa.filters.chroma(22050, 4096)
>>> # Use quarter-tones instead of semitones
>>> chroma_fbq = librosa.filters.chroma(22050, 4096, n_chroma=24)
>>> # Down-weight the high and low frequencies
>>> chroma_fb = librosa.filters.chroma(22050, 4096, ctroct=5, octwidth=2)
:parameters:
- sr : int
audio sampling rate
- n_fft : int
FFT window size
- n_chroma : int
number of chroma bins
- A440 : float
Reference frequency for A440
- ctroct : float
- octwidth : float
These parameters specify a dominance window - Gaussian
weighting centered on ctroct (in octs, re A0 = 27.5Hz) and
with a gaussian half-width of octwidth.
Defaults to halfwidth = inf, i.e. flat.
:returns:
- wts : ndarray, shape=(n_chroma, 1 + n_fft / 2)
Chroma filter matrix
"""
wts = np.zeros((n_chroma, n_fft))
# Get the FFT bins, not counting the DC component
frequencies = np.linspace(0, sr, n_fft, endpoint=False)[1:]
fftfrqbins = n_chroma * librosa.hz_to_octs(frequencies, A440)
# make up a value for the 0 Hz bin = 1.5 octaves below bin 1
# (so chroma is 50% rotated from bin 1, and bin width is broad)
fftfrqbins = np.concatenate( ( [fftfrqbins[0] - 1.5 * n_chroma],
fftfrqbins))
binwidthbins = np.concatenate(
(np.maximum(fftfrqbins[1:] - fftfrqbins[:-1], 1.0), [1]))
D = np.tile(fftfrqbins, (n_chroma, 1)) \
- np.tile(np.arange(0, n_chroma, dtype='d')[:, np.newaxis],
(1, n_fft))
n_chroma2 = round(n_chroma / 2.0)
# Project into range -n_chroma/2 .. n_chroma/2
# add on fixed offset of 10*n_chroma to ensure all values passed to
# rem are +ve
D = np.remainder(D + n_chroma2 + 10*n_chroma, n_chroma) - n_chroma2
# Gaussian bumps - 2*D to make them narrower
wts = np.exp(-0.5 * (2*D / np.tile(binwidthbins, (n_chroma, 1)))**2)
# normalize each column
wts /= np.tile(np.sqrt(np.sum(wts**2, 0)), (n_chroma, 1))
# Maybe apply scaling for fft bins
if octwidth is not None:
wts *= np.tile(
np.exp(-0.5 * (((fftfrqbins/n_chroma - ctroct)/octwidth)**2)),
(n_chroma, 1))
# remove aliasing columns
return wts[:, :(1 + n_fft/2)]
