class Sound(BaseModel):
"""
An audio processing pipeline that computes a frequency domain representation
of the sound that follows a geometric scale
"""
bark = zounds.ArrayWithUnitsFeature(zounds.BarkBands,
samplerate=samplerate,
stop_freq_hz=samplerate.nyquist,
needs=BaseModel.fft,
store=True)
long_windowed = zounds.ArrayWithUnitsFeature(
zounds.SlidingWindow,
wscheme=zounds.SampleRate(frequency=zounds.Milliseconds(358),
duration=zounds.Milliseconds(716)),
wfunc=zounds.OggVorbisWindowingFunc(),
needs=BaseModel.resampled,
store=True)
long_fft = zounds.ArrayWithUnitsFeature(zounds.FFT,
needs=long_windowed,
store=True)
freq_adaptive = zounds.FrequencyAdaptiveFeature(
zounds.FrequencyAdaptiveTransform,
transform=np.fft.irfft,
scale=scale,
window_func=np.hanning,
needs=long_fft,
store=False)
rasterized = zounds.ArrayWithUnitsFeature(lambda fa: fa.rasterize(64),
needs=freq_adaptive,
store=False)
class SoundWithNoSettings(BaseModel):
short_windowed = zounds.ArrayWithUnitsFeature(
zounds.SlidingWindow,
wscheme=windowing_scheme,
wfunc=zounds.OggVorbisWindowingFunc(),
needs=BaseModel.resampled)
fft = zounds.ArrayWithUnitsFeature(
zounds.FFT,
needs=short_windowed)
geom = zounds.ArrayWithUnitsFeature(
spectrogram,
needs=fft,
store=True)
log_spectrogram = zounds.ArrayWithUnitsFeature(
zounds.SlidingWindow,
wscheme=zounds.SampleRate(
frequency=windowing_scheme.frequency * (spectrogram_duration // 2),
duration=windowing_scheme.frequency * spectrogram_duration * 3),
needs=geom)
ls = zounds.ArrayWithUnitsFeature(
zounds.SlidingWindow,
wscheme=spectrogram_sample_rate,
needs=geom)
class Sound(Resampled):
"""
A simple pipeline that computes a perceptually weighted modified discrete
cosine transform, and "persists" feature data in an in-memory store.
"""
windowed = zounds.ArrayWithUnitsFeature(
zounds.SlidingWindow,
needs=Resampled.resampled,
wscheme=zounds.HalfLapped(),
wfunc=zounds.OggVorbisWindowingFunc(),
store=True)
mdct = zounds.ArrayWithUnitsFeature(zounds.MDCT, needs=windowed)
weighted = zounds.ArrayWithUnitsFeature(lambda x: x * zounds.AWeighting(),
needs=mdct)
def spectrogram(x):
x = apply_scale(np.abs(x.real),
scale,
window=zounds.OggVorbisWindowingFunc())
x = zounds.log_modulus(x * 100)
return x * zounds.AWeighting()
MDCT invertibility requirements, is tricky, and requires some low level
knowledge that zounds' Scale attempts to abstract away.
See section 3.3 Setting MDCT Sizes for information about what we're fudging/
glossing over in this implementation. We instead use the DCT2 transform, which
makes inversion easier, at the cost of more redundancy.
"""
from __future__ import division
import zounds
import scipy
samplerate = zounds.SR11025()
BaseModel = zounds.stft(resample_to=samplerate)
windowing_func = zounds.OggVorbisWindowingFunc()
scale = zounds.GeometricScale(300, 3030, 0.05, 100)
@zounds.simple_in_memory_settings
class Document(BaseModel):
bark = zounds.ArrayWithUnitsFeature(zounds.BarkBands,
samplerate=samplerate,
stop_freq_hz=samplerate.nyquist,
needs=BaseModel.fft,
store=True)
long_windowed = zounds.ArrayWithUnitsFeature(
zounds.SlidingWindow,
wscheme=zounds.SampleRate(frequency=zounds.Milliseconds(500),