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)
import featureflow as ff
import zounds
import numpy as np
windowing = zounds.HalfLapped()
# Segment audio files #########################################################
class Settings(ff.PersistenceSettings):
id_provider = ff.UserSpecifiedIdProvider(key='_id')
key_builder = ff.StringDelimitedKeyBuilder(seperator='|')
database = ff.LmdbDatabase(path='onsetdata', key_builder=key_builder)
event_log = ff.EventLog(
path='onsetdataevents', channel=ff.InMemoryChannel())
STFT = zounds.stft(
resample_to=zounds.SR11025(),
wscheme=windowing)
class WithOnsets(STFT, Settings):
bark = zounds.ArrayWithUnitsFeature(
zounds.BarkBands,
needs=STFT.fft,
store=True)
transience = zounds.ArrayWithUnitsFeature(
zounds.MeasureOfTransience,
needs=STFT.fft,
BaseModel = zounds.resampled(resample_to=samplerate, store_resampled=True)
scale_bands = 96
spectrogram_duration = 64
anchor_slice = slice(spectrogram_duration, spectrogram_duration * 2)
scale = zounds.GeometricScale(start_center_hz=50,
stop_center_hz=samplerate.nyquist,
bandwidth_ratio=0.115,
n_bands=scale_bands)
scale.ensure_overlap_ratio()
spectrogram_duration = 64
windowing_scheme = zounds.HalfLapped()
spectrogram_sample_rate = zounds.SampleRate(
frequency=windowing_scheme.frequency * (spectrogram_duration // 2),
duration=windowing_scheme.frequency * spectrogram_duration)
def spectrogram(x):
x = apply_scale(np.abs(x.real),
scale,
window=zounds.OggVorbisWindowingFunc())
x = zounds.log_modulus(x * 100)
return x * zounds.AWeighting()
@zounds.simple_lmdb_settings('spectrogram_embedding',
map_size=1e11,
def _complex_stft(cls, samples, window_sr=None):
window_sr = window_sr or zounds.HalfLapped()
return zounds.spectral.stft(samples, window_sr)
