def resampled(chunksize_bytes=DEFAULT_CHUNK_SIZE,
resample_to=SR44100(),
store_resampled=False):
"""
Create a basic processing pipeline that can resample all incoming audio
to a normalized sampling rate for downstream processing, and store a
convenient, compressed version for playback
:param chunksize_bytes: The number of bytes from the raw stream to process
at once
:param resample_to: The new, normalized sampling rate
:return: A simple processing pipeline
"""
class Resampled(BaseModel):
meta = JSONFeature(MetaData, store=True, encoder=AudioMetaDataEncoder)
raw = ByteStreamFeature(ByteStream,
chunksize=chunksize_bytes,
needs=meta,
store=False)
ogg = OggVorbisFeature(OggVorbis, needs=raw, store=True)
pcm = AudioSamplesFeature(AudioStream, needs=raw, store=False)
resampled = AudioSamplesFeature(Resampler,
needs=pcm,
samplerate=resample_to,
store=store_resampled)
return Resampled
def setUp(self):
self.samplerate = SR44100()
rs = resampled(resample_to=self.samplerate)
wscheme = HalfLapped()
@simple_in_memory_settings
class Document(rs):
windowed = ArrayWithUnitsFeature(
SlidingWindow,
wscheme=wscheme,
wfunc=OggVorbisWindowingFunc(),
needs=rs.resampled,
store=False)
fft = ArrayWithUnitsFeature(
FFT,
needs=windowed,
store=False)
centroid = ArrayWithUnitsFeature(
SpectralCentroid,
needs=fft,
store=True)
ss = SineSynthesizer(self.samplerate)
chunks = \
[ss.synthesize(Seconds(1), [440 * i]) for i in range(1, 6)]
self.audio = \
AudioSamples(ArrayWithUnits.concat(chunks), self.samplerate)
_id = Document.process(meta=self.audio.encode())
self.doc = Document(_id)
def frequency_adaptive(long_window_sample_rate,
scale,
store_freq_adaptive=False,
check_scale_overlap_ratio=False,
chunksize_bytes=DEFAULT_CHUNK_SIZE,
resample_to=SR44100(),
store_resampled=False):
BaseModel = resampled(chunksize_bytes, resample_to, store_resampled)
class FrequencyAdaptive(BaseModel):
long_windowed = ArrayWithUnitsFeature(SlidingWindow,
wscheme=long_window_sample_rate,
wfunc=OggVorbisWindowingFunc(),
needs=BaseModel.resampled,
store=False)
long_fft = ArrayWithUnitsFeature(FFT, needs=long_windowed, store=False)
freq_adaptive = FrequencyAdaptiveFeature(
FrequencyAdaptiveTransform,
transform=np.fft.irfft,
scale=scale,
check_scale_overlap_ratio=check_scale_overlap_ratio,
window_func=np.hanning,
needs=long_fft,
store=store_freq_adaptive)
return FrequencyAdaptive
def test_can_repr(self):
cs = ChunkSizeBytes(SR44100(), Seconds(30), channels=2, bit_depth=16)
s = cs.__repr__()
self.assertEqual(
'ChunkSizeBytes(samplerate=SR44100(f=2.2675736e-05, '
'd=2.2675736e-05), duration=30 seconds, channels=2, bit_depth=16)',
s)
def test_matches_fftfreq(self):
samplerate = SR44100()
n_bands = 2048
fft_freqs = np.fft.rfftfreq(n_bands, 1 / int(samplerate))
bands = LinearScale.from_sample_rate(samplerate, n_bands // 2)
linear_freqs = np.array([b.start_hz for b in bands])
np.testing.assert_allclose(linear_freqs, fft_freqs[:-1])
def audio_graph(chunksize_bytes=DEFAULT_CHUNK_SIZE,
resample_to=SR44100(),
store_fft=False):
"""
Produce a base class suitable as a starting point for many audio processing
pipelines. This class resamples all audio to a common sampling rate, and
produces a bark band spectrogram from overlapping short-time fourier
transform frames. It also compresses the audio into ogg vorbis format for
compact storage.
"""
band = FrequencyBand(20, resample_to.nyquist)
class AudioGraph(BaseModel):
meta = JSONFeature(MetaData, store=True, encoder=AudioMetaDataEncoder)
raw = ByteStreamFeature(ByteStream,
chunksize=chunksize_bytes,
needs=meta,
store=False)
ogg = OggVorbisFeature(OggVorbis, needs=raw, store=True)
pcm = AudioSamplesFeature(AudioStream, needs=raw, store=False)
resampled = AudioSamplesFeature(Resampler,
needs=pcm,
samplerate=resample_to,
store=False)
windowed = ArrayWithUnitsFeature(SlidingWindow,
needs=resampled,
wscheme=HalfLapped(),
wfunc=OggVorbisWindowingFunc(),
store=False)
dct = ArrayWithUnitsFeature(DCT, needs=windowed, store=True)
fft = ArrayWithUnitsFeature(FFT, needs=windowed, store=store_fft)
bark = ArrayWithUnitsFeature(BarkBands,
needs=fft,
frequency_band=band,
store=True)
centroid = ArrayWithUnitsFeature(SpectralCentroid,
needs=bark,
store=True)
chroma = ArrayWithUnitsFeature(Chroma,
needs=fft,
frequency_band=band,
store=True)
bfcc = ArrayWithUnitsFeature(BFCC, needs=fft, store=True)
return AudioGraph
def test_has_correct_sample_rate(self):
half_lapped = HalfLapped()
synth = DCTSynthesizer()
raw = np.zeros((100, 2048))
band = FrequencyBand(0, SR44100().nyquist)
scale = LinearScale(band, raw.shape[1])
timeseries = ArrayWithUnits(
raw, [TimeDimension(*half_lapped), FrequencyDimension(scale)])
output = synth.synthesize(timeseries)
self.assertIsInstance(output.samplerate, SR44100)
self.assertIsInstance(output, AudioSamples)
def test_can_do_multithreaded_resampling(self):
synth = SilenceSynthesizer(SR44100())
audio = [synth.synthesize(Seconds(5)) for _ in xrange(10)]
pool = ThreadPool(4)
def x(samples):
rs = Resample(int(SR44100()), int(SR11025()))
return rs(samples, end_of_input=True)
resampled = pool.map(x, audio)
self.assertEqual(10, len(resampled))
def test_can_get_all_even_sized_bands(self):
samplerate = SR44100()
scale = LinearScale.from_sample_rate(samplerate,
44100,
always_even=True)
log_scale = GeometricScale(20, 20000, 0.01, 64)
slices = [scale.get_slice(band) for band in log_scale]
sizes = [s.stop - s.start for s in slices]
self.assertTrue(
not any([s % 2 for s in sizes]),
'All slice sizes should be even but were {sizes}'.format(
**locals()))
def setUp(self):
@simple_in_memory_settings
class Document(stft(store_fft=True)):
pass
synth = NoiseSynthesizer(SR44100())
audio = synth.synthesize(Seconds(2))
_id = Document.process(meta=audio.encode())
doc = Document(_id)
non_doc = SomethingElse(11)
parser = FeatureParser(Document, locals())
self.document = Document
self.doc = doc
self.parser = parser
def windowed(wscheme,
chunksize_bytes=DEFAULT_CHUNK_SIZE,
resample_to=SR44100(),
store_resampled=True,
store_windowed=False,
wfunc=None):
rs = resampled(chunksize_bytes=chunksize_bytes,
resample_to=resample_to,
store_resampled=store_resampled)
class Sound(rs):
windowed = ArrayWithUnitsFeature(SlidingWindow,
wscheme=wscheme,
wfunc=wfunc,
needs=rs.resampled,
store=store_windowed)
return Sound
def test_can_encode_and_decode_variable_rate_time_Series(self):
class TimestampEmitter(ff.Node):
def __init__(self, needs=None):
super(TimestampEmitter, self).__init__(needs=needs)
self.pos = Picoseconds(0)
def _process(self, data):
td = data.dimensions[0]
frequency = td.frequency
timestamps = [self.pos + (i * frequency)
for i, d in enumerate(data)
if random() > 0.9]
slices = TimeSlice.slices(timestamps)
yield VariableRateTimeSeries(
(ts, np.zeros(0)) for ts in slices)
self.pos += frequency * len(data)
graph = stft(store_fft=True)
@simple_in_memory_settings
class Document(graph):
slices = TimeSliceFeature(
TimestampEmitter,
needs=graph.fft,
store=True)
pooled = VariableRateTimeSeriesFeature(
Pooled,
op=np.max,
axis=0,
needs=(slices, graph.fft),
store=False)
signal = NoiseSynthesizer(SR44100())\
.synthesize(Seconds(10))\
.encode()
_id = Document.process(meta=signal)
doc = Document(_id)
self.assertIsInstance(doc.pooled, VariableRateTimeSeries)
self.assertEqual(doc.fft.shape[1], doc.pooled.slicedata.shape[1])
def stft(chunksize_bytes=DEFAULT_CHUNK_SIZE,
resample_to=SR44100(),
wscheme=HalfLapped(),
store_fft=False,
fft_padding_samples=None,
store_windowed=False,
store_resampled=False):
class ShortTimeFourierTransform(BaseModel):
meta = JSONFeature(MetaData, store=True, encoder=AudioMetaDataEncoder)
raw = ByteStreamFeature(ByteStream,
chunksize=chunksize_bytes,
needs=meta,
store=False)
ogg = OggVorbisFeature(OggVorbis, needs=raw, store=True)
pcm = AudioSamplesFeature(AudioStream, needs=raw, store=False)
resampled = AudioSamplesFeature(Resampler,
needs=pcm,
samplerate=resample_to,
store=store_resampled)
windowed = ArrayWithUnitsFeature(SlidingWindow,
needs=resampled,
wscheme=wscheme,
wfunc=OggVorbisWindowingFunc(),
store=store_windowed)
fft = ArrayWithUnitsFeature(FFT,
padding_samples=fft_padding_samples,
needs=windowed,
store=store_fft)
return ShortTimeFourierTransform
def __init__(self, path):
super(MusicNet, self).__init__()
self.path = path
self._metadata = \
'https://homes.cs.washington.edu/~thickstn/media/musicnet_metadata.csv'
self._samplerate = SR44100()
import numpy as np
from featureflow import BaseModel, JSONFeature, ByteStream, ByteStreamFeature
from zounds.soundfile import \
MetaData, AudioMetaDataEncoder, OggVorbis, OggVorbisFeature, AudioStream, \
Resampler, ChunkSizeBytes
from zounds.segment import \
ComplexDomain, MovingAveragePeakPicker, TimeSliceFeature
from zounds.persistence import ArrayWithUnitsFeature, AudioSamplesFeature, \
FrequencyAdaptiveFeature
from zounds.timeseries import SR44100, HalfLapped, Stride, Seconds
from zounds.spectral import \
SlidingWindow, OggVorbisWindowingFunc, FFT, BarkBands, SpectralCentroid, \
Chroma, BFCC, DCT, FrequencyAdaptiveTransform, FrequencyBand
DEFAULT_CHUNK_SIZE = ChunkSizeBytes(samplerate=SR44100(),
duration=Seconds(30),
bit_depth=16,
channels=2)
def resampled(chunksize_bytes=DEFAULT_CHUNK_SIZE,
resample_to=SR44100(),
store_resampled=False):
"""
Create a basic processing pipeline that can resample all incoming audio
to a normalized sampling rate for downstream processing, and store a
convenient, compressed version for playback
:param chunksize_bytes: The number of bytes from the raw stream to process
at once
:param resample_to: The new, normalized sampling rate
def __init__(self, samplerate=None, needs=None):
super(Resampler, self).__init__(needs=needs)
self._samplerate = samplerate or SR44100()
self._resample = None
def test_can_invert_fft_44100(self):
self.can_invert_fft(SR44100())
def test_correct_output_with_stereo(self):
synth = SilenceSynthesizer(SR44100())
samples = synth.synthesize(Seconds(1)).stereo
rs = Resample(int(samples.samplerate), int(SR11025()), nchannels=2)
resampled = rs(samples, end_of_input=True)
self.assertEqual((11025, 2), resampled.shape)
def test_audible_range_lower_bound(self):
band = FrequencyBand.audible_range(SR44100())
self.assertEqual(20, band.start_hz)
def test_can_convert_to_integer_number_of_bytes(self):
cs = ChunkSizeBytes(SR44100(), Seconds(30), channels=2, bit_depth=16)
self.assertEqual(5292000, int(cs))
def test_correct_window_and_step_size_at_44100(self):
self._check(SR44100(), 2048, 1024)
def test_audible_range_upper_bound(self):
sr = SR44100()
band = FrequencyBand.audible_range(sr)
self.assertEqual(int(sr) // 2, band.stop_hz)
def soundfile(flo=None):
synth = NoiseSynthesizer(SR44100())
samples = synth.synthesize(Seconds(5)).stereo
flo = samples.encode(flo=flo)
return samples, flo
from soundfile import SoundFile
from zounds.timeseries import TimeSlice, AudioSamples, SR44100, HalfLapped, \
Seconds, Milliseconds, Stride
from zounds.persistence import ArrayWithUnitsFeature, AudioSamplesFeature
from zounds.soundfile import \
AudioStream, OggVorbis, OggVorbisFeature, Resampler
from zounds.spectral import \
SlidingWindow, OggVorbisWindowingFunc, FFT, Chroma, BarkBands, BFCC, \
FrequencyBand
from zounds.basic import Max
from zounds.util import simple_in_memory_settings
from featureflow import *
windowing_scheme = HalfLapped()
samplerate = SR44100()
band = FrequencyBand(20, samplerate.nyquist)
@simple_in_memory_settings
class Document(BaseModel):
raw = ByteStreamFeature(
ByteStream,
chunksize=2 * 44100 * 30 * 2,
store=True)
ogg = OggVorbisFeature(
OggVorbis,
needs=raw,
store=True)
def test_generates_correct_samplerate(self):
ss = SineSynthesizer(SR44100())
audio = ss.synthesize(Seconds(4), freqs_in_hz=[440.])
self.assertEqual(SR44100(), audio.samplerate)
def setUp(self):
self.samplerate = SR44100()
self.wscheme = HalfLapped()
self.STFT = stft(store_fft=True,
resample_to=self.samplerate,
wscheme=self.wscheme)
def x(samples):
rs = Resample(int(SR44100()), int(SR11025()))
return rs(samples, end_of_input=True)