Skip to content

Navigation Menu

Sign in
Appearance settings

Search code, repositories, users, issues, pull requests...

Provide feedback

We read every piece of feedback, and take your input very seriously.

Saved searches

Use saved searches to filter your results more quickly
Appearance settings

Commit bcfd691

Browse filesBrowse files
hshteingarthshteingart
committed
add stride trick to framesig and write corresponding tests.
Running time is reduces by a factor of ~4 for a long signal
1 parent a350d7a commit bcfd691
Copy full SHA for bcfd691

2 files changed

+80-31Lines changed: 80 additions & 31 deletions

File tree

Expand file treeCollapse file tree
Open diff view settings
Filter options
Expand file treeCollapse file tree
Open diff view settings
Collapse file

‎python_speech_features/sigproc.py‎

Copy file name to clipboardExpand all lines: python_speech_features/sigproc.py
+49-31Lines changed: 49 additions & 31 deletions
Original file line numberDiff line numberDiff line change
@@ -6,11 +6,19 @@
66
import math
77
import logging
88

9+
910
def round_half_up(number):
1011
return int(decimal.Decimal(number).quantize(decimal.Decimal('1'), rounding=decimal.ROUND_HALF_UP))
1112

1213

13-
def framesig(sig,frame_len,frame_step,winfunc=lambda x:numpy.ones((x,))):
14+
def rolling_window(a, window, step=1):
15+
# http://ellisvalentiner.com/post/2017-03-21-np-strides-trick
16+
shape = a.shape[:-1] + (a.shape[-1] - window + 1, window)
17+
strides = a.strides + (a.strides[-1],)
18+
return numpy.lib.stride_tricks.as_strided(a, shape=shape, strides=strides)[::step]
19+
20+
21+
def framesig(sig, frame_len, frame_step, winfunc=lambda x: numpy.ones((x,)), stride_trick=True):
1422
"""Frame a signal into overlapping frames.
1523
1624
:param sig: the audio signal to frame.
@@ -25,21 +33,26 @@ def framesig(sig,frame_len,frame_step,winfunc=lambda x:numpy.ones((x,))):
2533
if slen <= frame_len:
2634
numframes = 1
2735
else:
28-
numframes = 1 + int(math.ceil((1.0*slen - frame_len)/frame_step))
36+
numframes = 1 + int(math.ceil((1.0 * slen - frame_len) / frame_step))
2937

30-
padlen = int((numframes-1)*frame_step + frame_len)
38+
padlen = int((numframes - 1) * frame_step + frame_len)
3139

3240
zeros = numpy.zeros((padlen - slen,))
33-
padsignal = numpy.concatenate((sig,zeros))
41+
padsignal = numpy.concatenate((sig, zeros))
42+
if stride_trick:
43+
win = winfunc(frame_len)
44+
frames = rolling_window(padsignal, window=frame_len, step=frame_step)
45+
else:
46+
indices = numpy.tile(numpy.arange(0, frame_len), (numframes, 1)) + numpy.tile(
47+
numpy.arange(0, numframes * frame_step, frame_step), (frame_len, 1)).T
48+
indices = numpy.array(indices, dtype=numpy.int32)
49+
frames = padsignal[indices]
50+
win = numpy.tile(winfunc(frame_len), (numframes, 1))
3451

35-
indices = numpy.tile(numpy.arange(0,frame_len),(numframes,1)) + numpy.tile(numpy.arange(0,numframes*frame_step,frame_step),(frame_len,1)).T
36-
indices = numpy.array(indices,dtype=numpy.int32)
37-
frames = padsignal[indices]
38-
win = numpy.tile(winfunc(frame_len),(numframes,1))
39-
return frames*win
52+
return frames * win
4053

4154

42-
def deframesig(frames,siglen,frame_len,frame_step,winfunc=lambda x:numpy.ones((x,))):
55+
def deframesig(frames, siglen, frame_len, frame_step, winfunc=lambda x: numpy.ones((x,))):
4356
"""Does overlap-add procedure to undo the action of framesig.
4457
4558
:param frames: the array of frames.
@@ -54,68 +67,73 @@ def deframesig(frames,siglen,frame_len,frame_step,winfunc=lambda x:numpy.ones((x
5467
numframes = numpy.shape(frames)[0]
5568
assert numpy.shape(frames)[1] == frame_len, '"frames" matrix is wrong size, 2nd dim is not equal to frame_len'
5669

57-
indices = numpy.tile(numpy.arange(0,frame_len),(numframes,1)) + numpy.tile(numpy.arange(0,numframes*frame_step,frame_step),(frame_len,1)).T
58-
indices = numpy.array(indices,dtype=numpy.int32)
59-
padlen = (numframes-1)*frame_step + frame_len
70+
indices = numpy.tile(numpy.arange(0, frame_len), (numframes, 1)) + numpy.tile(
71+
numpy.arange(0, numframes * frame_step, frame_step), (frame_len, 1)).T
72+
indices = numpy.array(indices, dtype=numpy.int32)
73+
padlen = (numframes - 1) * frame_step + frame_len
6074

6175
if siglen <= 0: siglen = padlen
6276

6377
rec_signal = numpy.zeros((padlen,))
6478
window_correction = numpy.zeros((padlen,))
6579
win = winfunc(frame_len)
6680

67-
for i in range(0,numframes):
68-
window_correction[indices[i,:]] = window_correction[indices[i,:]] + win + 1e-15 #add a little bit so it is never zero
69-
rec_signal[indices[i,:]] = rec_signal[indices[i,:]] + frames[i,:]
81+
for i in range(0, numframes):
82+
window_correction[indices[i, :]] = window_correction[
83+
indices[i, :]] + win + 1e-15 # add a little bit so it is never zero
84+
rec_signal[indices[i, :]] = rec_signal[indices[i, :]] + frames[i, :]
7085

71-
rec_signal = rec_signal/window_correction
86+
rec_signal = rec_signal / window_correction
7287
return rec_signal[0:siglen]
7388

74-
def magspec(frames,NFFT):
89+
90+
def magspec(frames, NFFT):
7591
"""Compute the magnitude spectrum of each frame in frames. If frames is an NxD matrix, output will be Nx(NFFT/2+1).
7692
7793
:param frames: the array of frames. Each row is a frame.
7894
:param NFFT: the FFT length to use. If NFFT > frame_len, the frames are zero-padded.
7995
:returns: If frames is an NxD matrix, output will be Nx(NFFT/2+1). Each row will be the magnitude spectrum of the corresponding frame.
8096
"""
8197
if numpy.shape(frames)[1] > NFFT:
82-
logging.warn('frame length (%d) is greater than FFT size (%d), frame will be truncated. Increase NFFT to avoid.', numpy.shape(frames)[1], NFFT)
83-
complex_spec = numpy.fft.rfft(frames,NFFT)
98+
logging.warn(
99+
'frame length (%d) is greater than FFT size (%d), frame will be truncated. Increase NFFT to avoid.',
100+
numpy.shape(frames)[1], NFFT)
101+
complex_spec = numpy.fft.rfft(frames, NFFT)
84102
return numpy.absolute(complex_spec)
85103

86-
def powspec(frames,NFFT):
104+
105+
def powspec(frames, NFFT):
87106
"""Compute the power spectrum of each frame in frames. If frames is an NxD matrix, output will be Nx(NFFT/2+1).
88107
89108
:param frames: the array of frames. Each row is a frame.
90109
:param NFFT: the FFT length to use. If NFFT > frame_len, the frames are zero-padded.
91110
:returns: If frames is an NxD matrix, output will be Nx(NFFT/2+1). Each row will be the power spectrum of the corresponding frame.
92111
"""
93-
return 1.0/NFFT * numpy.square(magspec(frames,NFFT))
112+
return 1.0 / NFFT * numpy.square(magspec(frames, NFFT))
94113

95-
def logpowspec(frames,NFFT,norm=1):
114+
115+
def logpowspec(frames, NFFT, norm=1):
96116
"""Compute the log power spectrum of each frame in frames. If frames is an NxD matrix, output will be Nx(NFFT/2+1).
97117
98118
:param frames: the array of frames. Each row is a frame.
99119
:param NFFT: the FFT length to use. If NFFT > frame_len, the frames are zero-padded.
100120
:param norm: If norm=1, the log power spectrum is normalised so that the max value (across all frames) is 0.
101121
:returns: If frames is an NxD matrix, output will be Nx(NFFT/2+1). Each row will be the log power spectrum of the corresponding frame.
102122
"""
103-
ps = powspec(frames,NFFT);
104-
ps[ps<=1e-30] = 1e-30
105-
lps = 10*numpy.log10(ps)
123+
ps = powspec(frames, NFFT);
124+
ps[ps <= 1e-30] = 1e-30
125+
lps = 10 * numpy.log10(ps)
106126
if norm:
107127
return lps - numpy.max(lps)
108128
else:
109129
return lps
110130

111-
def preemphasis(signal,coeff=0.95):
131+
132+
def preemphasis(signal, coeff=0.95):
112133
"""perform preemphasis on the input signal.
113134
114135
:param signal: The signal to filter.
115136
:param coeff: The preemphasis coefficient. 0 is no filter, default is 0.95.
116137
:returns: the filtered signal.
117138
"""
118-
return numpy.append(signal[0],signal[1:]-coeff*signal[:-1])
119-
120-
121-
139+
return numpy.append(signal[0], signal[1:] - coeff * signal[:-1])
Collapse file

‎test/test_sigproc.py‎

Copy file name to clipboard
+31Lines changed: 31 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,31 @@
1+
from python_speech_features import sigproc
2+
import unittest
3+
import numpy as np
4+
import time
5+
6+
7+
class test_case(unittest.TestCase):
8+
def test_frame_sig(self):
9+
n = 10000124
10+
frame_len = 37
11+
frame_step = 13
12+
x = np.random.rand(n)
13+
t0 = time.time()
14+
y_old = sigproc.framesig(x, frame_len=frame_len, frame_step=frame_step, stride_trick=False)
15+
t1 = time.time()
16+
y_new = sigproc.framesig(x, frame_len=frame_len, frame_step=frame_step, stride_trick=True)
17+
t_new = time.time() - t1
18+
t_old = t1 - t0
19+
self.assertTupleEqual(y_old.shape, y_new.shape)
20+
np.testing.assert_array_equal(y_old, y_new)
21+
self.assertLess(t_new, t_old)
22+
print('new run time %3.2f < %3.2f sec' % (t_new, t_old))
23+
24+
def test_rolling(self):
25+
x = np.arange(10)
26+
y = sigproc.rolling_window(x, window=4, step=3)
27+
y_expected = np.array([[0, 1, 2, 3],
28+
[3, 4, 5, 6],
29+
[6, 7, 8, 9]]
30+
)
31+
y = np.testing.assert_array_equal(y, y_expected)

0 commit comments

Comments
0 (0)
Morty Proxy This is a proxified and sanitized view of the page, visit original site.