|
mi@0
|
1 # Copyright 2014 Jason Heeris, jason.heeris@gmail.com
|
|
mi@0
|
2 #
|
|
mi@0
|
3 # This file is part of the gammatone toolkit, and is licensed under the 3-clause
|
|
mi@0
|
4 # BSD license: https://github.com/detly/gammatone/blob/master/COPYING
|
|
mi@0
|
5 from __future__ import division
|
|
mi@0
|
6 import numpy as np
|
|
mi@0
|
7
|
|
mi@0
|
8 from .filters import make_erb_filters, centre_freqs, erb_filterbank
|
|
mi@0
|
9
|
|
mi@0
|
10 """
|
|
mi@0
|
11 This module contains functions for rendering "spectrograms" which use gammatone
|
|
mi@0
|
12 filterbanks instead of Fourier transforms.
|
|
mi@0
|
13 """
|
|
mi@0
|
14 '''
|
|
mi@0
|
15 Modified from the original toolbox. Jan 2015. --Mi
|
|
mi@0
|
16
|
|
mi@0
|
17 '''
|
|
mi@0
|
18
|
|
mi@0
|
19 def round_half_away_from_zero(num):
|
|
mi@0
|
20 """ Implement the round-half-away-from-zero rule, where fractional parts of
|
|
mi@0
|
21 0.5 result in rounding up to the nearest positive integer for positive
|
|
mi@0
|
22 numbers, and down to the nearest negative number for negative integers.
|
|
mi@0
|
23 """
|
|
mi@0
|
24 return np.sign(num) * np.floor(np.abs(num) + 0.5)
|
|
mi@0
|
25
|
|
mi@0
|
26
|
|
mi@0
|
27 def gtgram_strides(fs, gammatoneLen, step_rate, filterbank_cols):
|
|
mi@0
|
28 """
|
|
mi@0
|
29 Calculates the window size for a gammatonegram.
|
|
mi@0
|
30
|
|
mi@0
|
31 @return a tuple of (window_size, hop_samples, output_columns)
|
|
mi@0
|
32 """
|
|
mi@0
|
33 step_samples = int(gammatoneLen * step_rate)
|
|
mi@0
|
34 columns = (1 + int(np.floor((filterbank_cols - gammatoneLen)/ step_samples)))
|
|
mi@0
|
35
|
|
mi@0
|
36 return (step_samples, columns)
|
|
mi@0
|
37
|
|
mi@0
|
38
|
|
mi@0
|
39 def gtgram_xe(wave, fs, f_max, channels, f_min):
|
|
mi@0
|
40 """ Calculate the intermediate ERB filterbank processed matrix """
|
|
mi@0
|
41 cfs = centre_freqs(f_max, channels, f_min)
|
|
mi@0
|
42 fcoefs = np.flipud(make_erb_filters(fs, cfs))
|
|
mi@0
|
43 xf = erb_filterbank(wave, fcoefs)
|
|
mi@0
|
44 return xf
|
|
mi@0
|
45 # xe = np.power(xf, 2)
|
|
mi@0
|
46 # return xe
|
|
mi@0
|
47
|
|
mi@0
|
48
|
|
mi@0
|
49 def gtgram(wave, fs, gammatoneLen, step_rate, channels, f_max, f_min):
|
|
mi@0
|
50 """
|
|
mi@0
|
51 Calculate a spectrogram-like time frequency magnitude array based on
|
|
mi@0
|
52 gammatone subband filters. The waveform ``wave`` (at sample rate ``fs``) is
|
|
mi@0
|
53 passed through an multi-channel gammatone auditory model filterbank, with
|
|
mi@0
|
54 lowest frequency ``f_min`` and highest frequency ``f_max``. The outputs of
|
|
mi@0
|
55 each band then have their energy integrated over windows of ``window_time``
|
|
mi@0
|
56 seconds, advancing by ``hop_time`` secs for successive columns. These
|
|
mi@0
|
57 magnitudes are returned as a nonnegative real matrix with ``channels`` rows.
|
|
mi@0
|
58
|
|
mi@0
|
59 | 2009-02-23 Dan Ellis dpwe@ee.columbia.edu
|
|
mi@0
|
60 |
|
|
mi@0
|
61 | (c) 2013 Jason Heeris (Python implementation)
|
|
mi@0
|
62 """
|
|
mi@0
|
63 xe = gtgram_xe(wave, fs, f_max, channels, f_min)
|
|
mi@0
|
64 # print 'xe', xe.shape
|
|
mi@0
|
65 step_samples, ncols = gtgram_strides(fs, gammatoneLen, step_rate, xe.shape[1])
|
|
mi@0
|
66 # print gammatoneLen, step_samples, channels, ncols
|
|
mi@0
|
67 y = np.zeros((channels, ncols))
|
|
mi@0
|
68
|
|
mi@0
|
69 for cnum in range(ncols):
|
|
mi@0
|
70 segment = xe[:, cnum * step_samples + np.arange(gammatoneLen)]
|
|
mi@0
|
71 segment = np.power(segment, 2)
|
|
mi@0
|
72 y[:, cnum] = np.sqrt(segment.mean(1))
|
|
mi@0
|
73
|
|
mi@0
|
74 return y.T
|
