Showing the changes to python (1 of 1 file). Show all
Unified · Split
| @@ -0,0 +1,110 @@ | |||
| 1 | #! /usr/bin/env python | ||
| 2 | |||
| 3 | # Implementation of the timescale algorithm according to Dan Ellis, *A Phase | ||
| 4 | # Vocoder in Matlab*. http://www.ee.columbia.edu/~dpwe/resources/matlab/pvoc/ | ||
| 5 | |||
| 6 | # This file follows the original implementation, with analysis in a first pass, | ||
| 7 | # and synthesis in a second pass. | ||
| 8 | |||
| 9 | import sys | ||
| 10 | from aubio import source, sink, pvoc, mfcc, cvec | ||
| 11 | from aubio import unwrap2pi, float_type | ||
| 12 | import numpy as np | ||
| 13 | |||
| 14 | win_s = 1024 | ||
| 15 | hop_s = win_s / 8 # 87.5 % overlap | ||
| 16 | |||
| 17 | warmup = win_s // hop_s - 1 | ||
| 18 | |||
| 19 | if len(sys.argv) < 3: | ||
| 20 | print("Usage: %s <source_filename> <output_filename> <rate> [samplerate]".format(sys.argv[0])) | ||
| 21 | print("""Examples: | ||
| 22 | # twice faster | ||
| 23 | {0} track_01.mp3 track_01_faster.wav 2.0 | ||
| 24 | # twice slower | ||
| 25 | {0} track_02.flac track_02_slower.wav 0.5 | ||
| 26 | # one and a half time faster, resampling first the input to 22050 | ||
| 27 | {0} track_02.flac track_02_slower.wav 1.5 22050""".format(sys.argv[0])) | ||
| 28 | sys.exit(1) | ||
| 29 | |||
| 30 | source_filename = sys.argv[1] | ||
| 31 | output_filename = sys.argv[2] | ||
| 32 | rate = float(sys.argv[3]) | ||
| 33 | |||
| 34 | samplerate = 0 if len(sys.argv) < 5 else int(sys.argv[4]) | ||
| 35 | source_in = source(source_filename, samplerate, hop_s) | ||
| 36 | samplerate = source_in.samplerate | ||
| 37 | p = pvoc(win_s, hop_s) | ||
| 38 | |||
| 39 | # allocate memory to store norms and phases | ||
| 40 | n_blocks = source_in.duration // hop_s + 1 | ||
| 41 | # adding an empty frame at end of spectrogram | ||
| 42 | norms = np.zeros((n_blocks + 1, win_s // 2 + 1), dtype = float_type) | ||
| 43 | phases = np.zeros((n_blocks + 1, win_s // 2 + 1), dtype = float_type) | ||
| 44 | |||
| 45 | block_read = 0 | ||
| 46 | while True: | ||
| 47 | # read from source | ||
| 48 | samples, read = source_in() | ||
| 49 | # compute fftgrain | ||
| 50 | spec = p(samples) | ||
| 51 | # store current grain | ||
| 52 | norms[block_read] = spec.norm | ||
| 53 | phases[block_read] = spec.phas | ||
| 54 | # until end of file | ||
| 55 | if read < hop_s: break | ||
| 56 | # increment block counter | ||
| 57 | block_read += 1 | ||
| 58 | |||
| 59 | # just to make sure | ||
| 60 | #source_in.close() | ||
| 61 | |||
| 62 | sink_out = sink(output_filename, samplerate) | ||
| 63 | |||
| 64 | # interpolated time steps (j = alpha * i) | ||
| 65 | steps = np.arange(0, n_blocks, rate, dtype = float_type) | ||
| 66 | # initial phase | ||
| 67 | phas_acc = phases[0] | ||
| 68 | # excepted phase advance in each bin | ||
| 69 | phi_advance = np.linspace(0, np.pi * hop_s, win_s / 2 + 1).astype (float_type) | ||
| 70 | |||
| 71 | new_grain = cvec(win_s) | ||
| 72 | |||
| 73 | for (t, step) in enumerate(steps): | ||
| 74 | |||
| 75 | frac = 1. - np.mod(step, 1.0) | ||
| 76 | # get pair of frames | ||
| 77 | t_norms = norms[int(step):int(step+2)] | ||
| 78 | t_phases = phases[int(step):int(step+2)] | ||
| 79 | |||
| 80 | # compute interpolated frame | ||
| 81 | new_grain.norm = frac * t_norms[0] + (1. - frac) * t_norms[1] | ||
| 82 | new_grain.phas = phas_acc | ||
| 83 | #print t, step, new_grain.norm | ||
| 84 | #print t, step, phas_acc | ||
| 85 | |||
| 86 | # psola | ||
| 87 | samples = p.rdo(new_grain) | ||
| 88 | if t > warmup: # skip the first few frames to warm up phase vocoder | ||
| 89 | # write to sink | ||
| 90 | sink_out(samples, hop_s) | ||
| 91 | |||
| 92 | # calculate phase advance | ||
| 93 | dphas = t_phases[1] - t_phases[0] - phi_advance | ||
| 94 | # unwrap angle to [-pi; pi] | ||
| 95 | dphas = unwrap2pi(dphas) | ||
| 96 | # cumulate phase, to be used for next frame | ||
| 97 | phas_acc += phi_advance + dphas | ||
| 98 | |||
| 99 | for t in range(warmup + 1): # purge the last frames from the phase vocoder | ||
| 100 | new_grain.norm[:] = 0 | ||
| 101 | new_grain.phas[:] = 0 | ||
| 102 | samples = p.rdo(new_grain) | ||
| 103 | sink_out(samples, read if t == warmup else hop_s) | ||
| 104 | |||
| 105 | # just to make sure | ||
| 106 | #sink_out.close() | ||
| 107 | |||
| 108 | format_out = "read {:d} blocks from {:s} at {:d}Hz and rate {:f}, wrote {:d} blocks to {:s}" | ||
| 109 | print (format_out.format(block_read, source_filename, samplerate, rate, | ||
| 110 | len(steps), output_filename)) | ||