Showing the changes to python (1 of 1 file). Show all
Unified · Split
| @@ -0,0 +1,143 @@ | ||
| 1 | +#! /usr/bin/env python | |
| 2 | + | |
| 3 | +""" | |
| 4 | +Script to evaluate pitch algorithms against TONAS database. | |
| 5 | + | |
| 6 | +See http://mtg.upf.edu/download/datasets/tonas/ | |
| 7 | + | |
| 8 | +Example run: | |
| 9 | + | |
| 10 | + $ ./eval_pitch /path/to/TONAS/*/*.wav | |
| 11 | + OK: 94.74% vx r: 96.87% vx f: 15.83% f0: 96.02% %12: 0.50% /path/to/TONAS/Deblas/01-D_AMairena.wav | |
| 12 | + OK: 89.89% vx r: 93.21% vx f: 13.81% f0: 90.74% %12: 1.51% /path/to/TONAS/Deblas/02-D_ChanoLobato.wav | |
| 13 | + OK: 96.02% vx r: 96.73% vx f: 10.91% f0: 96.42% %12: 0.00% /path/to/TONAS/Deblas/03-D_Chocolate.wav | |
| 14 | + [...] | |
| 15 | + OK: 82.35% vx r: 95.52% vx f: 67.09% f0: 89.80% %12: 0.95% /path/to/TONAS/Martinetes2/80-M2_Rancapinos.wav | |
| 16 | + OK: 61.97% vx r: 85.71% vx f: 22.03% f0: 55.63% %12: 8.57% /path/to/TONAS/Martinetes2/81-M2_SDonday.wav | |
| 17 | + OK: 75.26% vx r: 91.63% vx f: 27.27% f0: 75.99% %12: 5.05% /path/to/TONAS/Martinetes2/82-M2_TiaAnicalaPiriniaca.wav | |
| 18 | + OK: 82.77% vx r: 92.74% vx f: 38.27% f0: 87.33% %12: 1.67% 69 files, total_length: 1177.69s, total runtime: 25.91s | |
| 19 | + | |
| 20 | + | |
| 21 | +""" | |
| 22 | + | |
| 23 | +import sys | |
| 24 | +import time | |
| 25 | +import os.path | |
| 26 | +import numpy | |
| 27 | +from utils import array_from_text_file, array_from_yaml_file | |
| 28 | +from aubio import source, pitch, freqtomidi | |
| 29 | + | |
| 30 | +start = time.time() | |
| 31 | + | |
| 32 | +freq_tol = .50 # more or less half a tone | |
| 33 | + | |
| 34 | +methods = ["default", "yinfft", "mcomb", "yin", "fcomb", "schmitt", "specacf"] | |
| 35 | +method = methods[0] | |
| 36 | + | |
| 37 | +downsample = 1 | |
| 38 | +tolerance = 0.35 | |
| 39 | +silence = -40. | |
| 40 | +skip = 1 | |
| 41 | +if method in ["yinfft", "default"]: | |
| 42 | + downsample = 1 | |
| 43 | + tolerance = 0.45 | |
| 44 | +elif method == "mcomb": | |
| 45 | + downsample = 4 | |
| 46 | +elif method == "yin": | |
| 47 | + downsample = 4 | |
| 48 | + tolerance = 0.2 | |
| 49 | + | |
| 50 | +samplerate = 44100 / downsample | |
| 51 | +hop_s = 512 / downsample | |
| 52 | +win_s = 2048 / downsample | |
| 53 | + | |
| 54 | +def get_pitches (filename, samplerate = samplerate, win_s = win_s, hop_s = hop_s): | |
| 55 | + s = source(filename, samplerate, hop_s) | |
| 56 | + samplerate = s.samplerate | |
| 57 | + | |
| 58 | + p = pitch(method, win_s, hop_s, samplerate) | |
| 59 | + p.set_unit("freq") | |
| 60 | + p.set_tolerance(tolerance) | |
| 61 | + p.set_silence(silence) | |
| 62 | + | |
| 63 | + # list of pitches, in samples | |
| 64 | + pitches = [] | |
| 65 | + | |
| 66 | + # total number of frames read | |
| 67 | + total_frames = 0 | |
| 68 | + while True: | |
| 69 | + samples, read = s() | |
| 70 | + new_pitch = p(samples)[0] | |
| 71 | + pitches.append([total_frames/float(samplerate), new_pitch]) | |
| 72 | + total_frames += read | |
| 73 | + if read < hop_s: break | |
| 74 | + return numpy.array(pitches) | |
| 75 | + | |
| 76 | +total_correct_f0, total_correct_sil, total_missed, total_incorrect, total_fp, total_total = 0, 0, 0, 0, 0, 0 | |
| 77 | +total_correct_chroma, total_voiced = 0, 0 | |
| 78 | +for source_file in sys.argv[1:]: | |
| 79 | + ground_truth_file = source_file.replace('.wav', '.f0.Corrected') | |
| 80 | + if os.path.isfile(ground_truth_file): | |
| 81 | + ground_truth = array_from_text_file(ground_truth_file)[:,[0,2]] | |
| 82 | + experiment = get_pitches(source_file) | |
| 83 | + # check that we have the same length, more or less one frame | |
| 84 | + assert abs(len(ground_truth) - len(experiment)) < 2 | |
| 85 | + # align experiment by skipping first results | |
| 86 | + experiment = experiment[skip:] | |
| 87 | + experiment[:,0] -= experiment[0,0] | |
| 88 | + # trim to shortest list | |
| 89 | + maxlen = min(len(ground_truth), len(experiment)) | |
| 90 | + experiment = experiment[:maxlen] | |
| 91 | + ground_truth = ground_truth[:maxlen] | |
| 92 | + # get difference matrix | |
| 93 | + diffmat = abs(experiment - ground_truth) | |
| 94 | + # make sure we got the timing right | |
| 95 | + assert max(diffmat[:,0]) < 10e-4, source_file | |
| 96 | + truth_pitches = freqtomidi(ground_truth[:,1]) | |
| 97 | + exper_pitches = freqtomidi(experiment[:,1]) | |
| 98 | + | |
| 99 | + total = len(truth_pitches) | |
| 100 | + unvoiced = len(truth_pitches[truth_pitches == 0]) | |
| 101 | + voiced = total - unvoiced | |
| 102 | + correct_sil, fp, missed, correct_f0, correct_chroma, incorrect = 0, 0, 0, 0, 0, 0 | |
| 103 | + for a, b in zip(truth_pitches, exper_pitches): | |
| 104 | + if a == 0 and b == 0: | |
| 105 | + correct_sil += 1 | |
| 106 | + elif a == 0 and b != 0: | |
| 107 | + fp += 1 | |
| 108 | + elif a != 0 and b == 0: | |
| 109 | + missed += 1 | |
| 110 | + elif abs(b - a) < freq_tol: | |
| 111 | + correct_f0 += 1 | |
| 112 | + elif abs(b - a) % 12. < freq_tol: | |
| 113 | + correct_chroma += 1 | |
| 114 | + else: | |
| 115 | + incorrect += 1 | |
| 116 | + assert correct_sil + fp + missed + correct_f0 + correct_chroma + incorrect == total | |
| 117 | + assert unvoiced == correct_sil + fp | |
| 118 | + assert voiced == missed + correct_f0 + correct_chroma + incorrect | |
| 119 | + print "OK: %6s%%" % ("%.2f" % (100. * (correct_f0 + correct_sil) / total )), | |
| 120 | + print "vx r: %6s%%" % ("%.2f" % (100. - 100. * missed / voiced)), | |
| 121 | + print "vx f: %6s%%" % ("%.2f" % (100. * fp / unvoiced)), | |
| 122 | + print "f0: %6s%%" % ("%.2f" % (100. * correct_f0 / voiced)), | |
| 123 | + print "%%12: %6s%%" % ("%.2f" % (100. * correct_chroma / voiced)), | |
| 124 | + print source_file | |
| 125 | + total_correct_sil += correct_sil | |
| 126 | + total_correct_f0 += correct_f0 | |
| 127 | + total_correct_chroma += correct_chroma | |
| 128 | + total_missed += missed | |
| 129 | + total_incorrect += incorrect | |
| 130 | + total_fp += fp | |
| 131 | + total_voiced += voiced | |
| 132 | + total_total += total | |
| 133 | + else: | |
| 134 | + print "ERR", "could not find ground_truth_file", ground_truth_file | |
| 135 | + | |
| 136 | +print "OK: %6s%%" % ("%.2f" % (100. * (total_correct_f0 + total_correct_sil) / total_total )), | |
| 137 | +print "vx r: %6s%%" % ("%.2f" % (100. - 100. * total_missed / total_voiced)), | |
| 138 | +print "vx f: %6s%%" % ("%.2f" % (100. * (total_fp) / (total_correct_sil + total_fp))), | |
| 139 | +print "f0: %6s%%" % ("%.2f" % (100. * total_correct_f0 / total_voiced)), | |
| 140 | +print "%%12: %6s%%" % ("%.2f" % (100. * total_correct_chroma / total_voiced)), | |
| 141 | +print "%d files," % len(sys.argv[1:]), | |
| 142 | +print "total_length: %.2fs," % ((total_total * hop_s) / float(samplerate)), | |
| 143 | +print "total runtime: %.2fs" % (time.time() - start) | |