shithub: aubio

ref: 89bc6f4938d3c1e41e351d7cd8314ed3fb3cd433
dir: /python/tests/eval_pitch/

View raw version
#! /usr/bin/env python

"""
Script to evaluate pitch algorithms against TONAS database.

See http://mtg.upf.edu/download/datasets/tonas/

Example run:

    $ ./eval_pitch /path/to/TONAS/*/*.wav
    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
    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
    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
    [...]
    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
    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
    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
    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


"""

import sys
import time
import os.path
import numpy
from .utils import array_from_text_file, array_from_yaml_file
from aubio import source, pitch, freqtomidi

start = time.time()

freq_tol = .50 # more or less half a tone

methods = ["default", "yinfft", "mcomb", "yin", "fcomb", "schmitt", "specacf"]
method = methods[0]

downsample = 1
tolerance =  0.35
silence = -40.
skip = 1
if method in ["yinfft", "default"]:
    downsample = 1
    tolerance = 0.45
elif method == "mcomb":
    downsample = 4
elif method == "yin":
    downsample = 4
    tolerance = 0.2

samplerate = 44100 / downsample
hop_s = 512 / downsample
win_s = 2048 / downsample

def get_pitches (filename, samplerate = samplerate, win_s = win_s, hop_s = hop_s):
    s = source(filename, samplerate, hop_s)
    samplerate = s.samplerate

    p = pitch(method, win_s, hop_s, samplerate)
    p.set_unit("freq")
    p.set_tolerance(tolerance)
    p.set_silence(silence)

    # list of pitches, in samples
    pitches = []

    # total number of frames read
    total_frames = 0
    while True:
        samples, read = s()
        new_pitch = p(samples)[0]
        pitches.append([total_frames/float(samplerate), new_pitch])
        total_frames += read
        if read < hop_s: break
    return numpy.array(pitches)

total_correct_f0, total_correct_sil, total_missed, total_incorrect, total_fp, total_total = 0, 0, 0, 0, 0, 0
total_correct_chroma, total_voiced = 0, 0
for source_file in sys.argv[1:]:
    ground_truth_file = source_file.replace('.wav', '.f0.Corrected')
    if os.path.isfile(ground_truth_file):
        ground_truth = array_from_text_file(ground_truth_file)[:,[0,2]]
        experiment = get_pitches(source_file)
        # check that we have the same length, more or less one frame
        assert abs(len(ground_truth) - len(experiment)) < 2
        # align experiment by skipping first results
        experiment = experiment[skip:]
        experiment[:,0] -= experiment[0,0]
        # trim to shortest list
        maxlen = min(len(ground_truth), len(experiment))
        experiment = experiment[:maxlen]
        ground_truth = ground_truth[:maxlen]
        # get difference matrix
        diffmat = abs(experiment - ground_truth)
        # make sure we got the timing right
        assert max(diffmat[:,0]) < 10e-4, source_file
        truth_pitches = freqtomidi(ground_truth[:,1])
        exper_pitches = freqtomidi(experiment[:,1])

        total = len(truth_pitches)
        unvoiced = len(truth_pitches[truth_pitches == 0])
        voiced = total - unvoiced
        correct_sil, fp, missed, correct_f0, correct_chroma, incorrect = 0, 0, 0, 0, 0, 0
        for a, b in zip(truth_pitches, exper_pitches):
            if a == 0 and b == 0:
                correct_sil += 1
            elif a == 0 and b != 0:
                fp += 1
            elif a != 0 and b == 0:
                missed += 1
            elif abs(b - a) < freq_tol:
                correct_f0 += 1
            elif abs(b - a) % 12. < freq_tol:
                correct_chroma += 1
            else:
                incorrect += 1
        assert correct_sil + fp + missed + correct_f0 + correct_chroma + incorrect == total
        assert unvoiced == correct_sil + fp
        assert voiced == missed + correct_f0 + correct_chroma + incorrect
        print "OK: %6s%%" % ("%.2f" % (100. * (correct_f0 + correct_sil) / total )),
        print "vx r: %6s%%" % ("%.2f" % (100. - 100. * missed / voiced)),
        print "vx f: %6s%%" % ("%.2f" % (100. * fp / unvoiced)),
        print "f0: %6s%%" % ("%.2f" % (100. * correct_f0 / voiced)),
        print "%%12: %6s%%" % ("%.2f" % (100. * correct_chroma / voiced)),
        print source_file
        total_correct_sil += correct_sil
        total_correct_f0 += correct_f0
        total_correct_chroma += correct_chroma
        total_missed += missed
        total_incorrect += incorrect
        total_fp += fp
        total_voiced += voiced
        total_total += total
    else:
        print "ERR", "could not find ground_truth_file", ground_truth_file

print "OK: %6s%%" % ("%.2f" % (100. * (total_correct_f0 + total_correct_sil) / total_total )),
print "vx r: %6s%%" % ("%.2f" % (100. - 100. * total_missed / total_voiced)),
print "vx f: %6s%%" % ("%.2f" % (100. * (total_fp) / (total_correct_sil + total_fp))),
print "f0: %6s%%" % ("%.2f" % (100. * total_correct_f0 / total_voiced)),
print "%%12: %6s%%" % ("%.2f" % (100. * total_correct_chroma / total_voiced)),
print "%d files," % len(sys.argv[1:]),
print "total_length: %.2fs," % ((total_total * hop_s) / float(samplerate)),
print "total runtime: %.2fs" % (time.time() - start)