-
Notifications
You must be signed in to change notification settings - Fork 9
/
eval.py
161 lines (134 loc) · 7.25 KB
/
eval.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
import os
import glob
import pickle
from pathlib import Path
import argparse
import torchaudio
import numpy as np
# FFE metrics
import amfm_decompy.pYAAPT as pYAAPT
import amfm_decompy.basic_tools as basic
from librosa.util import normalize
import textgrid
# ASR metrics
import string
import whisper
import editdistance as ed
# Earth movers distance
from scipy.stats import wasserstein_distance as emd
from utils import interp
def get_yaapt(audio):
frame_length = 20.0 # ms
to_pad = int(frame_length / 1000 * 16000) // 2
audio = np.pad(audio, (to_pad, to_pad), "constant", constant_values=0)
audio = normalize(audio) * 0.95
audio = basic.SignalObj(audio, 16000)
pitch = pYAAPT.yaapt(audio, frame_length=frame_length, nccf_thresh1=0.25, frame_space=0.005 * 1000, tda_frame_length=25.0)
return pitch.samp_values
def calc_asr_er(ref, pred):
int_dict = {'0': 'zero', '1': 'one', '2': 'two', '3': 'three', '4': 'four', '5': 'five', '6': 'six', '7': 'seven',
'8': 'eight', '9': 'nine'}
gt_text = ref.lower().strip().translate(str.maketrans('', '', string.punctuation))
ref_w = gt_text.split()
ref_c = list(' '.join(ref_w))
pred_text = pred.lower().strip().translate(str.maketrans('', '', string.punctuation))
for k, v in int_dict.items():
pred_text = pred_text.replace(k, v)
syn_w = pred_text.split()
syn_c = list(' '.join(syn_w))
return ed.eval(syn_c, ref_c), len(ref_c), ed.eval(syn_w, ref_w), len(ref_w)
def aligned_ffe(int1, int2, pitch1, pitch2, sr=16000):
ffe = []
for i in range(len(int1)):
ref = pitch1[int(int1[i].minTime * sr * 0.005 * 2.5 + 2):int(int1[i].maxTime * sr * 0.005 * 2.5 + 2)]
syn = pitch2[int(int2[i].minTime * sr * 0.005 * 2.5 + 2):int(int2[i].maxTime * sr * 0.005 * 2.5 + 2)]
syn = interp(syn, ref.shape[0])
ffe.append((np.abs(((ref + 0.0001)/(syn + 0.0001))-1) > 0.2).mean())
return np.mean(ffe)
def calc_errors(asr_model, args):
gt_path = f'{args.base_path}/orig/'
gen_path = f'{args.base_path}/{args.method}/'
err_dict = {'wer_s': 0, 'wer_d': 0, 'cer_s': 0, 'cer_d': 0, 'len': [], 'emd': [], 'w_ffe': [], 'w_len': [],
'p_ffe': [], 'p_len': []}
for trg in args.target_speakers:
print(f'--- speaker {trg} -----')
for f in glob.glob(f'{gen_path}/{trg}/*.wav'):
if trg in f.split('/')[-1]: # don't evaluate reconstruction
continue
seq = f.split('_')[-1].split('.')[0]
if not os.path.isfile(f'{gt_path}/{trg}_{seq}.wav'):
print('No reference recording: ', f'{trg}_{seq}.wav')
continue
path = Path(f)
if path.stem.split('_')[0] == 'p270' and seq == '024':
print('p270_024 is a problematic sample where content varies notably!')
continue
yref, sr = torchaudio.load(f'{gt_path}/{trg}_{seq}.wav')
ysyn, _sr = torchaudio.load(f)
yref, ysyn = yref[0], ysyn[0]
assert sr == _sr, f"{sr} != {_sr}"
# Length Error
err_dict['len'].append(abs(len(yref) - len(ysyn)))
# ASR metrics
ref_text = open(f'{os.path.splitext(f)[0]}.txt', 'r').readline()
pred_text = asr_model.transcribe(f)['text']
res = calc_asr_er(ref_text, pred_text)
err_dict['cer_s'], err_dict['cer_d'], err_dict['wer_s'], err_dict['wer_d'] = err_dict['cer_s'] + res[0], err_dict['cer_d'] + res[1], err_dict['wer_s'] + res[2], err_dict['wer_d'] + res[3]
# Earth Movers Distance
syn_pitch = get_yaapt(ysyn.numpy())
ref_pitch = get_yaapt(yref.numpy())
if ref_pitch.shape[0] > syn_pitch.shape[0]: # Make pitch seqs the same length
syn_pitch = np.pad(syn_pitch, (0, ref_pitch.shape[0] - syn_pitch.shape[0]), constant_values=0)
elif yref.shape[0] < ysyn.shape[0]:
ref_pitch = np.pad(ref_pitch, (0, syn_pitch.shape[0] - ref_pitch.shape[0]), constant_values=0)
err_dict['emd'].append(emd(syn_pitch, ref_pitch))
# Forced alignment metrics
ref_grid = textgrid.TextGrid.fromFile(f'{gt_path}/txtgrid/{trg}_{seq}.TextGrid')
syn_grid = None # If the content is corrupted by the conversion MFA doen't manage to align
if os.path.isfile(path.parent / f'txtgrid/{path.stem}.TextGrid'):
syn_grid = textgrid.TextGrid.fromFile(path.parent / f'txtgrid/{path.stem}.TextGrid')
try:
phone_ref_grid = [f for f in ref_grid[1] if f.mark]
if syn_grid:
phone_syn_grid = [f for f in syn_grid[1] if f.mark]
else:
phone_syn_grid = [textgrid.Interval(ref_grid.maxTime / (len(ref_grid[1]) + 1) * i, ref_grid.maxTime / (len(ref_grid[1]) + 1) * (i + 1), inv.mark) for i, inv in enumerate(ref_grid[1]) if inv.mark]
err_dict['p_len'].append(np.abs((np.array([i.duration() for i in phone_ref_grid]) - np.array(
[i.duration() for i in phone_syn_grid]))).mean())
err_dict['p_ffe'].append(aligned_ffe(phone_ref_grid, phone_syn_grid, ref_pitch, syn_pitch, sr))
except ValueError:
pass
try:
word_ref_grid = [f for f in ref_grid[0] if f.mark]
if syn_grid:
word_syn_grid = [f for f in syn_grid[0] if f.mark]
else:
word_syn_grid = [textgrid.Interval(ref_grid.maxTime / (len(ref_grid[0]) + 1) * i, ref_grid.maxTime / (len(ref_grid[0]) + 1) * (i + 1), inv.mark) for i, inv in enumerate(ref_grid[0]) if inv.mark]
err_dict['w_len'].append(np.abs((np.array([i.duration() for i in word_ref_grid]) - np.array([i.duration() for i in word_syn_grid]))).mean())
err_dict['w_ffe'].append(aligned_ffe(word_ref_grid, word_syn_grid, ref_pitch, syn_pitch, sr))
except ValueError:
pass
return err_dict
def log_results(err_dict, args, sr=16000):
with open(f'{args.base_path}/{args.method}_results.pkl', 'wb') as f:
pickle.dump(err_dict, f)
print('WER: ', err_dict['wer_s'] / err_dict['wer_d'])
print('CER: ', err_dict['cer_s'] / err_dict['cer_d'])
print('EMD: ', np.mean(err_dict['emd']))
print('Len Error: ', np.mean(err_dict['len']) / sr)
print('Word Len Error: ', np.mean(err_dict['w_len']))
print('Char Len Error: ', np.mean(err_dict['p_len']))
print('Word FFE: ', np.mean(err_dict['w_ffe']))
print('Character FFE: ', np.mean(err_dict['p_ffe']))
if __name__ == '__main__':
parser = argparse.ArgumentParser()
parser.add_argument('--base_path', default='../results/vctk/', help='Base path to all conversion methods')
parser.add_argument('--method', default='sr', help='Name of conversion type, as in folder name')
parser.add_argument('--device', default='cuda:0', help='Torch device')
parser.add_argument('--target_speakers', nargs='+', default=['p231', 'p239', 'p245', 'p270'], help='Target speakers for VC. If none random speakers are used')
args = parser.parse_args()
model = whisper.load_model("medium.en") # used for ASR metrics
model.eval()
model = model.to(args.device)
errs = calc_errors(model, args)
log_results(errs, args)