linearvc/speaker_similarity.py at main · kamperh/linearvc · GitHub

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
162
163
164
165
#!/usr/bin/env python

"""
Evaluate speaker similarity using EER (lower corresponds to better conversion).

Author: Benjamin van Niekerk, Herman Kamper
Date: 2023, 2024
"""

from pathlib import Path
from scipy.interpolate import interp1d
from scipy.optimize import brentq
from scipy.spatial.distance import cosine
from sklearn.metrics import roc_curve
from speechbrain.inference.speaker import EncoderClassifier
from tqdm import tqdm
import argparse
import numpy as np
import pandas
import sys
import torch
import torchaudio
import torchaudio.functional as F

device = "cuda"


def check_argv():
    parser = argparse.ArgumentParser(description=__doc__.strip().split("\n")[0])
    parser.add_argument(
        "eval_csv", type=str, help="evaluation CSV listing utterance pairs"
    )
    parser.add_argument(
        "converted_dir",
        type=Path,
        help="converted speech directory",
    )
    parser.add_argument(
        "groundtruth_dir",
        type=Path,
        help="real speech directory; for LibriSpeech, this would end in "
        "e.g. `dev-clean`",
    )
    parser.add_argument(
        "--format", choices=["librispeech", "vctk"], default="vctk"
    )
    parser.add_argument(
        "--zero_positive",
        help="in some cases you want the 0 label in the evaluation CSV "
        "to indicate a positive sample (instead of the usual 1 label)",
        action="store_true",
    )
    return parser.parse_args()


def eer(y, y_score):
    fpr, tpr, _ = roc_curve(y, 1 - y_score, pos_label=1)
    return brentq(lambda x: 1.0 - x - interp1d(fpr, tpr)(x), 0.0, 1.0)


def speaker_similarity(args):
    classifier = EncoderClassifier.from_hparams(
        source="speechbrain/spkrec-xvect-voxceleb",
        savedir="pretrained_models/spkrec-xvect-voxceleb",
        run_opts={"device": "cuda"},
    )

    pairs = pandas.read_csv(args.eval_csv)

    # print(pairs)

    converted_pairs = pairs[pairs.label == 0]
    groundtruth_paris = pairs[pairs.label == 1]

    # print(pairs)
    # assert False

    scores = []

    # Converted similarities
    for _, (
        source_speaker,
        target_speaker,
        source_key,
        target_key,
        label,
    ) in tqdm(list(converted_pairs.iterrows())):
        source_wav_fn = (args.converted_dir / source_key).with_suffix(".wav")
        if args.format == "librispeech":
            target_wav_fn = (args.groundtruth_dir / target_key).with_suffix(
                ".flac"
            )
        elif args.format == "vctk":
            target_wav_fn = (args.groundtruth_dir / target_key).with_suffix(
                ".wav"
            )

        x, sr = torchaudio.load(source_wav_fn)
        y, sr = torchaudio.load(target_wav_fn)
        x = F.resample(x, sr, 16000)
        y = F.resample(y, sr, 16000)
        x = x.to(device)
        y = y.to(device)

        x = classifier.encode_batch(x).squeeze().cpu().numpy()
        y = classifier.encode_batch(y).squeeze().cpu().numpy()

        if args.zero_positive:
            label = 1

        scores.append((source_speaker, target_speaker, cosine(x, y), label))

    # Ground truth similarities
    for _, (
        source_speaker,
        target_speaker,
        source_key,
        target_key,
        label,
    ) in tqdm(list(groundtruth_paris.iterrows())):
        if args.format == "librispeech":
            source_wav_fn = (args.groundtruth_dir / source_key).with_suffix(
                ".flac"
            )
            target_wav_fn = (args.groundtruth_dir / target_key).with_suffix(
                ".flac"
            )
        elif args.format == "vctk":
            source_wav_fn = (args.groundtruth_dir / source_key).with_suffix(
                ".wav"
            )
            target_wav_fn = (args.groundtruth_dir / target_key).with_suffix(
                ".wav"
            )

        x, sr = torchaudio.load(source_wav_fn)
        y, sr = torchaudio.load(target_wav_fn)
        x = F.resample(x, sr, 16000)
        y = F.resample(y, sr, 16000)
        x = x.to(device)
        y = y.to(device)

        x = classifier.encode_batch(x).squeeze().cpu().numpy()
        y = classifier.encode_batch(y).squeeze().cpu().numpy()

        if args.zero_positive:
            label = 0

        scores.append((source_speaker, target_speaker, cosine(x, y), label))

    scores = pandas.DataFrame(
        scores, columns=["src_speaker", "tgt_speaker", "score", "label"]
    )

    sim = (
        scores.groupby("tgt_speaker")
        .apply(lambda x: eer(x.label, x.score))
        .reset_index(name="eer")
    )
    print(sim.agg(mean=("eer", np.mean), std=("eer", np.std)))


if __name__ == "__main__":
    args = check_argv()
    speaker_similarity(args)