File size: 4,802 Bytes
0883aa1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
# Copyright (c) 2023 Amphion.
#
# This source code is licensed under the MIT license found in the
# LICENSE file in the root directory of this source tree.

import os
import json
import torchaudio
from tqdm import tqdm
from glob import glob
from collections import defaultdict

from utils.util import has_existed


def libritts_statistics(data_dir):
    speakers = []
    distribution2speakers2pharases2utts = defaultdict(
        lambda: defaultdict(lambda: defaultdict(list))
    )

    distribution_infos = glob(data_dir + "/*")

    for distribution_info in distribution_infos:
        distribution = distribution_info.split("/")[-1]
        print(distribution)

        speaker_infos = glob(distribution_info + "/*")

        if len(speaker_infos) == 0:
            continue

        for speaker_info in speaker_infos:
            speaker = speaker_info.split("/")[-1]

            speakers.append(speaker)

            pharase_infos = glob(speaker_info + "/*")

            for pharase_info in pharase_infos:
                pharase = pharase_info.split("/")[-1]

                utts = glob(pharase_info + "/*.wav")

                for utt in utts:
                    uid = utt.split("/")[-1].split(".")[0]
                    distribution2speakers2pharases2utts[distribution][speaker][
                        pharase
                    ].append(uid)

    unique_speakers = list(set(speakers))
    unique_speakers.sort()

    print("Speakers: \n{}".format("\t".join(unique_speakers)))
    return distribution2speakers2pharases2utts, unique_speakers


def main(output_path, dataset_path):
    print("-" * 10)
    print("Preparing samples for libritts...\n")

    save_dir = os.path.join(output_path, "libritts")
    os.makedirs(save_dir, exist_ok=True)
    train_output_file = os.path.join(save_dir, "train.json")
    test_output_file = os.path.join(save_dir, "test.json")
    singer_dict_file = os.path.join(save_dir, "singers.json")
    utt2singer_file = os.path.join(save_dir, "utt2singer")
    if has_existed(train_output_file):
        return
    utt2singer = open(utt2singer_file, "w")

    # Load
    libritts_path = dataset_path

    distribution2speakers2pharases2utts, unique_speakers = libritts_statistics(
        libritts_path
    )

    # We select pharases of standard spekaer as test songs
    train = []
    test = []

    train_index_count = 0
    test_index_count = 0

    train_total_duration = 0
    test_total_duration = 0

    for distribution, speakers2pharases2utts in tqdm(
        distribution2speakers2pharases2utts.items()
    ):
        for speaker, pharases2utts in tqdm(speakers2pharases2utts.items()):
            pharase_names = list(pharases2utts.keys())

            for chosen_pharase in pharase_names:
                for chosen_uid in pharases2utts[chosen_pharase]:
                    res = {
                        "Dataset": "libritts",
                        "Singer": speaker,
                        "Uid": "{}#{}#{}#{}".format(
                            distribution, speaker, chosen_pharase, chosen_uid
                        ),
                    }
                    res["Path"] = "{}/{}/{}/{}.wav".format(
                        distribution, speaker, chosen_pharase, chosen_uid
                    )
                    res["Path"] = os.path.join(libritts_path, res["Path"])
                    assert os.path.exists(res["Path"])

                    waveform, sample_rate = torchaudio.load(res["Path"])
                    duration = waveform.size(-1) / sample_rate
                    res["Duration"] = duration

                    if not "train" in distribution:
                        res["index"] = test_index_count
                        test_total_duration += duration
                        test.append(res)
                        test_index_count += 1
                    else:
                        res["index"] = train_index_count
                        train_total_duration += duration
                        train.append(res)
                        train_index_count += 1

                    utt2singer.write("{}\t{}\n".format(res["Uid"], res["Singer"]))

    print("#Train = {}, #Test = {}".format(len(train), len(test)))
    print(
        "#Train hours= {}, #Test hours= {}".format(
            train_total_duration / 3600, test_total_duration / 3600
        )
    )

    # Save train.json and test.json
    with open(train_output_file, "w") as f:
        json.dump(train, f, indent=4, ensure_ascii=False)
    with open(test_output_file, "w") as f:
        json.dump(test, f, indent=4, ensure_ascii=False)

    # Save singers.json
    singer_lut = {name: i for i, name in enumerate(unique_speakers)}
    with open(singer_dict_file, "w") as f:
        json.dump(singer_lut, f, indent=4, ensure_ascii=False)