Hugging Face's logo

ayousanz
/
yodas-ja000-WADA-SNR-filtering

Japanese
Model card Files Community
yodas-ja000-WADA-SNR-filtering
File size: 5,162 Bytes 
c113ea5
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
 
from datasets import load_dataset
import librosa
import IPython.display as ipd
from IPython.display import Audio, display
import random
from concurrent.futures import ProcessPoolExecutor
import numpy as np

import json


ds0 = load_dataset('espnet/yodas', 'ja000')
print("finished loading ja000")

def wada_snr(wav):
    # Direct blind estimation of the SNR of a speech signal.
    #
    # Paper on WADA SNR:
    #   http://www.cs.cmu.edu/~robust/Papers/KimSternIS08.pdf
    #
    # This function was adapted from this matlab code:
    #   https://labrosa.ee.columbia.edu/projects/snreval/#9

    # init
    eps = 1e-10
    # next 2 lines define a fancy curve derived from a gamma distribution -- see paper
    db_vals = np.arange(-20, 101)
    g_vals = np.array([0.40974774, 0.40986926, 0.40998566, 0.40969089, 0.40986186, 0.40999006, 0.41027138, 0.41052627, 0.41101024, 0.41143264, 0.41231718, 0.41337272, 0.41526426, 0.4178192 , 0.42077252, 0.42452799, 0.42918886, 0.43510373, 0.44234195, 0.45161485, 0.46221153, 0.47491647, 0.48883809, 0.50509236, 0.52353709, 0.54372088, 0.56532427, 0.58847532, 0.61346212, 0.63954496, 0.66750818, 0.69583724, 0.72454762, 0.75414799, 0.78323148, 0.81240985, 0.84219775, 0.87166406, 0.90030504, 0.92880418, 0.95655449, 0.9835349 , 1.01047155, 1.0362095 , 1.06136425, 1.08579312, 1.1094819 , 1.13277995, 1.15472826, 1.17627308, 1.19703503, 1.21671694, 1.23535898, 1.25364313, 1.27103891, 1.28718029, 1.30302865, 1.31839527, 1.33294817, 1.34700935, 1.3605727 , 1.37345513, 1.38577122, 1.39733504, 1.40856397, 1.41959619, 1.42983624, 1.43958467, 1.44902176, 1.45804831, 1.46669568, 1.47486938, 1.48269965, 1.49034339, 1.49748214, 1.50435106, 1.51076426, 1.51698915, 1.5229097 , 1.528578  , 1.53389835, 1.5391211 , 1.5439065 , 1.54858517, 1.55310776, 1.55744391, 1.56164927, 1.56566348, 1.56938671, 1.57307767, 1.57654764, 1.57980083, 1.58304129, 1.58602496, 1.58880681, 1.59162477, 1.5941969 , 1.59693155, 1.599446  , 1.60185011, 1.60408668, 1.60627134, 1.60826199, 1.61004547, 1.61192472, 1.61369656, 1.61534074, 1.61688905, 1.61838916, 1.61985374, 1.62135878, 1.62268119, 1.62390423, 1.62513143, 1.62632463, 1.6274027 , 1.62842767, 1.62945532, 1.6303307 , 1.63128026, 1.63204102])

    # peak normalize, get magnitude, clip lower bound
    wav = np.array(wav)
    max_val = np.abs(wav).max()
    if max_val == 0:
        max_val = eps
        
    wav = wav / max_val
    
    abs_wav = np.abs(wav)
    abs_wav[abs_wav < eps] = eps

    # calcuate statistics
    # E[|z|]
    v1 = max(eps, abs_wav.mean())
    # E[log|z|]
    v2 = np.log(abs_wav).mean()
    # log(E[|z|]) - E[log(|z|)]
    v3 = np.log(v1) - v2

    # table interpolation
    wav_snr_idx = None
    if any(g_vals < v3):
        wav_snr_idx = np.where(g_vals < v3)[0].max()
    # handle edge cases or interpolate
    if wav_snr_idx is None:
        wav_snr = db_vals[0]
    elif wav_snr_idx == len(db_vals) - 1:
        wav_snr = db_vals[-1]
    else:
        wav_snr = db_vals[wav_snr_idx] + \
            (v3-g_vals[wav_snr_idx]) / (g_vals[wav_snr_idx+1] - \
            g_vals[wav_snr_idx]) * (db_vals[wav_snr_idx+1] - db_vals[wav_snr_idx])

    # Calculate SNR
    dEng = sum(wav**2)
    dFactor = 10**(wav_snr / 10)
    dNoiseEng = dEng / (1 + dFactor) # Noise energy
    dSigEng = dEng * dFactor / (1 + dFactor) # Signal energy
    snr = 10 * np.log10(dSigEng / dNoiseEng)

    return snr


def preprocess_audio(data):
    # �?ータが整数型�?�場合、浮動小数点型に変換
    if data.dtype == np.int16:
        data = data.astype(np.float32) / np.iinfo(np.int16).max
    elif data.dtype == np.int32:
        data = data.astype(np.float32) / np.iinfo(np.int32).max

    # ス�?レオをモノラルに変換?���?要があれば?�?
    if len(data.shape) == 2:
        data = data.mean(axis=1)

    return data

# 音声データの前処理とSNR計算を行う関数
def process_audio_data(item):
    # 音声データの前処理
    audio_data = item['audio']['array']

    # 音声データが空でないことを確認
    if len(audio_data) == 0:
        return None

    preprocessed_data = preprocess_audio(audio_data)
    
    # WADA-SNRを計算
    snr = wada_snr(preprocessed_data)

    # データからidを取得
    uuid = item['utt_id']
    transcription = item['text']

    return {
        "ファイル名": uuid,
        "SNR値": snr,
        "トランスクリプション": transcription
    }

import os

if __name__ == '__main__':
    ds = load_dataset('espnet/yodas', 'ja000', trust_remote_code=True)

    print("データ数: ", ds['train'].dataset_size)

    # CPUのコア数を取得
    cpu_count = os.cpu_count()

    # 並列�?��?で関数を実�?
    with ProcessPoolExecutor(max_workers=cpu_count) as executor:
        results = list(executor.map(process_audio_data, ds['train']))

    # Noneを除去
    results = [result for result in results if result is not None]

    # 結果をJSONファイルに保存
    with open('audio_analysis_results.json', 'w') as f:
        json.dump(results, f, ensure_ascii=False, indent=4)

    print("JSONファイルが保存されました")