Zhou's home Zhou's home
首页
说话人识别
算法题库
编程相关
语音识别
关于
  • 分类
  • 标签
  • 归档
  • 常用工具
  • 友情链接

ZhouWJ

吾生也有涯,而知也无涯
首页
说话人识别
算法题库
编程相关
语音识别
关于
  • 分类
  • 标签
  • 归档
  • 常用工具
  • 友情链接
  • Kaldi Toolbox

    • Kaldi的基本框架和逻辑
    • 使用预训练Xvector模型进行测试
    • Kaldi中的PLDA自适应
    • Kaldi声纹识别代码详解|egs/aishell
    • 源码阅读|Kaldi中的PLDA打分
    • PLDA相关代码详解
    • iVector相关代码详解
      • ivector-normalize-length
    • Voxceleb源码阅读
    • Kaldi相关项目及碎片知识
    • Kaldi中的nnet3
  • 算法原理梳理

  • 文献记录

  • 我的工作

  • 声纹识别
  • Kaldi Toolbox
xugaoyi
2021-11-06
目录

iVector相关代码详解

# ivector-normalize-length

将iVectors长度归一化,并乘上缩放因子sqrt(feature-dimension)

Usage:  ivector-normalize-length [options] <ivector-rspecifier> <ivector-wspecifier>
e.g.:
 ivector-normalize-length ark:ivectors.ark ark:normalized_ivectors.ark

Options:
  --normalize: Set this to false to disable normalization (bool, default = true)
  --scaleup: 归一化后的vec的缩放因子为'sqrt(dim)' (bool, default = true)
1
2
3
4
5
6
7
int main(int argc, char *argv[]) {
  using namespace kaldi;
  typedef kaldi::int32 int32;
  try {
    ParseOptions po(usage);
    bool normalize = true;

    po.Register("normalize", &normalize,
                "Set this to false to disable normalization");

    bool scaleup = true;
    po.Register("scaleup", &scaleup,
                "If 'true', the normalized iVector is scaled-up by 'sqrt(dim)'");

    po.Read(argc, argv);

    if (po.NumArgs() != 2) {
      po.PrintUsage();
      exit(1);
    }

    std::string ivector_rspecifier = po.GetArg(1),
        ivector_wspecifier = po.GetArg(2);


    int32 num_done = 0;

    double tot_ratio = 0.0, tot_ratio2 = 0.0;

    SequentialBaseFloatVectorReader ivector_reader(ivector_rspecifier);
    BaseFloatVectorWriter ivector_writer(ivector_wspecifier);


    for (; !ivector_reader.Done(); ivector_reader.Next()) {
      std::string key = ivector_reader.Key();
      Vector<BaseFloat> ivector = ivector_reader.Value();
      BaseFloat norm = ivector.Norm(2.0); // L2-norm,取绝对值之和并开根
      BaseFloat ratio = norm / sqrt(ivector.Dim()); // 它比正常情况下的预期值大多少
      if (!scaleup) ratio = norm;

      KALDI_VLOG(2) << "Ratio for key " << key << " is " << ratio;
      if (ratio == 0.0) {
        KALDI_WARN << "Zero iVector";
      } else {
        if (normalize) ivector.Scale(1.0 / ratio);
      }
      ivector_writer.Write(key, ivector);
      tot_ratio += ratio;
      tot_ratio2 += ratio * ratio;
      num_done++;
    }

    KALDI_LOG << "Processed " << num_done << " iVectors.";
    if (num_done != 0) {
      BaseFloat avg_ratio = tot_ratio / num_done,
          ratio_stddev = sqrt(tot_ratio2 / num_done - avg_ratio * avg_ratio);
      KALDI_LOG << "Average ratio of iVector to expected length was "
                << avg_ratio << ", standard deviation was " << ratio_stddev;
    }
    return (num_done != 0 ? 0 : 1);
  } catch(const std::exception &e) {
    std::cerr << e.what();
    return -1;
  }
}
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
  • 为什么与sqrt(dim)有关?可能如果特征数过多,会导致数值无穷小,所以再乘以缩放因子
  • 为什么有个--normalize的开关,本来就是归一化函数,也不会有人选false吧?
#Kaldi
上次更新: 2024/04/10, 22:12:29
PLDA相关代码详解
Voxceleb源码阅读

← PLDA相关代码详解 Voxceleb源码阅读→

最近更新
01
方言语音识别综述
04-05
02
主流ASR模型概述
03-26
03
一些参考图表及表述
05-28
更多文章>
Theme by Vdoing | Copyright © 2019-2024 Evan Xu | MIT License
  • 跟随系统
  • 浅色模式
  • 深色模式
  • 阅读模式