<?xml version="1.0" encoding="utf-8" ?><rss version="2.0"><channel><title><![CDATA[libeicuo8108的博客]]></title><description><![CDATA[]]></description><link>https://blog.youkuaiyun.com/libeicuo8108</link><language>zh-cn</language><generator>https://blog.youkuaiyun.com/</generator><copyright><![CDATA[Copyright &copy; libeicuo8108]]></copyright><item><title><![CDATA[letter2phone]]></title><link>https://blog.youkuaiyun.com/libeicuo8108/article/details/109298661</link><guid>https://blog.youkuaiyun.com/libeicuo8108/article/details/109298661</guid><author>libeicuo8108</author><pubDate>Mon, 26 Oct 2020 20:52:55 +0800</pubDate><description><![CDATA[https://www.aclweb.org/anthology/P10-1080.pdf
]]></description><category></category></item><item><title><![CDATA[kaldi-gop]]></title><link>https://blog.youkuaiyun.com/libeicuo8108/article/details/108730688</link><guid>https://blog.youkuaiyun.com/libeicuo8108/article/details/108730688</guid><author>libeicuo8108</author><pubDate>Tue, 22 Sep 2020 20:53:55 +0800</pubDate><description><![CDATA[shell脚本位置kaldi-master/egs/gop/s5/run.sh

跑完log:


feat-to-len scp:../s1/data/test-1000_hires/feats.scp ark,t:data/test_10short/tmp.len 
utils/subset_data_dir.sh: reducing #utt from 1000 to 10
steps/nnet3/compute_output.sh --cmd run.pl --nj 1 --online-ivect]]></description><category></category></item><item><title><![CDATA[kaldi cmvn]]></title><link>https://blog.youkuaiyun.com/libeicuo8108/article/details/108568477</link><guid>https://blog.youkuaiyun.com/libeicuo8108/article/details/108568477</guid><author>libeicuo8108</author><pubDate>Mon, 14 Sep 2020 00:29:03 +0800</pubDate><description><![CDATA[cur_sum.AddRowSumMat(1.0, input_part , 0.0); // input 列求和得到cur_sum

cur_sumsq.AddDiagMat2(1.0, input_part, kTrans, 0.0); //Add the diagonal of a matrix times itself *this = diag(M^T M) + beta * *this (if trans == kTrans).

variance.Scale(1.0 / window_f...]]></description><category></category></item><item><title><![CDATA[gstreamer]]></title><link>https://blog.youkuaiyun.com/libeicuo8108/article/details/108493659</link><guid>https://blog.youkuaiyun.com/libeicuo8108/article/details/108493659</guid><author>libeicuo8108</author><pubDate>Wed, 09 Sep 2020 19:55:58 +0800</pubDate><description><![CDATA[mp3 转 wav


gst-launch-1.0 filesrc location=1.mp3 ! decodebin ! audioresample \
! audioconvert ! audio/x-raw,format=S24LE,rate=48000 ! wavenc \
! filesink location=out.wav

利用speex 编解码 插件:

源代码出自https://stackoverflow.com/questions/32564312/problems-with-us]]></description><category></category></item><item><title><![CDATA[音频知识点]]></title><link>https://blog.youkuaiyun.com/libeicuo8108/article/details/107758979</link><guid>https://blog.youkuaiyun.com/libeicuo8108/article/details/107758979</guid><author>libeicuo8108</author><pubDate>Mon, 03 Aug 2020 11:02:54 +0800</pubDate><description><![CDATA[ll *.mp3
-rw-rw-r-- 1 joe joe 39204 7月  30 13:44 0000452f-5aff-4a48-b9a9-4e24218753a1.mp3
-rw-rw-r-- 1 joe joe 30062 7月  30 19:46 9945b5b6-7b3c-4d7e-9d68-7d82953d124f.mp3
-rw-rw-r-- 1 joe joe  8672 7月  29 19:43 b81fc3c5-6dfd-4176-b9ee-d107f3a12763.mp3

fi.]]></description><category></category></item><item><title><![CDATA[shell编程计算1-1000中所有3或5的倍数之和]]></title><link>https://blog.youkuaiyun.com/libeicuo8108/article/details/107181927</link><guid>https://blog.youkuaiyun.com/libeicuo8108/article/details/107181927</guid><author>libeicuo8108</author><pubDate>Tue, 07 Jul 2020 15:08:33 +0800</pubDate><description><![CDATA[#!/bin/bash
sum=0 
int=1
while((int&lt;=1000))
do
    if (( $int%3==0 )) || (( $int%5==0 ));then
        sum=$((sum+int))
    fi
	let int++
done
echo $sum;


 #!/bin/bash
 sum=0
for((i=1;i&lt;=1000;i++));
do
    if (( $i%3==0 )) || (( $i%5==0 ));then
   ..]]></description><category></category></item><item><title><![CDATA[master_worker]]></title><link>https://blog.youkuaiyun.com/libeicuo8108/article/details/106986433</link><guid>https://blog.youkuaiyun.com/libeicuo8108/article/details/106986433</guid><author>libeicuo8108</author><pubDate>Sat, 27 Jun 2020 18:54:06 +0800</pubDate><description><![CDATA[见过很多获取服务器本地IP的代码，个人觉得都不是很好。从网上搜索到一个靠谱的脚本，分享一下！

通过 UDP 获取本机 IP，目前见过最优雅的方法

这个方法是目前见过最优雅获取本机服务器的IP方法了。没有任何的依赖，也没有去猜测机器上的网络设备信息。

而且是利用 UDP 协议来实现的，生成一个UDP包，把自己的 IP 放如到 UDP 协议头中，然后从UDP包中获取本机的IP。

这个方法并不会真实的向外部发包，所以用抓包工具是看不到的。但是会申请一个 UDP 的端口，所以如果经常调用也会比较耗时的，这里]]></description><category></category></item><item><title><![CDATA[录音切分]]></title><link>https://blog.youkuaiyun.com/libeicuo8108/article/details/106530044</link><guid>https://blog.youkuaiyun.com/libeicuo8108/article/details/106530044</guid><author>libeicuo8108</author><pubDate>Fri, 05 Jun 2020 10:41:04 +0800</pubDate><description><![CDATA[m4a 及其他转MP3 流程

首先 ffmpeg 转成wav

ffmpeg -y -i apple.m4a -ac 1 -ar 16000 apple.wav   # 苹果m4a格式 转成 单通道 16k采样率 的wav格式音频

lameapple.wav apple.mp3

切分录音:

sox apple.wav apple_1.wav trim 0 5.5 # 对apple.wav 切分0s到 5.5s部分
sox apple.wav apple_1.wav trim 62...]]></description><category></category></item><item><title><![CDATA[实时语音识别GStreamer+kaldi 搭建介绍]]></title><link>https://blog.youkuaiyun.com/libeicuo8108/article/details/106357009</link><guid>https://blog.youkuaiyun.com/libeicuo8108/article/details/106357009</guid><author>libeicuo8108</author><pubDate>Tue, 26 May 2020 21:13:55 +0800</pubDate><description><![CDATA[一.搭建 见博客

    http://www.luyixian.cn/news_show_274565.aspx

二. 结构

  1) master worker client 利用websocket交互

  2) gst 数据流媒体框架

  3) kaldi 模块,先不介绍

三. ws4py websocket

   官方文档:https://ws4py.readthedocs.io/en/latest/sources/requirements/

  ...]]></description><category></category></item><item><title><![CDATA[docker 学习]]></title><link>https://blog.youkuaiyun.com/libeicuo8108/article/details/106284327</link><guid>https://blog.youkuaiyun.com/libeicuo8108/article/details/106284327</guid><author>libeicuo8108</author><pubDate>Fri, 22 May 2020 21:03:56 +0800</pubDate><description><![CDATA[启动容器:


docker run ubuntu:15.10 /bin/echo "Hello world"

查看运行:

docker ps


runoob@runoob:~$ docker ps
CONTAINER ID        IMAGE                  COMMAND              ...  
5917eac21c36        ubuntu:15.10           "/bin/sh -c 'while t…"    ...

CONTAINE.]]></description><category></category></item><item><title><![CDATA[Deterministic Pytorch： pytorch如何保证可重复性]]></title><link>https://blog.youkuaiyun.com/libeicuo8108/article/details/106200700</link><guid>https://blog.youkuaiyun.com/libeicuo8108/article/details/106200700</guid><author>libeicuo8108</author><pubDate>Mon, 18 May 2020 20:53:51 +0800</pubDate><description><![CDATA[1 代码

要想整个模型是确定性的，我们需要设定好随机种子，这样当程序运行到具有random的模块时，由于伪随机的种子是相同的，那么随机得到的结果也是相同的。需要设定的随机种子如下。


def set_seed(self, seed=0):
    random.seed(seed)
    np.random.seed(seed)
    torch.manual_seed(seed) #为CPU设置种子用于生成随机数，以使得结果是确定的
    torch.cuda.manual_seed(seed)]]></description><category></category></item><item><title><![CDATA[Fully Convolutional Speech Recognition]]></title><link>https://blog.youkuaiyun.com/libeicuo8108/article/details/106107184</link><guid>https://blog.youkuaiyun.com/libeicuo8108/article/details/106107184</guid><author>libeicuo8108</author><pubDate>Wed, 13 May 2020 21:28:37 +0800</pubDate><description><![CDATA[待续

]]></description><category></category></item><item><title><![CDATA[Kaldi当中语音数据增强有什么方法]]></title><link>https://blog.youkuaiyun.com/libeicuo8108/article/details/106062745</link><guid>https://blog.youkuaiyun.com/libeicuo8108/article/details/106062745</guid><author>libeicuo8108</author><pubDate>Mon, 11 May 2020 21:26:55 +0800</pubDate><description><![CDATA[整理 知乎https://www.zhihu.com/people/leonjin回答，如侵权立删



音速扰动sp：将音频速度扩展为0.9,1.0,1,1

参考脚本utils/data/perturb_data_dir_speed_3way.sh

音量扰动vp：

参考脚本utils/data/perturb_data_dir_volume.sh

上面两个都是用sox实现


有加混响和加性噪声，参考脚本

steps/data/reverberate_data_dir.py

steps/...]]></description><category></category></item><item><title><![CDATA[Wav2Letter++  自动分割标准ASG]]></title><link>https://blog.youkuaiyun.com/libeicuo8108/article/details/105980910</link><guid>https://blog.youkuaiyun.com/libeicuo8108/article/details/105980910</guid><author>libeicuo8108</author><pubDate>Thu, 07 May 2020 20:58:38 +0800</pubDate><description><![CDATA[查看原文https://www.infoq.cn/article/BK9LZ7b4txmsOTCQ-vLm



Wav2Letter 的核心是个声学模型，就像我们也许已经猜到的那样，它可以预测声波中的字母


]]></description><category></category></item><item><title><![CDATA[半监督分类的方式：self-training和co-training的算法简述]]></title><link>https://blog.youkuaiyun.com/libeicuo8108/article/details/105972601</link><guid>https://blog.youkuaiyun.com/libeicuo8108/article/details/105972601</guid><author>libeicuo8108</author><pubDate>Thu, 07 May 2020 15:00:22 +0800</pubDate><description><![CDATA[自我学习 (self-training)

自我训练就是通过一系列的步骤，用已有的有标签的数据(labelled data)，去对剩下的还未标记的数据打标签。从而使得训练数据(training data)更多。具体步骤如下：

初始化有标签数据和无标签数据
	重复
	a) 用有标签的数据去训练一个模型
b) 用这个模型来对没有标签的数据做预测

c) 只选取最有把握的预测结果来标记数据(unlab...]]></description><category></category></item><item><title><![CDATA[Self-training for end-to-end speech recognition]]></title><link>https://blog.youkuaiyun.com/libeicuo8108/article/details/105957823</link><guid>https://blog.youkuaiyun.com/libeicuo8108/article/details/105957823</guid><author>libeicuo8108</author><pubDate>Thu, 07 May 2020 10:48:08 +0800</pubDate><description><![CDATA[目录ABSTRACT1. INTRODUCTION3. SEMI-SUPERVISED SELF-TRAINING3.1. Filtering3.2. Ensembles4. EXPERIMENTSdata4.2. Experimental Setting4.3. Results4.3.1 Supervised Baseline4.3.2 Evaluating Beam Search4.3.3 C...]]></description><category></category></item><item><title><![CDATA[SpecAugment]]></title><link>https://blog.youkuaiyun.com/libeicuo8108/article/details/105847455</link><guid>https://blog.youkuaiyun.com/libeicuo8108/article/details/105847455</guid><author>libeicuo8108</author><pubDate>Wed, 29 Apr 2020 23:41:18 +0800</pubDate><description><![CDATA[论文地址：https://arxiv.org/pdf/1904.08779.pdf
Augmentation Policy
我们旨在构建一种直接作用于log mel频谱图的增强策略，以帮助网络学习有用的功能。 出于以下目的的动机：这些功能应对时间方向的变形，频率信息的部分丢失和小部分语音片段具有较强的鲁棒性，因此我们选择了以下变形来制定策略：
1、时间warp 是通过tensorflow的spar...]]></description><category></category></item><item><title><![CDATA[SEMI-SUPERVISED ASR BY END-TO-END SELF-TRAINING 端到端半监督训练]]></title><link>https://blog.youkuaiyun.com/libeicuo8108/article/details/105839331</link><guid>https://blog.youkuaiyun.com/libeicuo8108/article/details/105839331</guid><author>libeicuo8108</author><pubDate>Wed, 29 Apr 2020 15:46:39 +0800</pubDate><description><![CDATA[作者分别来自 俄亥俄州立大学，Salesforce， 亚马逊 alexa，还是值得一读
目录ABSTRACTINTRODUCTIONSUPERVISED LEARNING FOR ASREnd-to-end ASR with CTC
ABSTRACT
尽管基于深度学习的端到端自动语音识别（ASR）系统大大简化了建模流程，但它们却遭受数据稀疏性的困扰。本文提出了一种用于半监督ASR的端到端系统的自训...]]></description><category></category></item><item><title><![CDATA[贪心 大纲]]></title><link>https://blog.youkuaiyun.com/libeicuo8108/article/details/105495710</link><guid>https://blog.youkuaiyun.com/libeicuo8108/article/details/105495710</guid><author>libeicuo8108</author><pubDate>Mon, 13 Apr 2020 19:37:29 +0800</pubDate><description><![CDATA[https://blog.youkuaiyun.com/hestendelin/article/details/105384299

]]></description><category></category></item><item><title><![CDATA[端到端asr系统搭建]]></title><link>https://blog.youkuaiyun.com/libeicuo8108/article/details/105255637</link><guid>https://blog.youkuaiyun.com/libeicuo8108/article/details/105255637</guid><author>libeicuo8108</author><pubDate>Wed, 01 Apr 2020 21:15:06 +0800</pubDate><description><![CDATA[验证了一个端到端语音识别系统，目前cer结果还不错。考虑怎么实际使用：
1）服务器端： 利用kaldi的流式处理方法（http://www.luyixian.cn/news_show_274565.aspx），会用到gsteamer 和kaldi 以及docker 方便部署
2）客户端：需要做做模型的压缩，目前模型参数30M，实现本地部署还是很方便的，下一步可以自己尝试在ios实现本地asr系统。...]]></description><category></category></item></channel></rss>