1、2013.11.30,1.嘗試使用Alize+Spro+python構(gòu)建說(shuō)話人識(shí)別平臺(tái) (1)ALIZE version 3.x- http:/mistral.univ-avignon.fr/index_en.html-LIA_RAL- LIA_SpkDET (2)Spro4.0.1-http:/www.irisa.fr/metiss/guig/spro/index.html-Filter-bank cepstral features (3) python3.3.2 2.機(jī)器學(xué)習(xí)的哲學(xué)探索讀書(shū)筆記,參考素材,1.paper-ALIZE, a free toolkit for speaker r

2、ecognition 2.paper-ALIZE/SpkDet: a state-of-the-art open source software for speaker recognition 3.HTK -http:/htk.eng.cam.ac.uk/ 5.BILLs block”使用Alize等工具構(gòu)建說(shuō)話人識(shí)別平臺(tái)”-http:/ibillxia.github.io/blog/2013/04/26/building-speaker-recognition-system-using-alize-etc/ 6. ALIZE 3.0 - Open-source platform for sp

3、eaker recognition-/technical-committees/list/sl-tc/spl-nl/2013-05/ALIZE/,ALIZE介紹,The ALIZE project consists of a low level API (ALIZE) and a set of high level executables that form the LIA_RAL toolkit. The ensemble makes it possible to easily set up a speaker rec

4、ognition system for research purposes as well as develop industry based applications. LIA_RAL is a high level toolkit based on the low level ALIZE API. It consists of three sets of executables: LIA_SpkSeg, LIA_Utils and LIA_SpkDET. LIA_SpkSeg and LIA_Utils respectively include executables dedicated

5、to speaker segmentation and utility programs to handle ALIZE objects while LIA_SpkDet is developed to fulfil the main functions of a state-of-the-art speaker recognition system as described in the following figure.,ALIZE介紹,ALIZE does not include acoustic feature extraction but is compatible withSPro

6、 , HTK and RAW formats Score matrices can be exported in binary format easily handled by the BOSARIS toolkit,SPro介紹,spro is aimed at extracting features in the area of speaker recognition,you can extract features such as mfcc and lpc. SPro is a free speech signal processing toolkit which provides ru

7、ntime commands implementing standard feature extraction algorithms for speech related applications and a C library to implement new algorithms and to use SPro files within your own programs. SPro was originally designed for variable resolution spectral analysis but also provides for feature extracti

8、on techniques classically used in speech applications. There are commands for the following representations:filter-bank energies cepstral coefficients linear prediction derived representation,SPro介紹,Though the toolkit has been designed as a front-end for applications such as speech or speaker recogn

9、ition, we believe the library provides enough possibilities to implement various feature extraction algorithms easily (e.g. zero crossing rate). However, no command for such features is provided. The library, written in ANSI C, provides functions for the following: waveform signal input low-level si

10、gnal processing (FFT, LPC analysis, etc.) low-level feature processing (lifter, CMS, variance normalization, deltas, etc.) feature I/O,SPro介紹,The library does not provide for high-level feature extraction functions which directly converts a waveform into features, mainly because such functions would

11、 require a tremendous number of arguments in order to be versatile. However, it is rather trivial to write such a function for your particular needs using the SPro library.,SPro介紹,Filter-bank cepstral features The second filter-bank analysis tool,sfbcep, takes as input a waveform and output filter-b

12、ank derived cepstral features. The filter-bank processing is similar to what is done insfbank(see previous section). The cepstral coefficients are computed by DCTing the filter-bank log-magnitudes and possibly liftered. Optionally, the log-energy can be added to the feature vector. Insfbcep, the fra

13、me energy is calculated as the sum of the squared waveform samples after windowing. As for the magnitudes in the filter-bank, the log-energy are thresholded to keep them positive or null. The log-energies may be scaled to avoid differences between recordings. Mean and variance normalization of the s

14、tatic cepstral coefficients can be specified with the global-cmsand-normalizeoptions but do not apply to log-energies. The normalizations can be global (default) or based on a sliding window whose length is specified with-segment-length. Finally, first and second order derivatives of the cepstral co

15、efficients and of the log-energies can be appended to the feature vectors. When using delta features, the absolute log-energy can be suppressed using the-no-static-energyoption,第1步，特征提取MFCC,sfbcep.exe（MFCC）,第2步， Silence removal 靜音檢測(cè)和去除,NormFeat.exe 先能量規(guī)整 EnergyDetector.exe 基于能量檢測(cè)的靜音去除,第3步, Features

16、Normalization 特征規(guī)整,NormFeat.exe 再使用這個(gè)工具進(jìn)行特征規(guī)整,第4步， World model training,TrainWorld.exe 訓(xùn)練UBM,第5步， Target model training,TrainWorld.exe 在訓(xùn)練好UBM的基礎(chǔ)上，訓(xùn)練training set和testing set的GMM,第6步， Testing,ComputeTest.exe 將testing set 的GMM在training set的GMM上進(jìn)行測(cè)試和打分,第7步， Score Normalization,ComputeNorm.exe 將得分進(jìn)行規(guī)整,第

17、8步， Compute EER 計(jì)算等錯(cuò)誤率,可以查查計(jì)算EER的matlab代碼，NIST SRE的官網(wǎng)上有下載（/iad/mig/tools/DETware_v2.1.targz.htm）,others,關(guān)于各步驟中參數(shù)的問(wèn)題，可以在命令行“工具 -help”來(lái)查看該工具個(gè)參數(shù)的具體含義，另外還可參考Alize源碼中各個(gè)工具的test目錄中提供的實(shí)例， 而關(guān)于每個(gè)工具的作用及理論知識(shí)則需要查看相關(guān)論文。 常見(jiàn)問(wèn)題及解答: http:/mistral.univ-avignon.fr/mediawiki/index.php/Frequently_as

18、ked_questions 更多問(wèn)題請(qǐng)?jiān)贕oogle論壇（,Others-ALIZE中用到的功能（其它功能作用待研究）,Others-淺談Python程序和C程序的整合,利用 ctypes 模塊整合 Python 程序和 C 程序 ctypes 是 Python 的一個(gè)標(biāo)準(zhǔn)模塊，它包含在 Python2.3 及以上的版本里。ctypes 是一個(gè) Python 的高級(jí)外部函數(shù)接口，它使得 Python 程序可以調(diào)用 C 語(yǔ)言編譯的靜態(tài)鏈接庫(kù)和動(dòng)態(tài)鏈接庫(kù)。運(yùn)用 ctypes 模塊，能夠在 Python 源程序中創(chuàng)建，訪問(wèn)和操作簡(jiǎn)單的或復(fù)雜的 C 語(yǔ)言數(shù)據(jù)類(lèi)型。最為重要的是 ctypes 模塊能夠在

19、多個(gè)平臺(tái)上工作，包括 Windows，Windows CE，Mac OS X，Linux，Solaris，F(xiàn)reeBSD，OpenBSD。,機(jī)器學(xué)習(xí)的哲學(xué)探索-A.學(xué)科前沿P32,使用CiteSpace 2描繪知識(shí)圖譜【附圖】,機(jī)器學(xué)習(xí)的哲學(xué)探索-A.學(xué)科前沿P32,1. 知識(shí)圖譜顯示的學(xué)科前沿：【附圖】,機(jī)器學(xué)習(xí)的哲學(xué)探索-A.學(xué)科前沿P32,2.重要的作者【附圖】,機(jī)器學(xué)習(xí)的哲學(xué)探索-A.學(xué)科前沿P32,3.前沿知識(shí)群：增強(qiáng)學(xué)習(xí)；分類(lèi)技術(shù)；數(shù)據(jù)挖掘 4.對(duì)知識(shí)群的分析得出，機(jī)器學(xué)習(xí)研究的兩種目的：增強(qiáng)自身的性能+學(xué)習(xí)到人類(lèi)可以理解的知識(shí),機(jī)器學(xué)習(xí)的哲學(xué)探索-B.機(jī)器學(xué)習(xí)研究的演化路徑P58,1.“增強(qiáng)學(xué)習(xí)”是機(jī)器學(xué)習(xí)研究中相對(duì)獨(dú)立的一個(gè)領(lǐng)域，是布魯克斯創(chuàng)立的行為主義人工智能研究范式在機(jī)器學(xué)習(xí)研究中的繼續(xù)。 2.“數(shù)據(jù)挖掘”將兩個(gè)曾經(jīng)對(duì)立的研究范式融合在一起，采用一種實(shí)用主義的態(tài)度共同解決實(shí)踐中的問(wèn)題。 以數(shù)據(jù)挖掘?yàn)楹诵牡慕y(tǒng)計(jì)學(xué)習(xí)處理的一個(gè)最重要或者最基本的問(wèn)題是“分類(lèi)”?；蛟S有人會(huì)認(rèn)為“分類(lèi)”是一個(gè)微不足道的過(guò)程，然而“分類(lèi)”卻遍布于智能的理解過(guò)程之中，及時(shí)向“機(jī)器人規(guī)劃”這樣的活動(dòng)，也能夠構(gòu)建成為“分類(lèi)”問(wèn)題?；蛘哒f(shuō)，目前整個(gè)機(jī)器學(xué)習(xí)的核心問(wèn)題就是一個(gè)