Sinovoice-2014-12-10 - 版本历史

2014年2月11日 (二) 14:00 Cslt

2014-02-11T14:00:37Z

2014-02-10T08:12:52Z

2014-02-10T08:12:40Z

2014-02-10T06:56:32Z

‎DNN Decoder

2014-02-10T06:45:35Z

‎6000 hour 16k training

2014-02-10T06:41:34Z

‎6000 hour 16k training

2014-02-10T06:38:39Z

‎DNN training

2014-02-10T05:47:14Z

2014-02-10T05:46:37Z

2014-02-10T05:42:21Z

←上一版本		2014年2月11日 (二) 14:00的版本
第66行：		第66行：
	:* std::exp/std::log result in very slow computation in train203. Solved the problem by replacing to standard exp() and log().		:* std::exp/std::log result in very slow computation in train203. Solved the problem by replacing to standard exp() and log().
	:* The RT of the latest decoder on train203 is 0.25		:* The RT of the latest decoder on train203 is 0.25
		+
		+	* Online decoder
		+	:* Chao will focus on interface change and CMN adaptation.

@@ 第64行： / 第64行： @@
 * Faster decoder
-* std::exp/std::log result in very slow computation in train203. Solved the problem by replacing to standard exp() and log().
+:* std::exp/std::log result in very slow computation in train203. Solved the problem by replacing to standard exp() and log().
-* The RT of the latest decoder on train203 is 0.25
+:* The RT of the latest decoder on train203 is 0.25

@@ 第62行： / 第62行： @@
 :* CLG decoder uses less memory in decoding
 :* HCLG is faster and more accurate than CLG, and more amiable to beam control [http://cslt.riit.tsinghua.edu.cn/cgi-bin/cvss/cvss_request.pl?step=view_request&cvssid=156 here]
-:* std::exp/std::log result in very slow computation in train203. Solved the problem by replacing to standard exp() and log().

@@ 第61行： / 第61行： @@
 * Comparison between CLG and HCLG decoder
 :* CLG decoder uses less memory in decoding
-:* HCLG is faster and more accurate than HCLG, and more amiable to beam control [http://cslt.riit.tsinghua.edu.cn/cgi-bin/cvss/cvss_request.pl?step=view_request&cvssid=156 here]
+:* HCLG is faster and more accurate than CLG, and more amiable to beam control [http://cslt.riit.tsinghua.edu.cn/cgi-bin/cvss/cvss_request.pl?step=view_request&cvssid=156 here]
 :* std::exp/std::log result in very slow computation in train203. Solved the problem by replacing to standard exp() and log().

@@ 第34行： / 第34行： @@
 ! Model !! WER !! RT
 |-
-|small LM, it 4,  -5/-9  ||15.80 || -
+|small LM, it 4,  -5/-9  ||15.80 || 1.18
 |-
-|large LM, it 4, -5/-9   || 15.30 || -
+|large LM, it 4, -5/-9   || 15.30 || 1.50
 |-
-|large LM,  it 4, -6/-9   || 15.36 || -
+|large LM,  it 4, -6/-9   || 15.36 || 1.30
 |-
-|large LM, it 4, -7/-9    || 15.25 || -
+|large LM, it 4, -7/-9    || 15.25 || 1.30
 |-
-|large LM, it 5, -5/-9    || 14.17 || -
+|large LM, it 5, -5/-9    || 14.17 || 1.10
 |-
-|large LM, it 5,  -5/-10 || 13.77 || -
+|large LM, it 5,  -5/-10 || 13.77 || 1.29
 |-
 |}

@@ 第29行： / 第29行： @@
 * Ran CE DNN to iteration 5 (8400 states, 80000 pdf)
-* Testing results go down to 13% WER.
+* Testing results go down to 13.77% WER (Sinovoice results: 11.78).
 {| class="wikitable"

@@ 第3行： / 第3行： @@
 ==Environment setting==
-* Another 3 3T disks are ready for RADI-0.
+* Another 3 3T disks are ready for RAID-0.
-* Another GPU machine was purchased.
+* Another GPU machine was purchased. Add 4 3T disks to construct RAID-0.
 ==Corpora==
-* Scripts for confidence generation is ready for auto transcription
+* Scripts for confidence generation is ready for auto transcription.
-* 300h telephone speech data (Sinovoice recording) were done
+* 300h telephone speech data (Sinovoice recording) were done.
-* Adaptation data ready
+* Adaptation data 900 sentences ready.
 ==470 hour 8k training==

@@ 第55行： / 第55行： @@
 * Hidden layer adaptation is better than input and output adaptation
 * Before Linear adaptation is better than after-linear adaptation
-* Results are [here http://cslt.riit.tsinghua.edu.cn/cgi-bin/cvss/cvss_request.pl?step=view_request&cvssid=158]
+* Results are [http://cslt.riit.tsinghua.edu.cn/cgi-bin/cvss/cvss_request.pl?step=view_request&cvssid=158 here]
 =DNN Decoder=
@@ 第61行： / 第61行： @@
 * Comparison between CLG and HCLG decoder
 :* CLG decoder uses less memory in decoding
-:* HCLG is faster and more accurate than HCLG, and more amiable to beam control [here http://cslt.riit.tsinghua.edu.cn/cgi-bin/cvss/cvss_request.pl?step=view_request&cvssid=156]
+:* HCLG is faster and more accurate than HCLG, and more amiable to beam control [http://cslt.riit.tsinghua.edu.cn/cgi-bin/cvss/cvss_request.pl?step=view_request&cvssid=156 here]
 :* std::exp/std::log result in very slow computation in train203. Solved the problem by replacing to standard exp() and log().