2014-07-05 - 版本历史

Cslt：/* Multilingual ASR */

2014-07-09T00:16:46Z

‎Multilingual ASR

2014年7月9日 (三) 00:16 Cslt

2014-07-09T00:16:29Z

Cslt：以内容“==Resoruce Building== == Leftover questions== * Asymmetric window: Great improvement on training set(WER 34% to 24%), however the improvement is lost on test. * Multi...”创建新页面

2014-07-09T00:14:59Z

以内容“==Resoruce Building== == Leftover questions== * Asymmetric window: Great improvement on training set(WER 34% to 24%), however the improvement is lost on test. * Multi...”创建新页面

新页面

==Resoruce Building==

== Leftover questions==
* Asymmetric window: Great improvement on training set(WER 34% to 24%), however the improvement is lost on test.
* Multi GPU training: Error encountered
* Multilanguage training
* Investigating LOUDS FST.
* CLG embedded decoder plus online compiler.
* DNN-GMM co-training

== AM development ==

=== Sparse DNN ===
* GA-based block sparsity (+++++++++)

===Noise training===
:* Journal paper writing on going

===Multilingual ASR===

<pre>
HW 27h (HW TR LM not involved) HW27h (HW TR LM involved)
Fbank stream (monolang) 21.64 20.72
FBank non-stream (MPE4) 22.23 21.38
FBank stream (MPE4) 21.99 -
</pre>

===Denoising & Farfield ASR===

* Reverberant data delivered
* global CMN based spectrum checking done. Seems the signal/feature transform with DNN is not a very reasonable way[http://cslt.riit.tsinghua.edu.cn/cgi-bin/cvss/cvss_request.pl?account=wangd&step=view_request&cvssid=238 check here].

===VAD===

* Waiting for engineering work

===Scoring===

* Refine the acoustic model with AMIDA database. problem solved by involving both wsj and AMIDA.

===Embedded decoder===

* WER vs RT vs graph size done.
* The first deliver is Emb201407_BG_v0.0
* Demo done

==LM development==

===Domain specific LM===

h2. Domain specific LM construction

h3. Mixture LM

* TAG model: 127h HuaWei tag analysis done.
* Performance on the NUM-tagged model under testing.

==Word2Vector==

===W2V based doc classification===

* Good performance obtained with the SSA (semantic space allocation). That is, train a general GMM, and then represent each doc as the vector of the GMM weight.
* APSIPA paper submitted

==Semantic word tree==

:* Version v2.0 released (filter with query log)
:* Please deliver to /nfs/disk/perm/data/corpora/semanticTree (Xingchao)
:* Version v3.0 under going. Further refinement with Baidu Baike hierarchy

==NN LM==

* Character-based NNLM (6700 chars, 7gram), 500M data training done.
:* Inconsistent pattern in WER were found on Tenent test sets
:* probably need to use another test set to do investigation.

* Investigate MS RNN LM training

@@ 第20行： / 第20行： @@
 <pre>
-      AM\\testset    |JS27H_100| JS_126H |  JS_2h  |ShanXi_2h|ShaanXi2h|Hubei2h|   ENG   |
+      AM\\testset   |JS27H_100| JS_126H |  JS_2h  |ShanXi_2h|ShaanXi2h|Hubei2h|   ENG   |
- Tel201406.v1.0.S  |         |         |    -    |    -    |    -    |    -    |    -    |
+Tel201406.v1.0.S   |         |         |    -    |    -    |    -    |    -    |    -    |
- Tel201406.v1.1.S  |    -    |         |    -    |    -    |    -    |    -    |    -    |
+Tel201406.v1.1.S   |    -    |         |    -    |    -    |    -    |    -    |    -    |
 Tel201406.HW.v2.0.B|  20.51  |  18.30  |  17.61  |  24.18  |  23.04  |  22.51  |  56.26  |
 Tel201406.HW.v2.0.S|  20.07  |  17.80  |  17.75  |  23.79  |  22.44  |  22.53  |  36.77  |