“Sinovoice-2014-01-20”版本间的差异

来自cslt Wiki
跳转至： 导航、 搜索
（以内容“=Project management= * Xiaoming and Xiao Na were added into the mail list * Potential Huawei conference-transcribing project was discussed =DNN training= ==Environme...”创建新页面）
 
 
（相同用户的2个中间修订版本未显示）
第1行： 第1行：
−
=Project management=
 
−
 
−
* Xiaoming and Xiao Na were added into the mail list
 
−
* Potential Huawei conference-transcribing project was discussed
 
−
 
 
=DNN training=
 
=DNN training=
  
 
==Environment setting==
 
==Environment setting==
  
−
* New disk space (3T) was created and mounted at /nfs/disk1
+
* Accounts re-arrangement done on the SGE cluster. NO ROOT TO WORK.
−
* Jobs with 100 threads work fine on the cluster
+
* Changed NFS server to 40 processes, hope to increase disk reading.
 +
* Agree to withdraw root/sudo privilege.
 +
* Agree to create a RAID-0 with another 3 3T disks
  
 
==Corpora==
 
==Corpora==
−
* 60 hour data were cut this week
+
* Changed the data labeling strategy: gender and noise length will not be labelled for the following several corpora.
−
* Just send out to vendors for labeling
+
* Automatic labeling
−
* Waiting for out-source platform construction
+
:* Xiaoming will work with Zhiyong to discover how to generate transcriptions with confidence score held.
−
* We assume 60 hour data per week in future
+
:* The first step is to investigate the raw accuracy on the domain-dependent test, and then decide if it is appropriate to use automatic labeling
 +
* Xiao Na will prepare 300h telephone speech data (Sinovoice recording). This will be used to improve the 8k model.
 +
 
  
 
==470 hour 8k training==
 
==470 hour 8k training==
  
−
* CE training done
+
* MPE training done
−
* MPE training partially done
+
  
 
{| class="wikitable"
 
{| class="wikitable"
 
! Model !! CE !! MPE1!! MPE2 !! MPE3 !! MPE4
 
! Model !! CE !! MPE1!! MPE2 !! MPE3 !! MPE4
 
|-
 
|-
−
|4k states||23.27/22.85 || 21.35/18.87 || 21.18/18.76 || 21.07/18.54
+
|4k states ||23.27/22.85 || 21.35/18.87 || 21.18/18.76 || 21.07/18.54 || 20.93/18.32
 
|-
 
|-
−
|8k states ||22.16/22.22 || - ||20.36/17.94 || - ||
+
|8k states ||22.16/22.22 || 20.55/18.03 ||20.36/17.94 || 20.32/17.78 || 20.29/17.80
 
|-
 
|-
 
|}
 
|}
第33行： 第31行：
 
==6000 hour 16k training==
 
==6000 hour 16k training==
  
−
* Audio files ready. Files with incorrect sampling rates were removed
+
* Feature extraction done: solved several problems in the data: (1) short wave (2) unmatched file length (3) unmatched sample rate.
−
* Lexicon and LM were ready
+
* Training has gone to tri4b, quick increase of states/pdfs.
−
* Making MFCC features
+
* DNN training will be started on Tuesday.
−
* Initial model (6 iterations etc) can be delivered before the spring holiday
+
  
 
=DNN Decoder=
 
=DNN Decoder=
−
* Initial trail of DNN decoder based on the Sinovoice code was failed, largely due to FST compiler
+
 
−
* Change the strategy to an integrated approach: use the sinovoice system to control connections, and use Kaldi base for asr engine
+
* Sinovoice decoder: some errors in FST building. Many triphones were lost after C composing. Problems in cdgen?
−
* Xiaoming will do some investigation on the Sinovoice FST compiler, while Liu Chao will focus on the Kaldi-based decoder
+
* Kaldi decoder:  
 +
:* A minor difference between CLG/HCLG results was found. Debugging into the problem.
 +
:* CLG RT is comparable to the HCLG, roughly 0.3-0.4 in CSLT grid-2.
 +
:* Additional optimization on pdf-pre-computing will be investigated.
 +
:* Code deliver today.

2014年1月20日 (一) 10:12的最后版本

DNN training

Environment setting

  • Accounts re-arrangement done on the SGE cluster. NO ROOT TO WORK.
  • Changed NFS server to 40 processes, hope to increase disk reading.
  • Agree to withdraw root/sudo privilege.
  • Agree to create a RAID-0 with another 3 3T disks

Corpora

  • Changed the data labeling strategy: gender and noise length will not be labelled for the following several corpora.
  • Automatic labeling
  • Xiaoming will work with Zhiyong to discover how to generate transcriptions with confidence score held.
  • The first step is to investigate the raw accuracy on the domain-dependent test, and then decide if it is appropriate to use automatic labeling
  • Xiao Na will prepare 300h telephone speech data (Sinovoice recording). This will be used to improve the 8k model.


470 hour 8k training

  • MPE training done
Model CE MPE1 MPE2 MPE3 MPE4
4k states 23.27/22.85 21.35/18.87 21.18/18.76 21.07/18.54 20.93/18.32
8k states 22.16/22.22 20.55/18.03 20.36/17.94 20.32/17.78 20.29/17.80

6000 hour 16k training

  • Feature extraction done: solved several problems in the data: (1) short wave (2) unmatched file length (3) unmatched sample rate.
  • Training has gone to tri4b, quick increase of states/pdfs.
  • DNN training will be started on Tuesday.

DNN Decoder

  • Sinovoice decoder: some errors in FST building. Many triphones were lost after C composing. Problems in cdgen?
  • Kaldi decoder:
  • A minor difference between CLG/HCLG results was found. Debugging into the problem.
  • CLG RT is comparable to the HCLG, roughly 0.3-0.4 in CSLT grid-2.
  • Additional optimization on pdf-pre-computing will be investigated.
  • Code deliver today.
取自“http://cslt.org/mediawiki/index.php?title=Sinovoice-2014-01-20&oldid=9107”