2013-12-27 - 版本历史

Cslt：/* Efficient DNN training */

2013-12-27T02:31:27Z

‎Efficient DNN training

Cslt：/* Speech QA */

2013-12-27T02:29:37Z

‎Speech QA

Cslt：/* Speech QA */

2013-12-27T02:29:04Z

‎Speech QA

Cslt：/* Embedded development */

2013-12-27T02:28:19Z

‎Embedded development

Cslt：/* NN LM */

2013-12-27T02:27:45Z

‎NN LM

Cslt：/* Optimal phoneset */

2013-12-27T02:27:18Z

‎Optimal phoneset

Cslt：/* Efficient DNN training */

2013-12-27T01:55:02Z

‎Efficient DNN training

Cslt：/* Efficient DNN training */

2013-12-27T01:54:47Z

‎Efficient DNN training

2013年12月27日 (五) 01:54 Cslt

2013-12-27T01:54:26Z

2013年12月27日 (五) 01:49 Cslt

2013-12-27T01:49:41Z

@@ 第17行： / 第17行： @@
 # Moment-based training. With m=0.2 performs the best on WER. Results are not so consistent. For 1900, m=0.2 is the best; For online1 and online2, m=0 is the best.
 # Asymmetric window: Great improvement on training set(WER 34% to 24%), however the improvement is lost on test. Overfitting?
-# Frame-skipping. Skipping 1 frame does not impact performance too much and speed up decoding in a consistent way. Skipping more frames lead to further performance degradation (no acceptable).
+# Frame-skipping. Skipping 1 frame speeds up decoding in a consistent way while retaining the accuracy largely. Skipping more frames lead to unacceptable performance degradation.
 <pre>

@@ 第121行： / 第121行： @@
 ==Speech QA==
 * Use N-best to expand match in QA. Better performance were obtained.
@@ 第126行： / 第127行： @@
 :* 10-best matches 102/121
-* Use N-best to recover errors in entity check
+* Use N-best to recover errors in entity check. Working on.
-* Use Pinyin to recover errors in entity check
+* Use Pinyin to recover errors in entity check. Future work.

@@ 第122行： / 第122行： @@
 ==Speech QA==
-* Use N-best to expand match in QA.
+* Use N-best to expand match in QA. Better performance were obtained.
 :* 1-best matches 96/121
 :* 10-best matches 102/121

@@ 第112行： / 第112行： @@
 ==Embedded development==
-* Narrow and deep small scale NN trained. Some errors.
+* Narrow and deep small scale NN trained. Investigating some bugs.
 * Embedded stream mode on progress.
 * On-the-fly grammar compiler

@@ 第107行： / 第107行： @@
 ===NN LM===
-* Collecting more lexicon: 40k words related to music, 56k words from official dictionary.
+* Collecting a bigger lexicon: 40k words related to music, 56k words from an official dictionary.
 * Working on NN LM based on word2vector.

@@ 第28行： / 第28行： @@
 === Optimal phoneset===
-# Experiment 3 phone sets: Tencent, CSLT, PQ
+* Experiment 3 phone sets: Tencent, CSLT, PQ
 <pre>
@@ 第52行： / 第58行： @@
-Puqiang：
+PQ：
 map 9: %WER 24.25 [ 3547 / 14628, 115 ins, 428 del, 3004 sub ]
@@ 第91行： / 第97行： @@
 </pre>
 ===Engine optimization===

@@ 第20行： / 第20行： @@
 <pre>
-                     mom0.05  mom0.1  mom0.3  mom0.4  mom0.5  mom0.6  mom0.8  fs_1  fs_2  fs_3
+                      mom0.05  mom0.1  mom0.3  mom0.4  mom0.5  mom0.6  mom0.8  fs_1  fs_2  fs_3
              -------------------------------------------------------------------------------------
              avg_time | 4500     4175    3380    3460    3448    3521    4212    3149   2692  2716

@@ 第15行： / 第15行： @@
 === Efficient DNN training ===
-# Moment-based training. With m=0.2 performs the best on WER. 6.8% improvement on WER. Other settings are tried on 0.05,0.1,0.2,..0.6,0.8,1.0.
+# Moment-based training. With m=0.2 performs the best on WER. Results are not so consistent. For 1900, m=0.2 is the best; For online1 and online2, m=0 is the best.
-# Asymmetric window: left 20, right 5. NN accuracy increase by 7%, however WER is a bit worse than the baseline. Move back to Tencent 100h training.
+# Asymmetric window: Great improvement on training set(WER 34% to 24%), however the improvement is lost on test. Overfitting?
-# Frame-skipping is on implementation.
+# Frame-skipping. Skipping 1 frame does not impact performance too much and speed up decoding in a consistent way. Skipping more frames lead to further performance degradation (no acceptable).
 === Optimal phoneset===
 # Experiment 3 phone sets: Tencent, CSLT, PQ
-# Some errors occur in pure CHS experiments
@@ 第34行： / 第103行： @@
 ===NN LM===
-* Trained with 500M QA data, 110k vocabulary.
+* Collecting more lexicon: 40k words related to music, 56k words from official dictionary.
-* Tested on number of hidden layers (DNN), performance is better for some tests, but not for others.
+* Working on NN LM based on word2vector.
-* Tested on larger projection layer, from 256 to 384, the performance is consistently improved.
-−
 ==Embedded development==
 * Embedded stream mode on progress.
@@ 第46行： / 第114行： @@
 ==Speech QA==
-* SP-QA accuracy 45.14% in all the input (18*199).
+* Use N-best to expand match in QA.
-* Investigate the error patterns:
+:* 1-best matches 96/121
-:* 70% errors are caused by incorrect name entity recognition. Working on entity recovery (character, pinyin, ... distance penalty).
+:* 10-best matches 102/121
-:* 8% errors are caused by English names. Use class-based LM to solve the problem. Ready to work.
-:* Use N-best to recover errors in QA.
+* Use N-best to recover errors in entity check