2013-07-26 - 版本历史

Wangd：/* Experiments */

2013-07-29T02:57:10Z

‎Experiments

2013-07-29T02:53:57Z

‎Experiments

2013-07-29T02:49:05Z

‎Embedded progress

2013-07-26T10:56:18Z

‎Experiments

2013-07-26T10:55:58Z

‎Experiments

2013-07-26T10:40:40Z

‎Confidence estimation

2013-07-26T10:38:58Z

2013-07-26T10:38:22Z

2013-07-26T10:37:39Z

2013-07-26T10:36:14Z

@@ 第31行： / 第31行： @@
 |-
 |}
 * Sparse DNN on the ARM board

@@ 第14行： / 第14行： @@
 ! TASK !!cross-entropy (original)  !! MPE (it1) !! MPE (it2) !! MPE (it3) !! MPE (it4)
 |-
-|map ||22.98|| 23.91 ||- || - || -
+|map ||22.98|| 23.91 || 23.26 || 22.84 || -
 |-
-|2044 ||21.94 || 25.92 ||- || - || -
+|2044 ||21.94 || 25.92 ||24.47 || 24.10  || -
 |-
-|notetp3|| 14.73 || 21.64 ||- || - || -
+|notetp3|| 14.73 || 21.64 || 18.83 || 19.16 || -
 |-
-|record1900||8.45 || 8.93 ||- || - || -
+|record1900||8.45 || 8.93 || 7.60 || 8.46 || -
 |-
-|general||34.0  || 35.29 ||- || - || -
+|general||34.0  || 35.29 || 33.72 || 33.62 || -
 |-
-|online1||34.16  || 31.70||- || - || -
+|online1||34.16  || 31.70|| 31.45 || 31.33 || -
 |-
-|online2||27.10  || 24.56||- || - || -
+|online2||27.10  || 24.56|| 24.42 ||  24.37 || -
 |-
-|speedup|| 24.1  || 22.93 ||- || - || -
+|speedup|| 24.1  || 22.93 || 21.86 || 21.60 || -
 |-
 |}

@@ 第92行： / 第92行： @@
 *To be done
 # Shrink the NN structure (4 layer to 2 layer), and test the performance
 # The Kaldi decoder costs a lot when the graph is large. Need to improve the indexing of the FST structure.
 # Integrate the DNN FE with the pocket-sphinx decoder.

@@ 第9行： / 第9行： @@
 * Discriminative DNN
-Use sequential MPE (with the DNN-based alignment/denlattices). The network structure: 100 + 4 X 800 + 2100:
+Use sequential MPE (with the DNN-based alignment/denlattices). 100-hour training, network structure: 100 + 4 X 800 + 2100:
 {| class="wikitable"

@@ 第9行： / 第9行： @@
 * Discriminative DNN
-Use sequential MPE (with the DNN-based alignment/denlattices). The network structure: 100 + 4 X 8000 + 2100:
+Use sequential MPE (with the DNN-based alignment/denlattices). The network structure: 100 + 4 X 800 + 2100:
 {| class="wikitable"

@@ 第70行： / 第70行： @@
 ==Confidence estimation==
+===DNN confidence===
 * We are interested in confidence estimation from DNN output directly. This confidence is naturally 'posterior' and does not rely on graphs so simply to generalize, e.g., when examine which output is the best from multiple decoders.
 * The first design employs the best path in decoding/alignment, based on the state-posterior matrix directly. Code finished and the intuitive testing seems ok.
@@ 第79行： / 第79行： @@
 # CI Phone posterior-based (instead of state posterior-based) full path(instead of best path) confidence estimation.
+===Multi-graph decoding based on DNN confidence===
 * Due to the DNN-based confidence which is independent of decoding graphs, it is simple to compare or integrate results from various graphs. E.g., decoding can be performed on a general big graph and a user-specific graph, and then compare the confidence of the two results and select the best one.
@@ 第86行： / 第86行： @@
 # Coding finished.
 # Debuging and Test next week.
 == Embedded progress ==

@@ 第13行： / 第13行： @@
 ! cross-entropy (original)  !! MPE (it1) !! MPE (it2) !! MPE (it3) !! MPE (it4)
 |-
-|map ||22.98|| 23.91 || ||
+|map ||22.98|| 23.91 ||- || - || -
 |-
-||21.94 || 25.92 || ||
+||21.94 || 25.92 ||- || - || -
 |-
-notetp3|| 14.73 || 21.64 || ||
+notetp3|| 14.73 || 21.64 ||- || - || -
 |-
-record1900||8.45 || 8.93 || ||
+record1900||8.45 || 8.93 ||- || - || -
 |-
-general||34.0  || 35.29 || ||
+general||34.0  || 35.29 ||- || - || -
 |-
-online1||34.16  || 31.70|| ||
+online1||34.16  || 31.70||- || - || -
 |-
-online2||27.10  || 24.56|| ||
+online2||27.10  || 24.56||- || - || -
 |-
-speedup|| 24.1  || 22.93 || ||
+speedup|| 24.1  || 22.93 ||- || - || -
 |-
 |}