Merge pull request #55 from kuke/update_benchmark

Retune hyper-parameters and update benchmark results for English models due to #50

README.md

@@ -505,13 +505,13 @@ Language Model | Training Data | Token-based | Size | Descriptions
 
 Test Set                | LibriSpeech Model | BaiduEN8K Model
 :---------------------  | ---------------:  | -------------------:
-LibriSpeech Test-Clean  |   7.73            |   6.63
+LibriSpeech Test-Clean  |   6.85            |   5.73
-LibriSpeech Test-Other  |   23.15           |   16.59
+LibriSpeech Test-Other  |   21.18           |   14.47
-VoxForge American-Canadian | 12.30          |   7.46
+VoxForge American-Canadian | 12.12          |   7.37
-VoxForge Commonwealth   |   20.03           |   16.23
+VoxForge Commonwealth   |   19.82           |   15.58
-VoxForge European       |   30.31           |   20.47
+VoxForge European       |   30.15           |   19.44
-VoxForge Indian         |   55.47           |   28.15
+VoxForge Indian         |   53.73           |   26.15
-Baidu Internal Testset  |   44.71           |   8.92
+Baidu Internal Testset  |   40.75           |   8.82
 
 For reproducing benchmark results on VoxForge data, we provide a script to download data and generate VoxForge dialect manifest files. Please go to ```data/voxforge``` and execute ```sh run_data.sh``` to get VoxForge dialect manifest files. Notice that VoxForge data may keep updating and the generated manifest files may have difference from those we evaluated on.
 

examples/librispeech/run_infer.sh

@@ -21,8 +21,8 @@ python -u infer.py \
 --num_conv_layers=2 \
 --num_rnn_layers=3 \
 --rnn_layer_size=2048 \
---alpha=2.15 \
+--alpha=2.5 \
---beta=0.35 \
+--beta=0.3 \
 --cutoff_prob=1.0 \
 --cutoff_top_n=40 \
 --use_gru=False \

examples/librispeech/run_infer_golden.sh

@@ -30,8 +30,8 @@ python -u infer.py \
 --num_conv_layers=2 \
 --num_rnn_layers=3 \
 --rnn_layer_size=2048 \
---alpha=2.15 \
+--alpha=2.5 \
---beta=0.35 \
+--beta=0.3 \
 --cutoff_prob=1.0 \
 --cutoff_top_n=40 \
 --use_gru=False \

examples/librispeech/run_test.sh

@@ -22,8 +22,8 @@ python -u test.py \
 --num_conv_layers=2 \
 --num_rnn_layers=3 \
 --rnn_layer_size=2048 \
---alpha=2.15 \
+--alpha=2.5 \
---beta=0.35 \
+--beta=0.3 \
 --cutoff_prob=1.0 \
 --cutoff_top_n=40 \
 --use_gru=False \

examples/librispeech/run_test_golden.sh

@@ -31,8 +31,8 @@ python -u test.py \
 --num_conv_layers=2 \
 --num_rnn_layers=3 \
 --rnn_layer_size=2048 \
---alpha=2.15 \
+--alpha=2.5 \
---beta=0.35 \
+--beta=0.3 \
 --cutoff_prob=1.0 \
 --cutoff_top_n=40 \
 --use_gru=False \

examples/tiny/run_infer.sh

@@ -21,8 +21,8 @@ python -u infer.py \
 --num_conv_layers=2 \
 --num_rnn_layers=3 \
 --rnn_layer_size=2048 \
---alpha=2.15 \
+--alpha=2.5 \
---beta=0.35 \
+--beta=0.3 \
 --cutoff_prob=1.0 \
 --cutoff_top_n=40 \
 --use_gru=False \

examples/tiny/run_infer_golden.sh

@@ -30,8 +30,8 @@ python -u infer.py \
 --num_conv_layers=2 \
 --num_rnn_layers=3 \
 --rnn_layer_size=2048 \
---alpha=2.15 \
+--alpha=2.5 \
---beta=0.35 \
+--beta=0.3 \
 --cutoff_prob=1.0 \
 --cutoff_top_n=40 \
 --use_gru=False \

examples/tiny/run_test.sh

@@ -22,8 +22,8 @@ python -u test.py \
 --num_conv_layers=2 \
 --num_rnn_layers=3 \
 --rnn_layer_size=2048 \
---alpha=2.15 \
+--alpha=2.5 \
---beta=0.35 \
+--beta=0.3 \
 --cutoff_prob=1.0 \
 --cutoff_top_n=40 \
 --use_gru=False \

examples/tiny/run_test_golden.sh

@@ -31,8 +31,8 @@ python -u test.py \
 --num_conv_layers=2 \
 --num_rnn_layers=3 \
 --rnn_layer_size=2048 \
---alpha=2.15 \
+--alpha=2.5 \
---beta=0.35 \
+--beta=0.3 \
 --cutoff_prob=1.0 \
 --cutoff_top_n=40 \
 --use_gru=False \