工作日记 19/11/22

work vivo daily

0. TODO

1. 阅读论文 [ISCA Speech][2018]Attention-based End-to-End Models for Small-Footprint Keyword Spotting
2. 使用CNN+GRU的RCNN结构分类'hi, jovi'和'小v小v'
3. *阅读
   
   • ATTENTION-BASED MODELS FOR TEXT-DEPENDENT SPEAKER VERIFICATION
   • SMALL-FOOTPRINT KEYWORD SPOTTING USING DEEP NEURAL NETWORKS

1. 论文

1.1 Model

1. 模型结构:
   
   • Attention-Based
   • Encoder对出入的序列进行特征抽取
   • 输出层对编码后的特征进行Attention
2. Attention(待阅读)
   
   • ATTENTION-BASED MODELS FOR TEXT-DEPENDENT SPEAKER VERIFICATION
3. KWS结构参数
   

   SMALL-FOOTPRINT KEYWORD SPOTTING USING DEEP NEURAL NETWORKS
   

   • The feed-forward DNN model had 3 hidden layers and 64 hidden nodes per layer with rectified linear unit (ReLU) non-linearity.
   • An input window with 15 left frames and 5 right frameswasused.
   • The LSTM and GRU models were built with 2 hidden layers and 64 hidden nodes per layer.
   • For the GRU KWS model, the final GRU layer was followed by a fully connected layer with ReLU non-linearity.
   • There were no stacked frames in the input for the LSTM and GRU models.
   • The smoothing window for Deep KWS was set to 20 frames.
   • We also trained a TDNN-basedacousticmodelusing∼3000hoursofspeechdata to perform frame-level alignment before KWS model training.

1.2 Data

训练集：18W正样本，100W负样本
验证机：1W正样本，5W负样本
测试集：3W正样本，3W负样本

序列：40-channel Mel-filterbank，25ms window，10ms frame shift，filterbank feature was converted to per-channel energy normalized (PCEN) Mel-spectrograms.

1.3 Conclusion

CNN + GRU + Attention

2. CNN + GRU

2.1 二分类

2.1.1 hi, jovi vs 小v

• Training Set: 2W + 2W
• Dev Set: 2K + 2K
• GRU Acc:
  
  • 84%
  • 86%
• CNN + GRU Acc:
  
  • 93%
  • 91%
• 结果分析：
  CNN具有局部特征抽取的能力，减轻了GRU的记忆压力

2.1.2 hi, jovi vs 噪声

• Training Set: 2W + 1W
• Dev Set: 2K + 1K
• Model: /home/vivoadmin/work/project/training/trigger_word/models/cnn-gru-2-class/saved-model-100-0.9957.h5.jovi-Vs-noise
• Acc:
  
  • 99%+
  • 99%+

2.1.3 小v vs 噪声

• Training Set: 2W + 1W
• Dev Set: 2K + 1K
• Model: /home/vivoadmin/work/project/training/trigger_word/models/cnn-gru-2-class/saved-model-100-0.9927.h5.xiaov-Vs-noise
• Acc:
  
  • 99%+
  • 99%+

2.2 三分类

2.2.1 小数据量

• Training Set: 2W + 2W + 1W
• Dev Set: 2K + 2K + 1K
• Model: /home/vivoadmin/work/project/training/trigger_word/models/cnn-gru-3-class/saved-model-2000-0.9056.h5.2w2w1w
• Acc:
  
  • 90.6%
  • 90.5%

2.2.1 噪声数据增多

• Training Set: 2W + 2W + 4W
• Dev Set: 2K + 2K + 6K
• Model: /home/vivoadmin/work/project/training/trigger_word/models/cnn-gru-3-class/saved-model-2000-0.9388.h5.2w2w4w
• Acc:
  
  • 91.0%
  • 93.8%