KWS 派系類別

原創: [email protected]
時間: 2020/05/19

之前查看論文得方法有誤，鄙人還沾沾自喜，引以爲戒，paper非經典或必看切勿通篇閱讀
parper 閱讀手冊：你的Paper閱讀能力合格了嗎

一、LVCSR

• The LVCSR based systems need to generate rich lattices and high computational resources.

二、HMM-GMM

三、語音到文本

3.1 DNN

• 2014 | Small-footprint keyword spotting using deep neural networks
• arXiv:1709.03665 | attention | DNN + CTC
  • 不需要幀對齊，需要full-sized encoders
• arXiv:1812.02802 | Streaming + End-to-End
  • Singular value decomposition SVD，奇異值分解 減小模型
• 2019 | Multitask Learning of Deep Neural Network Based Keyword Spotting for IoT Devices | DNN + HMM
• 2019 | Time-Delayed Bottleneck Highway Networks Using a DFT Feature for Keyword Spotting

3.2 CNN | 考慮時間和頻率得相關性

• 2015 | Convolutional neural networks for small-footprint keyword spotting
• arXiv:1907.01448 | Sub-band CNN
• 2019 | A Small-Footprint End-to-End KWS System in Low Resources | CTC + End-to-End
• arXiv:1811.07684 | Dilated Convolutions | end-to-end

3.3 CRNN

• arXiv:1703.05390
  • 解碼窗1.5秒，不能做到真正得實時
• arXiv:1911.01803 | CRNN + temporal feedback connections

3.4 RNN

• arXiv:1512.08903 | LSTM + CTC

• arXiv:1611.09405 | RNN + CTC (End-to-End)

• arXiv:1705.02411 | LSTM + Max Pooling loss

  • 不需要依賴於音素級信息對齊，但是受限於解碼

• arXiv:1803.10916 | Attention | Encoding - Decoding | End-to-End

• 2019 | Adversarial examples for improving end-to-end attention-based small-footprint keyword spotting (End-to-End)

• 2019 | DesenNet-BiLSTM

• arXiv:1912.07575 | Encoding - Decoding

• arXiv:2002.10851 | Quantized LSTM + CTC

• ResNet| 更大得感受野

  • arXiv:1710.10361
  • arXiv:1904.03814 | TCNet | hyperconnect/TC-ResNet
  • arXiv:1912.05124 | CENet-GCN
  • arXiv:2004.08531 | MatchboxNet (end-to-end)

3.5 Other

• TDNN

  • 2017 | Compressed time delay neural network for small-footprint keyword spotting | SVD
  • 2019 | A Time Delay Neural Network with Shared Weight Self-Attention for
    Small-Footprint Keyword Spotting

• DSConv (Depthwise Separable Convolutions)

  • arXiv:1911.02086 | SincConv + DSConv | Raw audio

    arXiv:1808.00158 | ASR + SincConv | github

  • arXiv:2004.12200 | DSConv + ResNet

四、Query-by-Example

• LSTM

  • 2015 | Query-by-example keyword spotting using long short-term memory networks

• arXiv:1811.10736 | DONUT | CTC | 後驗概率

• RNN-T with attention

  • arXiv:1710.09617 | Squence-to-sequence | End-to-End | keyword/filler
    • 不需要幀對齊，需要full-sized encoders

• arXiv:1910.05171 | query enrollment and testing | user-specific queries

五、Other

5.1 Optimization model

• Compression methods
  • arXiv:1412.6115 | CV端用得比較多
  • 2016 | Model compression applied to small-footprint keyword spotting
  • arXiv:1711.07128
  • arXiv:1712.05877
  • arXiv:1902.05026
• Another optimization approach，non-streaming model convert to streaming model
  • arXiv:1811.07684
  • arXiv:2005.06720 | google-research/kws_streaming
• Quantized Distillation
  • 2016 | Knowledge Distillation for Small-footprint Highway Networks
  • 2018 | Compression of End-to-End Models
  • arXiv:1907.00873
• Low rank
  • 2016 | Model compression applied to small-footprint keyword spotting
  • 2017| Compressed time delay neural network for small-footprint keyword
    spotting

5.2 Loss

• CTC loss

  • 2006 | Connection-ist temporal classifification: labelling unsegmented sequence data with recurrent neural networks

• Max-pooling loss

  • arXiv:1705.02411
  • arXiv:2001.09246 | A smoothed maxpooling loss (end-to-end)

5.3 Dataset

5.3.1 Enhance

• 2019 | Adversarial Examples for Improving End-to-end Attention-based Small-footprint Keyword Spotting (End-to-End)

5.3.2 Small dataset

• 2018 | Fast asr-free and almost zero-resource keyword spotting using dtw and cnns for humanitarian monitoring
  • use DTW to augment the data
• 2019 | Meta learning for few-shot keyword spotting
  • suggests
    a few-shot meta-learning approach

5.4 Other

• arXiv:2005.03633 | Far-field
• 2019 | Improving keyword spotting and language identifification via neural architecture search at scale
• 2017 | Hey Siri: An On-device DNN-powered Voice Trigger for Apple’s Personal Assistan