Small-footprint high-performance deep neural network-based speech recognition using split-VQ

By Yongqiang Wang, Jinyu Li, Yifan Gong. 2015

41 citations recorded by the source corpus.

Subfields: Artificial Intelligence, Efficiency & Compression, Speech & Audio, Signal Processing

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

1. Jinyu Li
2. Dong Yu
3. Tara N. Sainath
4. Rohit Prabhavalkar
5. Yoshua Bengio
6. Li Deng
7. Yu Zhang
8. Shinji Watanabe

In Efficiency & Compression

1. Jinyu Li
2. Rohit Prabhavalkar
3. Dong Yu
4. Carolina Parada
5. Ian McGraw
6. Ouais Alsharif
7. Antoine Bruguier
8. Raziel Alvarez

In Speech & Audio

1. Jinyu Li
2. Yoshua Bengio
3. Rohit Prabhavalkar
4. Tara N. Sainath
5. Dong Yu
6. Mirco Ravanelli
7. Kanishka Rao
8. Shinji Watanabe

In Signal Processing

1. Jinyu Li
2. Dong Yu
3. Rohit Prabhavalkar
4. Yoshua Bengio
5. Tara N. Sainath
6. Mirco Ravanelli
7. Ian McGraw
8. Li Deng

Citations in the corpus

• Recent progresses in deep learning based acoustic models (2017, background): “Therefore, refining with quantization is important for both very low-bit quantization [177], [178] and vector quantization [179] so that the training and testing are consistent.”
• Twin Regularization for online speech recognition (2018, methodology): “In particular, the use of feed-forward Deep Neural Networks (DNNs), including both fully-connected and convolutional architectures, has been largely investigated in the literature [7, 8], especially in the context of online ASR performed on small-footprint devices [9–13].”
• On the compression of recurrent neural networks with an application to LVCSR acoustic modeling for embedded speech recognition (2016, methodology): “[21] which uses a combination of singular value decomposition (SVD) and vector quantization to compress acoustic models.”
• Survey on Deep Neural Networks in Speech and Vision Systems (2019, methodology): “Several studies [167-170, 173] have customized DNN, CNN, and recurrent LSTM architectures with compression and quantization schemes to achieve considerable reductions in memory and computational requirements.”
• Personalized speech recognition on mobile devices (2016, background): “Vector quantization has also been shown to significantly reduce model size with only small accuracy losses [10], however it is unclear whether this algorithm can be implemented in a computationally efficient manner while minimizing runtime memory footprint.”
• On Online Attention-Based Speech Recognition and Joint Mandarin Character-Pinyin Training (2016, unknown): “, SVD decomposition, projection layers or quantization [26, 27]) we note that both are English and Mandarin models are less than 64 MiB in size.”