Convolutional neural networks for small-footprint keyword spotting

By Tara N. Sainath, Carolina Parada. 2015

584 citations recorded by the source corpus.

Subfields: Artificial Intelligence, Model Architectures, Speech & Audio

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

In Model Architectures

In Speech & Audio

Citations in the corpus

Streaming End-to-end Speech Recognition for Mobile Devices (2018, background): “Prominent examples include wakeword detection (i.e., recognizing speciﬁc words or phrases) [4, 5, 6], as well as large vocabulary continuous speech recognition (LVCSR) [7, 8].”
Speech Commands: A Dataset for Limited-Vocabulary Speech Recognition (2018, methodology): “Training the default convolution model from the TensorFlow tutorial (based on Convolutional Neural Networks for Small-footprint Keyword Spotting[19]) using the V1 training data gave a TopOne score of 85.”
Very deep convolutional neural networks for robust speech recognition (2016, methodology): “• Rather than using the traditional 9×9 or 3×4 filters and 1×3 pooling [7, 8, 9], VDCNNs for speech recognition use filters of 3 × 3 (sometimes 1 × 3 and 3 × 1), and the pooling size is constrained to 1 × 2 or 2 × 2.”
Domain Generalization: A Survey (2021, methodology): “Speech recognition - Google Speech Command [83] 65 k 1,888 Speakers”
Training Keyword Spotting Models on Non-IID Data with Federated Learning (2020, background): “Training such an architecture traditionally required framelevel labels generated by LVCSR systems [27] to provide accurate timing information.”
Efficient Keyword Spotting Using Dilated Convolutions and Gating (2018, methodology): “For comparison, we also add as a CNN variant the base architecture trad-fpool3 from [5], a multi-class model with 4 output labels (“hey”, “sni”, “ps”, and background).”
Hello Edge: Keyword Spotting on Microcontrollers (2017, methodology): “Figure 5 shows the number of operations per inference, memory requirement and test accuracy of neural network models from prior work [5, 6, 7, 8] trained on Google speech commands dataset overlayed with the memory and compute bounding boxes for the neural network classes from section 4.”
Streaming small-footprint keyword spotting using sequence-to-sequence models (2017, background): “Many previous works propose to train neural networks to identify word targets in individual keywords: for example, using feed-forward deep neural networks [4, 5, 6], convolutional networks [7] or recurrent neural networks [8, 9, 10].”
Structured Transforms for Small-Footprint Deep Learning (2015, methodology): “Results with 11 different models are reported in Figure 4 (left) including the state of the art keyword spotting model developed in [23].”
Generalizing Across Domains via Cross-Gradient Training (2018, methodology): “We use the architecture of Sainath & Parada (2015) 7 .”
Survey on Deep Neural Networks in Speech and Vision Systems (2019, background): “CNN improvement over DNN, 2015 [109] 65.”
Speech Processing for Digital Home Assistants: Combining signal processing with deep-learning techniques (2019, methodology): “To improve detection accuracy, convolutional [47], timedelay [48], or recurrent network layers [49] have been proposed, as well as subword modeling of the keyword and background speech using a DNN-HMM architecture [50], all of which aim to exploit the temporal properties of the input signal for clas”
Streaming keyword spotting on mobile devices (2020, methodology): “The standard end to end model architecture applied for KWS [12, 14] consists of a speech feature extractor(optional) and a neural network based classifier.”