Multi-QuartzNet: Multi-Resolution Convolution for Speech Recognition with Multi-Layer Feature Fusion
arXiv:2011.13090
Abstract
In this paper, we propose an end-to-end speech recognition network based on Nvidia's previous QuartzNet model. We try to promote the model performance, and design three components: (1) Multi-Resolution Convolution Module, replaces the original 1D time-channel separable convolution with multi-stream convolutions. Each stream has a unique dilated stride on convolutional operations. (2) Channel-Wise Attention Module, calculates the attention weight of each convolutional stream by spatial channel-wise pooling. (3) Multi-Layer Feature Fusion Module, reweights each convolutional block by global multi-layer feature maps. Our experiments demonstrate that Multi-QuartzNet model achieves CER 6.77% on AISHELL-1 data set, which outperforms original QuartzNet and is close to state-of-art result.
will be presented in SLT 2021
References in corpus (16)
- Deep Speech 2: End-to-End Speech Recognition in English and Mandarin
- Squeeze-and-Excitation Networks
- Densely Connected Convolutional Networks
- SGDR: Stochastic Gradient Descent with Warm Restarts
- A Comparative Study on Transformer vs RNN in Speech Applications
- Feature Pyramid Networks for Object Detection
- CBAM: Convolutional Block Attention Module
- Wav2Letter: an End-to-End ConvNet-based Speech Recognition System
- Self-Attention Transducers for End-to-End Speech Recognition
- Stochastic Gradient Methods with Layer-wise Adaptive Moments for Training of Deep Networks
- Fully Convolutional Speech Recognition
- AugFPN: Improving Multi-scale Feature Learning for Object Detection
- XceptionTime: A Novel Deep Architecture based on Depthwise Separable Convolutions for Hand Gesture Classification
- Multi-encoder multi-resolution framework for end-to-end speech recognition
- Multi-head Monotonic Chunkwise Attention For Online Speech Recognition
- Multi-Stream End-to-End Speech Recognition