QuartzNet: Deep Automatic Speech Recognition with 1D Time-Channel Separable Convolutions

TL;DR

QuartzNet introduces a novel neural acoustic model utilizing 1D time-channel separable convolutions, achieving high accuracy with fewer parameters and effective fine-tuning capabilities for speech recognition tasks.

Contribution

The paper presents a new end-to-end speech recognition model with 1D separable convolutions, offering improved efficiency and comparable accuracy to state-of-the-art models.

Findings

Achieves near state-of-the-art accuracy on LibriSpeech and Wall Street Journal datasets.

Uses fewer parameters than competing models.

Can be effectively fine-tuned on new datasets.

Abstract

We propose a new end-to-end neural acoustic model for automatic speech recognition. The model is composed of multiple blocks with residual connections between them. Each block consists of one or more modules with 1D time-channel separable convolutional layers, batch normalization, and ReLU layers. It is trained with CTC loss. The proposed network achieves near state-of-the-art accuracy on LibriSpeech and Wall Street Journal, while having fewer parameters than all competing models. We also demonstrate that this model can be effectively fine-tuned on new datasets.