Qifusion-Net: Layer-adapted Stream/Non-stream Model for End-to-End Multi-Accent Speech Recognition

TL;DR

Qifusion-Net is a novel end-to-end multi-accent speech recognition model that adaptively fuses features without prior accent knowledge, improving accuracy across diverse accents.

Contribution

It introduces a layer-adapted fusion model with dynamic chunk strategy for streaming, multi-accent recognition without needing prior accent information.

Findings

Achieved 22.1% and 17.2% CER reduction on KeSpeech and MagicData-RMAC datasets.

Outperformed baseline models in multi-accent speech recognition tasks.

Enabled fine-grained feature fusion for improved recognition accuracy.

Abstract

Currently, end-to-end (E2E) speech recognition methods have achieved promising performance. However, auto speech recognition (ASR) models still face challenges in recognizing multi-accent speech accurately. We propose a layer-adapted fusion (LAF) model, called Qifusion-Net, which does not require any prior knowledge about the target accent. Based on dynamic chunk strategy, our approach enables streaming decoding and can extract frame-level acoustic feature, facilitating fine-grained information fusion. Experiment results demonstrate that our proposed methods outperform the baseline with relative reductions of 22.1$\%$ and 17.2$\%$ in character error rate (CER) across multi accent test datasets on KeSpeech and MagicData-RMAC.