Original title: Multi-Scale Residual Attention Network for Chinese Speech Recognition Through Collaborative Design
Yuying Li, Hongjie Wan, Akash Sutradhar, Jianing Qin · September 2026
Automatic speech recognition systems often struggle to catch speech patterns across different time scales. Many models also use inefficient attention steps. Meanwhile, lightweight models split data into groups, which traps information and stops the system from seeing the full picture. To fix this, researchers created a model called MSRDA-GSCA. It runs parallel convolutions, which are filters of different sizes, to gather sound patterns across varying time spans. Then, it uses an attention process that recalibrates channel and spatial data using element-wise max fusion. Next, it shuffles channels between attention steps so different groups can share clues. The researchers tested this setup on the THCHS-30 and ST-CMDS datasets. The new model beat existing methods on both character and word error rates. With only 4.65 M parameters, it stays small while delivering accurate Chinese speech recognition.
You are baking a cake and need to follow a recipe. If you only look at one ingredient at a time, you might forget how they combine. But if you also consider the whole recipe steps together, you avoid mistakes. In this paper, the model uses parallel processing to analyze speech at both fine and coarse time scales, capturing details and context simultaneously.
Most accurate speech recognizers are too large for phones or smart speakers. This model achieves top accuracy on Chinese speech with only 4.65 million parameters, making it practical for everyday devices. It also solves a key problem: capturing both short and long sound patterns without slowing down.
The abstract only evaluates on two specific Chinese speech datasets (THCHS-30 and ST-CMDS), so performance on other languages or real-world noisy conditions is not established. While the model is efficient with 4.65M parameters, it may not be the smallest possible, and comparisons are limited to existing methods on those datasets.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
Multi-Scale Residual Attention Network for Chinese Speech Recognition Through Collaborative Design — Yuying Li, Hongjie Wan, Akash Sutradhar et al. — Entropy — doi:10.3390/e28091037
Abstract licence: CC BY
Explanation generated by Synthesis from the abstract — it is not the original paper.