# A recurrent design aims to shrink speech-recognition models

> Researchers propose sharing weights across repeated blocks to cut ASR model size while retaining most of the accuracy of larger models.

Oossa · 2026-10-30 · https://oossa.com/en/a-recurrent-design-aims-to-shrink-speech-recognition-models

A team writing on arXiv proposes a speech-recognition model that reuses a small set of blocks instead of giving every layer its own parameters. The researchers say this approach, combined with training methods that transfer knowledge from larger models, uses about one-tenth as many parameters while retaining accuracy.

On Open ASR Leaderboard benchmarks, their model with two shared blocks recovered 97.4% of large models’ average accuracy, according to the abstract. The team also describes a method for adapting the compact model to individual users, but the abstract does not give device measurements or establish how it performs in real-world use.

## The facts

- The two-block model recovered 97.4% of large models’ average accuracy across Open ASR Leaderboard benchmarks, the researchers report.
- The paper was posted on arXiv on October 30, 2026.

## Why it matters

For people building speech recognition for watches and glasses, fewer model parameters could ease memory constraints; whether this design delivers useful speed or energy savings on actual devices remains unknown.

## Sources & references

1. [Depth through Recurrence: Towards Ultra-Efficient On-Device ASR](https://arxiv.org/abs/2610.37833) – arXiv cs.LG, 2026-10-30

Last updated: 2026-10-30