# Asset learns to compress speech encoders

Digest AI · Research · published 2026-10-14T04:00:00Z

Canonical: https://digestai.news/story/asset-learns-to-compress-speech-encoders

## Summary

It uses reinforcement learning (RL) to adaptively compress audio input sequences for better performance in large language models (LLMs). The framework divides the encoded segments into variable-duration parts and then combines them into a continuous audio token, aiming to maintain task-oriented information while reducing the number of tokens needed. This approach has shown promising results in automatic speech recognition (ASR) tasks, achieving comparable performance to an oracle that uses both character and phone boundaries with fewer audio tokens on English LibriSpeech data. In Mandarin AISHELL, it outperforms a phone-boundary oracle at a similar token rate. The team also conducted experiments demonstrating how the learned segmentation changes when applied to non-ASR tasks.

## Key points

- Introduces ASSET framework for adaptive speech compression
- Achieves comparable performance to oracle with fewer audio tokens
- Improves on phone-boundary oracle in Mandarin AISHELL

## Why it matters

This research could lead to more efficient LLMs by reducing the number of audio tokens required, potentially improving model scalability and resource usage.

## Sources

1. [ASSET: Adaptive Speech Segmentation via Reinforcement Learning](https://arxiv.org/abs/2610.14309) (arXiv cs.CL, 2026-10-14, primary source)

## Cite

Digest AI, "Asset learns to compress speech encoders", 14 October 2026, https://digestai.news/story/asset-learns-to-compress-speech-encoders

---

Written by Digest AI's editorial model from the linked sources; the sources are the record. Headlines, digests and key points are written by Digest AI and may be quoted with a link to the story page. Linked articles belong to their publishers. Terms: https://digestai.news/terms#reuse
JSON: https://digestai.news/story/asset-learns-to-compress-speech-encoders.json