---
title: "Lost but not erased: Finding traces of a forgotten language in neural speech models"
source: "arXiv AI/CL/LG（论文）"
source_tier: "t2"
url: "https://arxiv.org/abs/2608.25976v1"
permalink: "https://feed.astralor.com/items/itm_6f88303cc2fbc7339fa47b2a"
published: "2026-08-26T16:32:41.000Z"
score: 73
channel: "ai"
tags: ["Critical Period Hypothesis","Continual Learning","ASR","Representational Entrenchment"]
---

# Lost but not erased: Finding traces of a forgotten language in neural speech models

## AI commentary

> This paper is worth paying attention to not because it gives final answers, but because it translates the 'critical period' from a biological puzzle into a computable problem of representation entrenchment. It is especially instructive for researchers working on language acquisition, continual learning, and model plasticity; for those building speech recognition or foundation models, it also suggests that low-level features may carry surprising transfer value.

## At a glance

This study uses automatic speech recognition models as a proxy for international adoptees who cannot speak or comprehend their birth language. Models were trained on a first language, then abruptly switched to a second. Results show traces of the first language persist through second-language training, mainly in the lowest, pre-phonemic layers. These traces are functional: models with early exposure relearn their lost language 14% faster than naive baselines, and this advantage survives even when the early language is related, yet disappears if the lowest layers are replaced by a non-adopted model. The authors argue that critical-period effects reflect entrenchment of foundational representations, not a maturational loss of plasticity, emphasizing that experience shapes critical periods in language acquisition.

## Key points

- Traces of a first language persist in the lowest pre-phonemic layers after a switch to a second language.
- Early-exposed models relearn the forgotten language 14% faster than naive models.
- The advantage also holds for related-language early exposure but vanishes when the lowest layers are substituted.
- Suggests critical-period effects stem from representational entrenchment rather than lost plasticity.

## Full text

International adoptees retain phonological traces of a birth language they can no longer speak or comprehend, a persistence typically attributed to a biologically-timed critical period. We asked whether it could instead reflect the ordinary dynamics of learning, using automatic speech recognition models that simulate the international adoptee experience without maturational confounds. Models were trained on one language and then abruptly switched to a second. We found that traces of the first language persisted throughout second-language training, but mainly in the lowest, pre-phonemic layers. These traces were functional, as models with early exposure re-learned their lost first language 14% faster than naive models; this advantage held even against models adopted early from a related language and disappeared when the earliest layers were substituted from a non-adopted model. We argue that these critical-period effects reflect entrenchment of foundational representations rather than a maturational loss of plasticity, and that experience plays a central role in critical periods in language acquisition.

---

Verify against the original before citing; curated by Astral Feed: https://feed.astralor.com/items/itm_6f88303cc2fbc7339fa47b2a