@inproceedings{ferreira-etal-2026-echoes,
title = "Echoes in Plato{'}s Cave: Measuring Global and Local Alignment Between Speech and Language Representations",
author = "Ferreira, Alef Iury Siqueira and
Stefanel Gris, Lucas Rafael and
de Oliveira, F S and
Galv{\~a}o Filho, Arlindo Rodrigues and
Soares, Anderson Da Silva",
editor = "T.Y.S.S., Santosh and
Ghosh, Sreyan and
Duraiswami, Ramani and
Watanabe, Shinji and
Arora, Siddhant and
Kumar, Sonal and
Adi, Yossi and
Manocha, Dinesh and
Garcia Perera, Leibny Paola",
booktitle = "The Proceedings of Speech and Audio Language Models Workshop (2nd Edition)",
month = oct,
year = "2026",
address = "Budapest, Hungary",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.salma-1.10/",
pages = "134--155",
ISBN = "979-8-89176-496-5",
abstract = "We investigate representational alignment between nine independently trained speech encoders and ten text encoders. We build a controlled corpus in which the same sentences are rendered by different speakers and under different emotional conditions, so that prosody varies while linguistic content is held fixed. We compare every speech{--}text layer pair and calibrate the strongest match with a permutation procedure that prevents the search over pairs from inflating alignment scores. Local neighborhood agreement between modalities survives calibration throughout and is strongest between late speech layers and early text layers, providing strong evidence of genuine local alignment. Global similarity measures, however, do not provide reliable evidence that the two modalities share a common geometric space as a whole. Alignment is unaffected by speaker identity and prosodic variation, so the shared structure is tied to the content of the sentence itself rather than to how it is spoken. It also does not increase with model capability or with how well a text model reproduces human judgments of meaning, suggesting that this shared structure corresponds more closely to lexical and surface-level representations than to fine-grained sentence meaning. These findings indicate where speech and text representations can be connected, suggest that cross-modal transfer may benefit from exploiting neighborhood structure rather than assuming a shared global geometry, and provide new alignment methods with an existing lexical correspondence to build on, while pointing to sentence-level meaning as the structure that future speech representations could capture so that alignment extends beyond the lexical level."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="ferreira-etal-2026-echoes">
<titleInfo>
<title>Echoes in Plato’s Cave: Measuring Global and Local Alignment Between Speech and Language Representations</title>
</titleInfo>
<name type="personal">
<namePart type="given">Alef</namePart>
<namePart type="given">Iury</namePart>
<namePart type="given">Siqueira</namePart>
<namePart type="family">Ferreira</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Lucas</namePart>
<namePart type="given">Rafael</namePart>
<namePart type="family">Stefanel Gris</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">F</namePart>
<namePart type="given">S</namePart>
<namePart type="family">de Oliveira</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Arlindo</namePart>
<namePart type="given">Rodrigues</namePart>
<namePart type="family">Galvão Filho</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Anderson</namePart>
<namePart type="given">Da</namePart>
<namePart type="given">Silva</namePart>
<namePart type="family">Soares</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2026-10</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>The Proceedings of Speech and Audio Language Models Workshop (2nd Edition)</title>
</titleInfo>
<name type="personal">
<namePart type="given">Santosh</namePart>
<namePart type="family">T.Y.S.S.</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Sreyan</namePart>
<namePart type="family">Ghosh</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ramani</namePart>
<namePart type="family">Duraiswami</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Shinji</namePart>
<namePart type="family">Watanabe</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Siddhant</namePart>
<namePart type="family">Arora</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Sonal</namePart>
<namePart type="family">Kumar</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yossi</namePart>
<namePart type="family">Adi</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Dinesh</namePart>
<namePart type="family">Manocha</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Leibny</namePart>
<namePart type="given">Paola</namePart>
<namePart type="family">Garcia Perera</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">Budapest, Hungary</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
<identifier type="isbn">979-8-89176-496-5</identifier>
</relatedItem>
<abstract>We investigate representational alignment between nine independently trained speech encoders and ten text encoders. We build a controlled corpus in which the same sentences are rendered by different speakers and under different emotional conditions, so that prosody varies while linguistic content is held fixed. We compare every speech–text layer pair and calibrate the strongest match with a permutation procedure that prevents the search over pairs from inflating alignment scores. Local neighborhood agreement between modalities survives calibration throughout and is strongest between late speech layers and early text layers, providing strong evidence of genuine local alignment. Global similarity measures, however, do not provide reliable evidence that the two modalities share a common geometric space as a whole. Alignment is unaffected by speaker identity and prosodic variation, so the shared structure is tied to the content of the sentence itself rather than to how it is spoken. It also does not increase with model capability or with how well a text model reproduces human judgments of meaning, suggesting that this shared structure corresponds more closely to lexical and surface-level representations than to fine-grained sentence meaning. These findings indicate where speech and text representations can be connected, suggest that cross-modal transfer may benefit from exploiting neighborhood structure rather than assuming a shared global geometry, and provide new alignment methods with an existing lexical correspondence to build on, while pointing to sentence-level meaning as the structure that future speech representations could capture so that alignment extends beyond the lexical level.</abstract>
<identifier type="citekey">ferreira-etal-2026-echoes</identifier>
<location>
<url>https://aclanthology.org/2026.salma-1.10/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>134</start>
<end>155</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Echoes in Plato’s Cave: Measuring Global and Local Alignment Between Speech and Language Representations
%A Ferreira, Alef Iury Siqueira
%A Stefanel Gris, Lucas Rafael
%A de Oliveira, F. S.
%A Galvão Filho, Arlindo Rodrigues
%A Soares, Anderson Da Silva
%Y T.Y.S.S., Santosh
%Y Ghosh, Sreyan
%Y Duraiswami, Ramani
%Y Watanabe, Shinji
%Y Arora, Siddhant
%Y Kumar, Sonal
%Y Adi, Yossi
%Y Manocha, Dinesh
%Y Garcia Perera, Leibny Paola
%S The Proceedings of Speech and Audio Language Models Workshop (2nd Edition)
%D 2026
%8 October
%I Association for Computational Linguistics
%C Budapest, Hungary
%@ 979-8-89176-496-5
%F ferreira-etal-2026-echoes
%X We investigate representational alignment between nine independently trained speech encoders and ten text encoders. We build a controlled corpus in which the same sentences are rendered by different speakers and under different emotional conditions, so that prosody varies while linguistic content is held fixed. We compare every speech–text layer pair and calibrate the strongest match with a permutation procedure that prevents the search over pairs from inflating alignment scores. Local neighborhood agreement between modalities survives calibration throughout and is strongest between late speech layers and early text layers, providing strong evidence of genuine local alignment. Global similarity measures, however, do not provide reliable evidence that the two modalities share a common geometric space as a whole. Alignment is unaffected by speaker identity and prosodic variation, so the shared structure is tied to the content of the sentence itself rather than to how it is spoken. It also does not increase with model capability or with how well a text model reproduces human judgments of meaning, suggesting that this shared structure corresponds more closely to lexical and surface-level representations than to fine-grained sentence meaning. These findings indicate where speech and text representations can be connected, suggest that cross-modal transfer may benefit from exploiting neighborhood structure rather than assuming a shared global geometry, and provide new alignment methods with an existing lexical correspondence to build on, while pointing to sentence-level meaning as the structure that future speech representations could capture so that alignment extends beyond the lexical level.
%U https://aclanthology.org/2026.salma-1.10/
%P 134-155
Markdown (Informal)
[Echoes in Plato’s Cave: Measuring Global and Local Alignment Between Speech and Language Representations](https://aclanthology.org/2026.salma-1.10/) (Ferreira et al., SALMA 2026)
ACL