@inproceedings{hb-etal-2026-aura,
title = "{AURA}-{ST}: Acoustic-Unconstrained Residual Architecture for Speech Translation",
author = "HB, Barathi Ganesh and
Ptaszynski, Michal and
R, Jairam and
Unnikrishnan, Reshma",
editor = "Salesky, Elizabeth and
Anastasopoulos, Antonios and
Negri, Matteo and
Federico, Marcello",
booktitle = "Proceedings of the 23rd International Conference on Spoken Language Translation ({IWSLT} 2026)",
month = jul,
year = "2026",
address = "San Diego, USA (in-person and online)",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.iwslt-1.28/",
pages = "247--254",
ISBN = "979-8-89176-411-8",
abstract = "We present AURA-ST, a three-stage modular pipeline for low-resource speech-to-text translation submitted to the IWSLT 2026 African-Celtic Track 1. The architecture bypasses traditional cross-attention between audio and text modalities by treating projected acoustic representations as a native token prefix to a frozen large language model. A dual-stream encoder captures linguistic and paralinguistic features via a jointly trained semantic and a paralinguistic encoder. A convolutional subsampler then bridges the modality gap through a 4x temporal compression and a linear projection into the LLM embedding space. Finally, a MLP-targeted Low-Rank Adaptation adapter fine-tunes the frozen Gemma-4-E2B backbone for translation without catastrophic forgetting of base language model knowledge. We further identify and resolve the incompatibility between standard PEFT attention-level adapter injection and the Gemma-4 Per-Layer Embedding architecture that tends to cause gradient isolation. Trained on the IWSLT 2026 Track 1 data covering Hausa, Igbo, and Yoruba, the final system achieves a best proxy teacher-forced SacreBLEU of 91.29 on the validation set at Phase 3, with Phase 1 speech encoder validation loss converging to 0.651."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="hb-etal-2026-aura">
<titleInfo>
<title>AURA-ST: Acoustic-Unconstrained Residual Architecture for Speech Translation</title>
</titleInfo>
<name type="personal">
<namePart type="given">Barathi</namePart>
<namePart type="given">Ganesh</namePart>
<namePart type="family">HB</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Michal</namePart>
<namePart type="family">Ptaszynski</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jairam</namePart>
<namePart type="family">R</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Reshma</namePart>
<namePart type="family">Unnikrishnan</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2026-07</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Proceedings of the 23rd International Conference on Spoken Language Translation (IWSLT 2026)</title>
</titleInfo>
<name type="personal">
<namePart type="given">Elizabeth</namePart>
<namePart type="family">Salesky</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Antonios</namePart>
<namePart type="family">Anastasopoulos</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Matteo</namePart>
<namePart type="family">Negri</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Marcello</namePart>
<namePart type="family">Federico</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">San Diego, USA (in-person and online)</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
<identifier type="isbn">979-8-89176-411-8</identifier>
</relatedItem>
<abstract>We present AURA-ST, a three-stage modular pipeline for low-resource speech-to-text translation submitted to the IWSLT 2026 African-Celtic Track 1. The architecture bypasses traditional cross-attention between audio and text modalities by treating projected acoustic representations as a native token prefix to a frozen large language model. A dual-stream encoder captures linguistic and paralinguistic features via a jointly trained semantic and a paralinguistic encoder. A convolutional subsampler then bridges the modality gap through a 4x temporal compression and a linear projection into the LLM embedding space. Finally, a MLP-targeted Low-Rank Adaptation adapter fine-tunes the frozen Gemma-4-E2B backbone for translation without catastrophic forgetting of base language model knowledge. We further identify and resolve the incompatibility between standard PEFT attention-level adapter injection and the Gemma-4 Per-Layer Embedding architecture that tends to cause gradient isolation. Trained on the IWSLT 2026 Track 1 data covering Hausa, Igbo, and Yoruba, the final system achieves a best proxy teacher-forced SacreBLEU of 91.29 on the validation set at Phase 3, with Phase 1 speech encoder validation loss converging to 0.651.</abstract>
<identifier type="citekey">hb-etal-2026-aura</identifier>
<location>
<url>https://aclanthology.org/2026.iwslt-1.28/</url>
</location>
<part>
<date>2026-07</date>
<extent unit="page">
<start>247</start>
<end>254</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T AURA-ST: Acoustic-Unconstrained Residual Architecture for Speech Translation
%A HB, Barathi Ganesh
%A Ptaszynski, Michal
%A R, Jairam
%A Unnikrishnan, Reshma
%Y Salesky, Elizabeth
%Y Anastasopoulos, Antonios
%Y Negri, Matteo
%Y Federico, Marcello
%S Proceedings of the 23rd International Conference on Spoken Language Translation (IWSLT 2026)
%D 2026
%8 July
%I Association for Computational Linguistics
%C San Diego, USA (in-person and online)
%@ 979-8-89176-411-8
%F hb-etal-2026-aura
%X We present AURA-ST, a three-stage modular pipeline for low-resource speech-to-text translation submitted to the IWSLT 2026 African-Celtic Track 1. The architecture bypasses traditional cross-attention between audio and text modalities by treating projected acoustic representations as a native token prefix to a frozen large language model. A dual-stream encoder captures linguistic and paralinguistic features via a jointly trained semantic and a paralinguistic encoder. A convolutional subsampler then bridges the modality gap through a 4x temporal compression and a linear projection into the LLM embedding space. Finally, a MLP-targeted Low-Rank Adaptation adapter fine-tunes the frozen Gemma-4-E2B backbone for translation without catastrophic forgetting of base language model knowledge. We further identify and resolve the incompatibility between standard PEFT attention-level adapter injection and the Gemma-4 Per-Layer Embedding architecture that tends to cause gradient isolation. Trained on the IWSLT 2026 Track 1 data covering Hausa, Igbo, and Yoruba, the final system achieves a best proxy teacher-forced SacreBLEU of 91.29 on the validation set at Phase 3, with Phase 1 speech encoder validation loss converging to 0.651.
%U https://aclanthology.org/2026.iwslt-1.28/
%P 247-254
Markdown (Informal)
[AURA-ST: Acoustic-Unconstrained Residual Architecture for Speech Translation](https://aclanthology.org/2026.iwslt-1.28/) (HB et al., IWSLT 2026)
ACL