@inproceedings{kolluri-etal-2026-role,
title = "On the Role of Encoder Depth: Pruning Whisper and {L}o{RA} Fine-Tuning in {SLAM}-{ASR}",
author = "Kolluri, Ganesh Pavan Kartikeya Bharadwaj and
Kampouridis, Michael and
Shekhar, Ravi",
editor = "Hosseini-Kivanani, Nina and
Brutti, Alessio and
Matassoni, Marco and
Dowerah, Sandipana and
Liga, Davide and
Schommer, Christoph",
booktitle = "Proceedings of Speech Language Models in Low-Resource Settings: Performance, Evaluation, and Bias Analysis ({SPEAKABLE}) @ {LREC} 2026",
month = may,
year = "2026",
address = "Palma, Mallorca (Spain)",
publisher = "ELRA Language Resources Association (ELRA)",
url = "https://aclanthology.org/2026.speakable-1.20/",
doi = "10.63317/2pbwm223zmqx",
pages = "183--193",
abstract = "Automatic speech recognition (ASR) has advanced rapidly in recent years, driven by large-scale pretrained models and end-to-end architectures such as SLAM-ASR. A key component of SLAM-ASR systems is the Whisper speech encoder, which provides robust acoustic representations. While model pruning has been explored for the full Whisper encoder{--}decoder architecture, its impact within the SLAM-ASR setting remains under-investigated. In this work, we analyze the effects of layer pruning in the Whisper encoder when used as the acoustic backbone of SLAM-ASR. We further examine the extent to which LoRA-based fine-tuning can recover performance degradation caused by pruning. Experiments conducted across three Whisper variants (Small, Medium, Large-v2), three languages representing distinct resource levels (Danish, Dutch, English), and over 200 training runs demonstrate that pruning two encoder layers causes only 2{--}4{\%} WER degradation, and that combining this pruning with LoRA adaptation consistently outperforms the unpruned baseline while reducing total parameters by 7{--}14{\%}. Moreover, our error analysis reveals that LoRA primarily compensates through the language model{'}s linguistic priors, reducing total word errors by 18.2{\%}, with substitution errors showing the largest reduction. However, for low-resource Danish, LoRA introduces increased insertion errors, indicating that compensation effectiveness depends on the LLM{'}s pre-existing language proficiency and available training data."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="kolluri-etal-2026-role">
<titleInfo>
<title>On the Role of Encoder Depth: Pruning Whisper and LoRA Fine-Tuning in SLAM-ASR</title>
</titleInfo>
<name type="personal">
<namePart type="given">Ganesh</namePart>
<namePart type="given">Pavan</namePart>
<namePart type="given">Kartikeya</namePart>
<namePart type="given">Bharadwaj</namePart>
<namePart type="family">Kolluri</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Michael</namePart>
<namePart type="family">Kampouridis</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ravi</namePart>
<namePart type="family">Shekhar</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2026-05</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Proceedings of Speech Language Models in Low-Resource Settings: Performance, Evaluation, and Bias Analysis (SPEAKABLE) @ LREC 2026</title>
</titleInfo>
<name type="personal">
<namePart type="given">Nina</namePart>
<namePart type="family">Hosseini-Kivanani</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Alessio</namePart>
<namePart type="family">Brutti</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Marco</namePart>
<namePart type="family">Matassoni</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Sandipana</namePart>
<namePart type="family">Dowerah</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Davide</namePart>
<namePart type="family">Liga</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Christoph</namePart>
<namePart type="family">Schommer</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>ELRA Language Resources Association (ELRA)</publisher>
<place>
<placeTerm type="text">Palma, Mallorca (Spain)</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
</relatedItem>
<abstract>Automatic speech recognition (ASR) has advanced rapidly in recent years, driven by large-scale pretrained models and end-to-end architectures such as SLAM-ASR. A key component of SLAM-ASR systems is the Whisper speech encoder, which provides robust acoustic representations. While model pruning has been explored for the full Whisper encoder–decoder architecture, its impact within the SLAM-ASR setting remains under-investigated. In this work, we analyze the effects of layer pruning in the Whisper encoder when used as the acoustic backbone of SLAM-ASR. We further examine the extent to which LoRA-based fine-tuning can recover performance degradation caused by pruning. Experiments conducted across three Whisper variants (Small, Medium, Large-v2), three languages representing distinct resource levels (Danish, Dutch, English), and over 200 training runs demonstrate that pruning two encoder layers causes only 2–4% WER degradation, and that combining this pruning with LoRA adaptation consistently outperforms the unpruned baseline while reducing total parameters by 7–14%. Moreover, our error analysis reveals that LoRA primarily compensates through the language model’s linguistic priors, reducing total word errors by 18.2%, with substitution errors showing the largest reduction. However, for low-resource Danish, LoRA introduces increased insertion errors, indicating that compensation effectiveness depends on the LLM’s pre-existing language proficiency and available training data.</abstract>
<identifier type="citekey">kolluri-etal-2026-role</identifier>
<identifier type="doi">10.63317/2pbwm223zmqx</identifier>
<location>
<url>https://aclanthology.org/2026.speakable-1.20/</url>
</location>
<part>
<date>2026-05</date>
<extent unit="page">
<start>183</start>
<end>193</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T On the Role of Encoder Depth: Pruning Whisper and LoRA Fine-Tuning in SLAM-ASR
%A Kolluri, Ganesh Pavan Kartikeya Bharadwaj
%A Kampouridis, Michael
%A Shekhar, Ravi
%Y Hosseini-Kivanani, Nina
%Y Brutti, Alessio
%Y Matassoni, Marco
%Y Dowerah, Sandipana
%Y Liga, Davide
%Y Schommer, Christoph
%S Proceedings of Speech Language Models in Low-Resource Settings: Performance, Evaluation, and Bias Analysis (SPEAKABLE) @ LREC 2026
%D 2026
%8 May
%I ELRA Language Resources Association (ELRA)
%C Palma, Mallorca (Spain)
%F kolluri-etal-2026-role
%X Automatic speech recognition (ASR) has advanced rapidly in recent years, driven by large-scale pretrained models and end-to-end architectures such as SLAM-ASR. A key component of SLAM-ASR systems is the Whisper speech encoder, which provides robust acoustic representations. While model pruning has been explored for the full Whisper encoder–decoder architecture, its impact within the SLAM-ASR setting remains under-investigated. In this work, we analyze the effects of layer pruning in the Whisper encoder when used as the acoustic backbone of SLAM-ASR. We further examine the extent to which LoRA-based fine-tuning can recover performance degradation caused by pruning. Experiments conducted across three Whisper variants (Small, Medium, Large-v2), three languages representing distinct resource levels (Danish, Dutch, English), and over 200 training runs demonstrate that pruning two encoder layers causes only 2–4% WER degradation, and that combining this pruning with LoRA adaptation consistently outperforms the unpruned baseline while reducing total parameters by 7–14%. Moreover, our error analysis reveals that LoRA primarily compensates through the language model’s linguistic priors, reducing total word errors by 18.2%, with substitution errors showing the largest reduction. However, for low-resource Danish, LoRA introduces increased insertion errors, indicating that compensation effectiveness depends on the LLM’s pre-existing language proficiency and available training data.
%R 10.63317/2pbwm223zmqx
%U https://aclanthology.org/2026.speakable-1.20/
%U https://doi.org/10.63317/2pbwm223zmqx
%P 183-193
Markdown (Informal)
[On the Role of Encoder Depth: Pruning Whisper and LoRA Fine-Tuning in SLAM-ASR](https://aclanthology.org/2026.speakable-1.20/) (Kolluri et al., SPEAKABLE 2026)
ACL