@inproceedings{sohler-etal-2026-quantizing,
title = "Quantizing Whisper: How Design Choices Affect {ASR} Performance",
author = {S{\"o}hler, Arthur and
Irigoyen, Julian and
S{\o}eborg Kirkedal, Andreas},
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-1.5/",
doi = "10.63317/4f2o9uqpxhn7",
pages = "39--46",
abstract = "Large speech recognition models like OpenAI{'}s Whisper achieve high accuracy but are difficult to deploy in resource-constrained environments due to their high memory and computational demands. This matters for low-resource and on-device settings, where compute and memory constraints often limit the practical use and evaluation of ASR systems. To address this, we present a unified, cross-library evaluation of post-training quantization (PTQ) on Whisper-small, comparing supported configurations across quantization scheme, method, granularity, and bit-width. Our study is based on four libraries{---}PyTorch, Optimum-Quanto, HQQ, and bitsandbytes. Experiments on LibriSpeech test-clean and test-other show that dynamic int8 quantization with Optimum-Quanto offers the best trade-off, reducing model size by 57{\%} while lowering Word Error Rate below the baseline. Additional experiments on Whisper-base and Whisper-tiny confirm these trends, though with more pronounced degradation at lower bit-widths. Static quantization performed worse, likely due to the absence of efficient low-bit implementations for operations such as LayerNorm and Softmax. More aggressive formats (e.g., nf4, int3) achieved up to 71{\%} compression at the cost of accuracy in acoustically challenging conditions. Our results demonstrate that carefully chosen PTQ methods can substantially reduce model size and inference cost without retraining, enabling efficient deployment of Whisper on constrained hardware."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="sohler-etal-2026-quantizing">
<titleInfo>
<title>Quantizing Whisper: How Design Choices Affect ASR Performance</title>
</titleInfo>
<name type="personal">
<namePart type="given">Arthur</namePart>
<namePart type="family">Söhler</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Julian</namePart>
<namePart type="family">Irigoyen</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Andreas</namePart>
<namePart type="family">Søeborg Kirkedal</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>Large speech recognition models like OpenAI’s Whisper achieve high accuracy but are difficult to deploy in resource-constrained environments due to their high memory and computational demands. This matters for low-resource and on-device settings, where compute and memory constraints often limit the practical use and evaluation of ASR systems. To address this, we present a unified, cross-library evaluation of post-training quantization (PTQ) on Whisper-small, comparing supported configurations across quantization scheme, method, granularity, and bit-width. Our study is based on four libraries—PyTorch, Optimum-Quanto, HQQ, and bitsandbytes. Experiments on LibriSpeech test-clean and test-other show that dynamic int8 quantization with Optimum-Quanto offers the best trade-off, reducing model size by 57% while lowering Word Error Rate below the baseline. Additional experiments on Whisper-base and Whisper-tiny confirm these trends, though with more pronounced degradation at lower bit-widths. Static quantization performed worse, likely due to the absence of efficient low-bit implementations for operations such as LayerNorm and Softmax. More aggressive formats (e.g., nf4, int3) achieved up to 71% compression at the cost of accuracy in acoustically challenging conditions. Our results demonstrate that carefully chosen PTQ methods can substantially reduce model size and inference cost without retraining, enabling efficient deployment of Whisper on constrained hardware.</abstract>
<identifier type="citekey">sohler-etal-2026-quantizing</identifier>
<identifier type="doi">10.63317/4f2o9uqpxhn7</identifier>
<location>
<url>https://aclanthology.org/2026.speakable-1.5/</url>
</location>
<part>
<date>2026-05</date>
<extent unit="page">
<start>39</start>
<end>46</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Quantizing Whisper: How Design Choices Affect ASR Performance
%A Söhler, Arthur
%A Irigoyen, Julian
%A Søeborg Kirkedal, Andreas
%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 sohler-etal-2026-quantizing
%X Large speech recognition models like OpenAI’s Whisper achieve high accuracy but are difficult to deploy in resource-constrained environments due to their high memory and computational demands. This matters for low-resource and on-device settings, where compute and memory constraints often limit the practical use and evaluation of ASR systems. To address this, we present a unified, cross-library evaluation of post-training quantization (PTQ) on Whisper-small, comparing supported configurations across quantization scheme, method, granularity, and bit-width. Our study is based on four libraries—PyTorch, Optimum-Quanto, HQQ, and bitsandbytes. Experiments on LibriSpeech test-clean and test-other show that dynamic int8 quantization with Optimum-Quanto offers the best trade-off, reducing model size by 57% while lowering Word Error Rate below the baseline. Additional experiments on Whisper-base and Whisper-tiny confirm these trends, though with more pronounced degradation at lower bit-widths. Static quantization performed worse, likely due to the absence of efficient low-bit implementations for operations such as LayerNorm and Softmax. More aggressive formats (e.g., nf4, int3) achieved up to 71% compression at the cost of accuracy in acoustically challenging conditions. Our results demonstrate that carefully chosen PTQ methods can substantially reduce model size and inference cost without retraining, enabling efficient deployment of Whisper on constrained hardware.
%R 10.63317/4f2o9uqpxhn7
%U https://aclanthology.org/2026.speakable-1.5/
%U https://doi.org/10.63317/4f2o9uqpxhn7
%P 39-46
Markdown (Informal)
[Quantizing Whisper: How Design Choices Affect ASR Performance](https://aclanthology.org/2026.speakable-1.5/) (Söhler et al., SPEAKABLE 2026)
ACL
- Arthur Söhler, Julian Irigoyen, and Andreas Søeborg Kirkedal. 2026. Quantizing Whisper: How Design Choices Affect ASR Performance. In Proceedings of Speech Language Models in Low-Resource Settings: Performance, Evaluation, and Bias Analysis (SPEAKABLE) @ LREC 2026, pages 39–46, Palma, Mallorca (Spain). ELRA Language Resources Association (ELRA).