<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="tung-etal-2026-spikexl">
    <titleInfo>
        <title>SpikeXL: Language-Agnostic Spike Representations for Few-Shot Cross-Lingual Speech Command Recognition</title>
    </titleInfo>
    <name type="personal">
        <namePart type="given">Vu</namePart>
        <namePart type="given">Son</namePart>
        <namePart type="family">Tung</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Minh-Thuan</namePart>
        <namePart type="family">Bui</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Duc-Tho</namePart>
        <namePart type="family">Mai</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>Findings of the Association for Computational Linguistics: EMNLP 2026</title>
        </titleInfo>
        <name type="personal">
            <namePart type="given">Hung-Yi</namePart>
            <namePart type="family">Lee</namePart>
            <role>
                <roleTerm authority="marcrelator" type="text">editor</roleTerm>
            </role>
        </name>
        <name type="personal">
            <namePart type="given">Sunipa</namePart>
            <namePart type="family">Dev</namePart>
            <role>
                <roleTerm authority="marcrelator" type="text">editor</roleTerm>
            </role>
        </name>
        <name type="personal">
            <namePart type="given">Jessy</namePart>
            <namePart type="family">Li</namePart>
            <role>
                <roleTerm authority="marcrelator" type="text">editor</roleTerm>
            </role>
        </name>
        <name type="personal">
            <namePart type="given">Desmond</namePart>
            <namePart type="family">Elliott</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">979-8-89176-462-0</identifier>
    </relatedItem>
    <abstract>Deploying speech command recognition on multilingual edge devices demands models that are both energy-efficient and capable of cross-lingual transfer without per-language retraining—two goals rarely addressed jointly. Spiking Neural Networks (SNNs) offer substantial energy savings via sparse binary communication, yet their cross-lingual generalization remains unstudied. We present SpikeXL, a framework that investigates whether SNN spike sparsity provides an inductive bias toward language-agnostic representations. Our experiments suggest that the V_th firing threshold of Leaky Integrate-and-Fire (LIF) neurons suppresses language-specific surface phonetics while preserving command-level acoustic structure, yielding embeddings that transfer across languages more readily than dense ANN activations. Building on this finding, we introduce Spike-space Contrastive Alignment (SCA), a contrastive objective adapted for spike-rate embeddings that aligns semantically equivalent commands across languages without requiring bilingual lexicons or parallel speech data. On Speech-MASSIVE across 12 typologically diverse languages, SpikeXL achieves a Cross-Lingual Semantic Similarity (CLSS) of 0.912—outperforming a parameter-matched Transformer with the same SCA objective (0.763) and SNN without SCA (0.721)—indicating that spike sparsity itself contributes an inductive bias beyond what contrastive training alone provides. Few-shot k-NN classification on target languages (115 support samples, frozen encoder) reaches 80.8 macro-average accuracy, corroborating alignment quality with a metric fully independent of the SCA objective. On English GSC-v2, SpikeXL reaches 92.76 accuracy with 2.5M parameters and 45M synaptic operations per inference.</abstract>
    <identifier type="citekey">tung-etal-2026-spikexl</identifier>
    <identifier type="doi">10.18653/2026.findings-emnlp.1760</identifier>
    <location>
        <url>https://aclanthology.org/2026.findings-emnlp.1760/</url>
    </location>
    <part>
        <date>2026-10</date>
        <extent unit="page">
            <start>37645</start>
            <end>37657</end>
        </extent>
    </part>
</mods>
</modsCollection>

%0 Conference Proceedings
%T SpikeXL: Language-Agnostic Spike Representations for Few-Shot Cross-Lingual Speech Command Recognition
%A Tung, Vu Son
%A Bui, Minh-Thuan
%A Mai, Duc-Tho
%Y Lee, Hung-Yi
%Y Dev, Sunipa
%Y Li, Jessy
%Y Elliott, Desmond
%S Findings of the Association for Computational Linguistics: EMNLP 2026
%D 2026
%8 October
%I Association for Computational Linguistics
%C Budapest, Hungary
%@ 979-8-89176-462-0
%F tung-etal-2026-spikexl
%X Deploying speech command recognition on multilingual edge devices demands models that are both energy-efficient and capable of cross-lingual transfer without per-language retraining—two goals rarely addressed jointly. Spiking Neural Networks (SNNs) offer substantial energy savings via sparse binary communication, yet their cross-lingual generalization remains unstudied. We present SpikeXL, a framework that investigates whether SNN spike sparsity provides an inductive bias toward language-agnostic representations. Our experiments suggest that the V_th firing threshold of Leaky Integrate-and-Fire (LIF) neurons suppresses language-specific surface phonetics while preserving command-level acoustic structure, yielding embeddings that transfer across languages more readily than dense ANN activations. Building on this finding, we introduce Spike-space Contrastive Alignment (SCA), a contrastive objective adapted for spike-rate embeddings that aligns semantically equivalent commands across languages without requiring bilingual lexicons or parallel speech data. On Speech-MASSIVE across 12 typologically diverse languages, SpikeXL achieves a Cross-Lingual Semantic Similarity (CLSS) of 0.912—outperforming a parameter-matched Transformer with the same SCA objective (0.763) and SNN without SCA (0.721)—indicating that spike sparsity itself contributes an inductive bias beyond what contrastive training alone provides. Few-shot k-NN classification on target languages (115 support samples, frozen encoder) reaches 80.8 macro-average accuracy, corroborating alignment quality with a metric fully independent of the SCA objective. On English GSC-v2, SpikeXL reaches 92.76 accuracy with 2.5M parameters and 45M synaptic operations per inference.
%R 10.18653/2026.findings-emnlp.1760
%U https://aclanthology.org/2026.findings-emnlp.1760/
%U https://doi.org/10.18653/2026.findings-emnlp.1760
%P 37645-37657

Markdown (Informal)

[SpikeXL: Language-Agnostic Spike Representations for Few-Shot Cross-Lingual Speech Command Recognition](https://aclanthology.org/2026.findings-emnlp.1760/) (Tung et al., Findings 2026)

• SpikeXL: Language-Agnostic Spike Representations for Few-Shot Cross-Lingual Speech Command Recognition (Tung et al., Findings 2026)

ACL

• Vu Son Tung, Minh-Thuan Bui, and Duc-Tho Mai. 2026. SpikeXL: Language-Agnostic Spike Representations for Few-Shot Cross-Lingual Speech Command Recognition. In Findings of the Association for Computational Linguistics: EMNLP 2026, pages 37645–37657, Budapest, Hungary. Association for Computational Linguistics.