Towards Zero-Shot SLU: An Empirical Study of Competing Architectural Paradigms

2026-06-29 · Source: Paper Index on ACL Anthology · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Natural Language Processing · Depth: Expert, quick

Summary

A comprehensive assessment of zero-shot Spoken Language Understanding (SLU) capabilities, specifically intent classification (IC), has been conducted across 13 languages. This study addresses the limitations of traditional supervised SLU models, which struggle to generalize to new domains due to prohibitive data annotation costs and difficulties in transferring domain-specific intents. Despite the promise of Large Language Models (LLMs) for zero-shot inference, their zero-shot SLU performance, particularly for speech-enabled LLMs, remained largely unexplored. The research systematically evaluates cascaded, end-to-end, and hybrid architectural paradigms for zero-shot SLU. It concludes that the hybrid approach is the most effective architectural design for end-to-end SLU, also assessing multilingual transfer capabilities. The findings detail challenges and opportunities, highlighting promising models and settings for zero-shot SLU.

Key takeaway

For NLP Engineers developing voice interaction systems, you should prioritize hybrid architectural designs for zero-shot Spoken Language Understanding (SLU). This approach has proven most effective for intent classification across 13 languages, mitigating high data annotation costs and improving domain generalization. Evaluate your models' multilingual transfer capabilities to ensure robust performance across diverse linguistic contexts.

Key insights

The hybrid architectural approach is most effective for zero-shot Spoken Language Understanding across multiple languages.

Principles

• Zero-shot SLU addresses data annotation costs.
• Multilingual transfer is key for SLU generalization.
• Architectural choice impacts zero-shot SLU efficacy.

Method

The study systematically evaluates cascaded, end-to-end, and hybrid architectures for zero-shot intent classification across 13 languages.

In practice

• Prioritize hybrid architectures for SLU.
• Test SLU models across diverse languages.
• Focus on intent classification as a sub-task.

Topics

• Spoken Language Understanding
• Zero-shot Learning
• Intent Classification
• Hybrid Architectures
• Multilingual Transfer
• Large Language Models

Best for: Research Scientist, AI Scientist, NLP Engineer, Machine Learning Engineer

Related on AIssential

• ArtNet: A JEPA-Like Articulatory Predictive Framework for Robust Zero-Shot Phoneme Recognition — ArtNet uses articulatory feature prediction and VIB to achieve robust zero-shot phoneme recognition across languages.
• Selective Capability Unlearning in End-to-End Spoken Language Understanding — Binding Subspace (BSU) framework effectively unlearns specific capabilities in SLU models by attenuating intent-conditioned representations, preventing reconstruction.
• The easiest way to improve prompts — Zero-shot and few-shot prompting are core techniques for controlling language model output based on guidance provided.
• CLP-Transfer: Cross-Lingual and Progressive Transfer Learning — CLP-Transfer simplifies cross-lingual model transfer using token overlap and a small helper model, but shows limited downstream task performance.
• Granite 4.0 1B Speech: Compact, Multilingual, and Built for the Edge — Granite 4.0 1B Speech offers high accuracy and multilingual support in a compact, efficient model for edge devices.
• Complex & Intelligent Systems, Volume 12, Issue 4, April 2026 — The issue highlights diverse AI advancements across networking, computer vision, NLP, and secure systems.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Paper Index on ACL Anthology.