LiSTEN: Learning Soft Token Embeddings for Neural Audio LLMs

2025-05-24

AI Summary

Large language models (LLMs) face significant challenges in general audio-language tasks, including substantial acoustic variability, poor cross-task generalization, and heavy reliance on large-scale ASR or captioning data. Method: We propose a dynamic soft prompt selection mechanism that employs a learnable key-value memory module to adaptively balance generic and task-specific knowledge; integrates soft token embeddings; and adopts single-stage end-to-end training—eliminating multi-stage fine-tuning. Contribution/Results: Our approach substantially reduces data dependency while enhancing cross-task interpretability and prompt discriminability. It achieves state-of-the-art performance across multiple speech-language understanding benchmarks, with fewer parameters and a more streamlined training process.

Technology Category

Machine Learning: Large Multimodal Models (LMMs)Natural Language Processing: (Large) Language ModelsPlanning, Routing, and Scheduling: Planning with Language Models

Application Category

User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendationSearch and Retrieval-Augmented AI: Search Tool Learning with LLM: Teaching LLMs to invoke search and make use of retrieved informationSemantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs

Abstract

Foundation models based on large language models (LLMs) have shown great success in handling various tasks and modalities. However, adapting these models for general-purpose audio-language tasks is challenging due to differences in acoustic environments and task variations. In this work, we introduce LiSTEN Learning Soft Token Embeddings for Neural Audio LLMs), a framework for adapting LLMs to speech and audio tasks. LiSTEN uses a dynamic prompt selection strategy with learnable key-value pairs, allowing the model to balance general and task-specific knowledge while avoiding overfitting in a multitask setting. Our approach reduces dependence on large-scale ASR or captioning datasets, achieves competitive performance with fewer trainable parameters, and simplifies training by using a single-stage process. Additionally, LiSTEN enhances interpretability by analyzing the diversity and overlap of selected prompts across different tasks.

Problem

• Adapting LLMs for general-purpose audio-language tasks
• Balancing general and task-specific knowledge in multitask settings
• Reducing dependence on large-scale ASR or captioning datasets

Innovation

• Dynamic prompt selection with learnable key-value pairs
• Balances general and task-specific knowledge efficiently
• Single-stage training with fewer parameters