SpeechPrune: Context-aware Token Pruning for Speech Information Retrieval

TL;DR

SpeechPrune is a novel token pruning method that enhances long-context speech information retrieval by efficiently discarding irrelevant tokens, significantly improving accuracy and scalability in speech understanding tasks.

Contribution

We introduce SpeechPrune, a training-free token pruning strategy for Speech LLMs, enabling efficient long-form speech processing with substantial accuracy gains.

Findings

SpeechPrune improves accuracy by up to 47% over baseline models.

It maintains performance even at 80% token pruning.

Demonstrates potential for scalable long-form speech understanding.

Abstract

We introduce Speech Information Retrieval (SIR), a new long-context task for Speech Large Language Models (Speech LLMs), and present SPIRAL, a 1,012-sample benchmark testing models' ability to extract critical details from approximately 90-second spoken inputs. While current Speech LLMs excel at short-form tasks, they struggle with the computational and representational demands of longer audio sequences. To address this limitation, we propose SpeechPrune, a training-free token pruning strategy that uses speech-text similarity and approximated attention scores to efficiently discard irrelevant tokens. In SPIRAL, SpeechPrune achieves accuracy improvements of 29% and up to 47% over the original model and the random pruning model at a pruning rate of 20%, respectively. SpeechPrune can maintain network performance even at a pruning level of 80%. This approach highlights the potential of token-level pruning for efficient and scalable long-form speech understanding.