Toward Elastic Speech Inference: Training-Free Wake-Word Detection from Pretrained ASR

2026-10-01

AI Summary

This study addresses the high resource consumption and limited generalization of existing wake word detection systems, which typically rely on fine-tuning or dedicated modules. We propose a gradient-free fine-tuning paradigm leveraging a shared ASR backbone. Building upon Parakeet and Moonshine models, our approach integrates PCA with SliceGPT-based structured pruning to directly derive task-specific compact encoders from pretrained models without additional training. Experimental results demonstrate that wake word detection performance remains stable even when the encoder channel dimension is reduced by 50%. By effectively balancing detection efficiency with model generalization, this work offers a promising new direction for lightweight voice interaction systems.

Abstract

Recent ASR development has placed growing emphasis on generalization across diverse domains and acoustic conditions. Existing approaches typically adapt pretrained ASR models to front-end functions such as wake-up word (WuW) detection through additional training or task-specific modules. In this work, we explore the use of a shared pretrained ASR backbone for WuW detection without gradient-based fine-tuning and examine whether a compact encoder can be extracted using the PCA-based structured pruning approach of SliceGPT. Experiments with Parakeet-TDT-0.6B-v3 and Moonshine-base show that WuW detection performance remains relatively stable when the encoder channel dimension is reduced by 50%. These results suggest that task-relevant compact encoders can be derived from pretrained ASR models without fine-tuning.

Problem

• Wake-Word Detection
• Pretrained ASR
• Training-Free
• Structured Pruning
• Compact Encoder

Innovation

• Training-Free
• Wake-Word Detection
• Structured Pruning
• SliceGPT
• Pretrained ASR