Efficient method compresses speech for better zero-shot language model use
Align, Integrate, and Fire: Efficient Token-Level Alignment for Zero-Shot SpeechLLMs
Computation and Language
Summary
Understanding spoken language with large language models is hard because speech data is long and complicated. The authors designed a way to squash speech sounds into shorter bits that match the exact words without heavy computation. They first align sounds to text timings, then train a simpler system to link speech to meaning without needing the whole language model each time. This makes teaching the model to understand speech faster and less costly. Their experiments show better speech recognition and translation compared to older methods.
What this means in practice
- •For speech recognition engineers: Improve automatic speech-to-text systems by efficiently aligning speech frames to text tokens using less computation.
- •For real-time translation developers: Build speech translation tools that better convert spoken words to text with fewer resources and faster training.
Authors
Abderrahmane Issam, Yusuf Can Semerci, Jan Scholtes, Gerasimos Spanakis
Abstract
While Large Language Models excel in natural language processing, efficiently extending their capabilities to spoken input remains a significant challenge. Existing methods for building SpeechLLMs often rely on computationally expensive full-model fine-tuning, or employ parameter-efficient projectors that suffer from inefficient token sequence lengths and costly full-model supervision. In this paper, we introduce Aligned Continuous Integrate-and-Fire, a highly efficient framework for zero-shot speech processing. Our method dynamically compresses continuous acoustic frames into the exact discrete token length of the target text utilizing explicit Dynamic Time Warping alignments. This allows our initial training stage to establish a robust acoustic-to-semantic bridge using lightweight distance metrics, entirely bypassing the computationally expensive LLM forward pass. For subsequent fine-tuning, we propose a memory-efficient knowledge distillation objective that targets a single LLM layer, performing competitively with full-model cross-entropy training at a fraction of the computational cost. Through extensive evaluations on Automatic Speech Recognition and Speech Translation, we demonstrate that our method achieves superior performance compared to prior parameter-efficient baselines.