Chang-Jae Chun

Junseok Lee, Nahoon Kim, Sangyong Lee et al.

Knowledge distillation is one of the most effective methods for model compression. Previous studies have focused on the student model effectively training the predictive distribution of the teacher model. However, during training, the student model may inherit the shortcomings of the teacher model, which can lead to a decline in generalization capacity. To mitigate this issue, we propose adaptive self-knowledge distillation (ASKD), which dynamically reduces the dependence of the teacher model to improve the self-training capacity, and performs the self-knowledge distillation method to improve the generalization capacity of the student model. We further distill the Whisper model into a smaller variant, called FastWhisper. In our post-training setting, FastWhisper achieved a word error rate of 1.07% lower than the teacher model Whisper, and its relative inference time was 5 times faster.

Junseok Lee, Sangyong Lee, Chang-Jae Chun

Although Multimodal Large Language Models (MLLMs) have demonstrated remarkable capabilities in vision, language, and video understanding tasks, scaling them to long-form speech remains a critical bottleneck due to the explosive growth of input tokens. Existing speech-language models typically project high-frame-rate acoustic features directly into the LLM input space, rendering long-context processing computationally prohibitive as audio duration increases. In this paper, we present FastSLM, a token-efficient architecture designed to overcome this scalability limit through extreme temporal compression. At its core is the Hierarchical Frame Querying Transformer (HFQ-Former), which progressively distills local acoustic details into compact, semantically rich representations across multiple temporal scales. This hierarchical abstraction reduces the speech representation rate to just 1.67 tokens per second, achieving a 93 percent reduction in tokens compared to standard frame-level adapters, while preserving the critical context required for complex reasoning. Experimental results demonstrate that FastSLM achieves competitive performance with state-of-the-art models on long-form benchmarks, despite operating with significantly lower FLOPs and parameter counts. Our findings establish that extreme token compression is a viable pathway to making real-time, long-context speech understanding feasible for LLMs, even under strict computational constraints. The source code and model checkpoints are available at https://anonymous.4open.science/r/FastSLM-8BD3