2022/11/17 by Tzu-Quan Lin, Lin, Tzu-Quan, Tsung-Huan Yang +11 · 1 citation
Computer Science · #Advanced Data Compression Techniques #Artificial intelligence #Compression (physics) #Computer science #Data compression #Data compression ratio #Image compression #Image processing #Inference #Machine learning #Speech Recognition and Synthesis #Speech and Audio Processing #Transformer
paper · pdf · doi:10.1109/asru65441.2025.11434686
openalex created_date 2022/11/28 · openalex publication_date 2025/12/06 · openalex updated_date 2026/08/05
Transformer-based self-supervised models have achieved remarkable success in speech processing, but their large size and high inference cost present significant challenges for real-world deployment. While numerous compression techniques have been proposed, inconsistent evaluation metrics make it difficult to compare their practical effectiveness. In this work, we conduct a comprehensive study of four common compression methods, including weight pruning, head pruning, low-rank approximation, and knowledge distillation on self-supervised speech Transformers. We evaluate each method under three key metrics: parameter count, multiply-accumulate operations, and real-time factor. Results show that each method offers distinct advantages. In addition, we contextualize recent compression techniques, comparing DistilHuBERT, FitHuBERT, LightHuBERT, ARMHuBERT, and STaRHuBERT under the same framework, offering practical guidance on compression for deployment.