🤖 AI Summary
To address the high computational complexity, low energy efficiency, and poor adaptability of Transformer-based speech enhancement models to streaming, low-power edge scenarios, this work proposes a model–hardware co-optimization framework. Methodologically, it introduces domain-aware and streaming-aware joint pruning, a softmax-free attention mechanism, and batch-normalization-enhanced Transformer architecture to improve model lightweighting and hardware friendliness; additionally, it designs a 1D configurable processing array coupled with an SRAM address remapping scheme that eliminates memory skips. Experimental results demonstrate a 93.9% reduction in model size, requiring only 207.8K logic gates and 53.75 KB SRAM. Operating at 62.5 MHz, the system achieves a mere 8.08 mW power consumption while enabling real-time streaming speech denoising—significantly outperforming state-of-the-art edge speech enhancement solutions.
📝 Abstract
Transformer-based speech enhancement models yield impressive results. However, their heterogeneous and complex structure restricts model compression potential, resulting in greater complexity and reduced hardware efficiency. Additionally, these models are not tailored for streaming and low-power applications. Addressing these challenges, this paper proposes a low-power streaming speech enhancement accelerator through model and hardware optimization. The proposed high performance model is optimized for hardware execution with the co-design of model compression and target application, which reduces 93.9% of model size by the proposed domain-aware and streaming-aware pruning techniques. The required latency is further reduced with batch normalization-based transformers. Additionally, we employed softmax-free attention, complemented by an extra batch normalization, facilitating simpler hardware design. The tailored hardware accommodates these diverse computing patterns by breaking them down into element-wise multiplication and accumulation (MAC). This is achieved through a 1-D processing array, utilizing configurable SRAM addressing, thereby minimizing hardware complexities and simplifying zero skipping. Using the TSMC 40nm CMOS process, the final implementation requires merely 207.8K gates and 53.75KB SRAM. It consumes only 8.08 mW for real-time inference at a 62.5MHz frequency.