Folding Attention: Memory and Power Optimization for On-Device Transformer-based Streaming Speech Recognition
This work addresses memory and power optimization for on-device streaming speech recognition, offering incremental improvements for efficient deployment.
The paper tackled the bottleneck of linear projection layers in Transformer-based streaming speech recognition models, which cause high memory and power usage, by proposing folding attention, resulting in up to 24% reduction in model size and 23% reduction in power consumption without accuracy loss.
Transformer-based models excel in speech recognition. Existing efforts to optimize Transformer inference, typically for long-context applications, center on simplifying attention score calculations. However, streaming speech recognition models usually process a limited number of tokens each time, making attention score calculation less of a bottleneck. Instead, the bottleneck lies in the linear projection layers of multi-head attention and feedforward networks, constituting a substantial portion of the model size and contributing significantly to computation, memory, and power usage. To address this bottleneck, we propose folding attention, a technique targeting these linear layers, significantly reducing model size and improving memory and power efficiency. Experiments on on-device Transformer-based streaming speech recognition models show that folding attention reduces model size (and corresponding memory consumption) by up to 24% and power consumption by up to 23%, all without compromising model accuracy or computation overhead.