Multilingual and Fully Non-Autoregressive ASR with Large Language Model Fusion: A Comprehensive Study
This addresses latency issues in multilingual ASR for real-time applications, though it is incremental as it builds on existing models like USM and PaLM 2.
The paper tackles the latency bottleneck in autoregressive decoding for ASR by proposing a non-autoregressive LM-fused system, achieving average relative WER improvements of 10.8% on FLEURS and 3.6% on YouTube captioning.
In the era of large models, the autoregressive nature of decoding often results in latency serving as a significant bottleneck. We propose a non-autoregressive LM-fused ASR system that effectively leverages the parallelization capabilities of accelerator hardware. Our approach combines the Universal Speech Model (USM) and the PaLM 2 language model in per-segment scoring mode, achieving an average relative WER improvement across all languages of 10.8% on FLEURS and 3.6% on YouTube captioning. Furthermore, our comprehensive ablation study analyzes key parameters such as LLM size, context length, vocabulary size, fusion methodology. For instance, we explore the impact of LLM size ranging from 128M to 340B parameters on ASR performance. This study provides valuable insights into the factors influencing the effectiveness of practical large-scale LM-fused speech recognition systems.