Delayed Fusion: Integrating Large Language Models into First-Pass Decoding in End-to-end Speech Recognition
This addresses efficiency and compatibility problems for ASR systems using LLMs, but it is incremental as it builds on existing fusion methods.
The paper tackles the computational cost and vocabulary mismatch issues of integrating large language models (LLM) into end-to-end speech recognition (E2E-ASR) by proposing 'delayed fusion,' which applies LLM scores with a delay during decoding, resulting in improved decoding speed and accuracy compared to shallow fusion and N-best rescoring on the LibriHeavy corpus with models like OpenLLaMA and Mistral.
This paper presents an efficient decoding approach for end-to-end automatic speech recognition (E2E-ASR) with large language models (LLMs). Although shallow fusion is the most common approach to incorporate language models into E2E-ASR decoding, we face two practical problems with LLMs. (1) LLM inference is computationally costly. (2) There may be a vocabulary mismatch between the ASR model and the LLM. To resolve this mismatch, we need to retrain the ASR model and/or the LLM, which is at best time-consuming and in many cases not feasible. We propose "delayed fusion," which applies LLM scores to ASR hypotheses with a delay during decoding and enables easier use of pre-trained LLMs in ASR tasks. This method can reduce not only the number of hypotheses scored by the LLM but also the number of LLM inference calls. It also allows re-tokenizion of ASR hypotheses during decoding if ASR and LLM employ different tokenizations. We demonstrate that delayed fusion provides improved decoding speed and accuracy compared to shallow fusion and N-best rescoring using the LibriHeavy ASR corpus and three public LLMs, OpenLLaMA 3B & 7B and Mistral 7B.