Reproduce “BR-ASR: Efficient and Scalable Bias Retrieval Framework for Contextual Biasing ASR in Speech LLM” on OpenResearch

About this paper

Xun Gong, Anqi Lv, Zhiming Wang, Huijia Zhu, Yanmin Qian

Researchers from Shanghai Jiao Tong University and AntGroup developed BR-ASR, a framework enabling efficient and scalable contextual biasing for Automatic Speech Recognition (ASR) systems leveraging Speech Large Language Models (SpeechLLMs). The method addresses limitations in handling large bias lists by achieving robust recognition of up to 200,000 bias words with a 2.8% Biased Word Error Rate (B-WER) on LibriSpeech `test-clean` while maintaining real-time latency.

Key insights

• Attention Pooling within the contrastive learning setup significantly improves retrieval efficiency and yields a 32% relative reduction in B-WER by better focusing on semantically rich speech features.
• Dynamic Curriculum Learning for homophone debiasing is crucial, reducing homophone retrieval recall by 30-40% and contributing an additional 23% relative reduction in B-WER.
• Encoding bias words directly using the textual LLM (`TextualBias`) generally outperforms synthesizing them into audio (`AcousticBias`), providing better semantic preservation and further reducing homophone confusion by 10%.

Summary by alphaXiv.