StreamHear pipeline enhances domain-adapted streaming speech recognition

Researchers have developed StreamHear, a novel semi-supervised learning pipeline designed to improve streaming automatic speech recognition (ASR) performance on domain-shifted audio. The system adapts a pretrained streaming student model by first fine-tuning an offline transducer teacher on available labeled data. This teacher then generates pseudo-labels for unlabeled audio, which are used to fine-tune the student. Additionally, StreamHear incorporates a realignment step to refine word placement using ASR hypothesis anchors, demonstrating consistent improvements across various datasets. AI

IMPACT Improves ASR performance on specialized audio by leveraging unlabeled data, potentially reducing costs for domain adaptation.

RANK_REASON The cluster contains a research paper detailing a new method for speech recognition. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CL →

• arXiv
• Hugging Face
• StreamHear

• paper
• other

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →