New research compares search complexity in end-to-end ASR models

A new research paper systematically compares the search space complexity and efficiency of end-to-end Automatic Speech Recognition (ASR) models, specifically Connectionist Temporal Classification (CTC) and neural transducer models. The study reveals that while these models typically operate with a small search space, they can experience significant temporary peaks in hypothesis count during moments of uncertainty. The research demonstrates that score-based pruning dynamically adjusts the search to model uncertainty, resulting in a substantially lower Real-Time Factor (RTF) and highlighting the limitations of fixed-size beam pruning. AI

IMPACT Provides practical guidelines for optimizing search efficiency in end-to-end ASR systems.

RANK_REASON The item is a research paper published on arXiv detailing a systematic comparison of ASR models. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CL →

• alphaXiv
• arXiv
• CatalyzeX Code Finder for Papers
• Connectionist Temporal Classification
• CORE Recommender
• DagsHub
• Gotit.pub
• Hidden Markov models
• Hugging Face
• Influence Flower
• language model
• ScienceCast

• paper
• other

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