论文解读
Lattice-Guided Consistency Regularization of Dual-Mode Transducers for Automatic Speech Recognition
语音识别 | 8.0/10
一致性正则化
论文解读
Reducing the Offline-Streaming Gap for Unified ASR Transducer with Consistency Regularization
本文旨在解决训练单一自动语音识别（ASR）模型同时高效支持高精度离线转写和低延迟流式识别这一挑战。现有统一模型在低延迟流式模式下性能下降明显。作者提出了一个统一的RNN-Transducer (RNNT) 框架，其核心是结合了**带右上下文的chunk限制注意力**和**动态chunk卷积（DCCo