LibriSpeech
PulseAugur coverage of LibriSpeech — every cluster mentioning LibriSpeech across labs, papers, and developer communities, ranked by signal.
2 天有情绪数据
-
量化研究使更小、更准确的Whisper-small ASR成为可能
一篇新发布的arXiv研究论文评估了用于Whisper-small自动语音识别模型的各种训练后量化（PTQ）技术。该研究测试了PyTorch、Optimum-Quanto、HQQ和bitsandbytes等库，发现使用Quanto的动态int8量化在压缩和准确性之间取得了最佳平衡。该方法将模型大小减少了57%，同时在LibriSpeech数据集上略微提高了词错误率，使得Whisper-small更容易部署在资源受限的设备上。
-
New framework uses calculus to optimize ASR vocabulary size
Researchers have developed a calculus-based framework to determine the optimal vocabulary size for end-to-end Automatic Speech Recognition (ASR) systems. Unlike traditional hybrid ASR, end-to-end systems derive their vo…
-
New research explores text-only data for faster encoder-dominated speech recognition models
This paper introduces novel methods for enhancing speech recognition models by leveraging text-only data. The research focuses on encoder-dominated architectures, demonstrating that a larger encoder paired with a smalle…