Libri-Adapt

本文介绍了一个新的数据集 Libri-Adapt，以支持对语音识别模型的无监督域自适应研究。 Libri-Adapt 建立在 LibriSpeech 语料库之上，包含在移动和嵌入式麦克风上录制的英语语音，跨越 72 个不同领域，代表了 ASR 模型遇到的具有挑战性的实际场景。更具体地说，Libri-Adapt 有助于研究 ASR 模型中由 a) 不同声学环境、b) 说话者口音的变化、c) 麦克风硬件和平台软件的异质性以及 d)上述三个班次。我们还提供了一些基线结果，量化了这些领域转移对 Mozilla DeepSpeech2 ASR 模型的影响。This paper introduces a novel dataset, Libri-Adapt, to support unsupervised domain adaptation research for speech recognition models. Libri-Adapt is built upon the LibriSpeech corpus and contains English speech recorded via mobile and embedded microphones, spanning 72 distinct domains that represent challenging real-world scenarios encountered by ASR models. More specifically, Libri-Adapt facilitates research on four types of domain shifts in ASR models: a) diverse acoustic environments, b) variations in speaker accents, c) heterogeneity of microphone hardware and platform software, and d) combinations of the aforementioned three shifts. We also provide baseline results that quantify the impact of these domain shifts on the Mozilla DeepSpeech2 ASR model.

搜集汇总

背景与挑战

背景概述

Libri-Adapt是一个基于LibriSpeech语料库构建的语音数据集，专为无监督域自适应研究设计，包含在移动和嵌入式麦克风上录制的英语语音，覆盖72个不同领域，以模拟实际场景中的声学环境、口音和硬件变化。该数据集由牛津大学等机构于2020年发布，用于支持语音识别模型的跨环境和跨设备适应研究。