Auto-KWS

Auto-KWS 是一个用于定制关键字发现的数据集，即检测口语关键字的任务。该数据集与现实世界的场景非常相似，因为每个录音机都被分配了一个独特的唤醒词，并且可以自由选择他们的录音环境和熟悉的方言。所有数据均由近场手机记录（位于扬声器前方约 0.2m 距离处）。每个样本都以 16kHz 的采样率记录在单通道、16 位流中。有 4 个数据集：训练数据集、练习数据集、反馈数据集和私有数据集。从大约 100 个记录器记录的训练数据集用于参与者开发 Auto-KWS 解决方案。练习数据集包含 5 个说话人数据，每个数据包含 5 个注册音频数据和几个测试音频。练习数据集与可下载的 docker 一起提供了平台如何调用参与者代码的示例。训练和练习数据集都可以下载用于本地调试。反馈数据集和私有数据集具有相同格式的练习数据集，用于最终评估，因此将对参与者隐藏。

Auto-KWS is a dataset for customized keyword spotting (KWS), which refers to the task of detecting spoken keywords. This dataset closely aligns with real-world scenarios: each recorder (participant providing recordings) is assigned a unique wake word, and is free to choose their own recording environment and familiar dialects. All data was recorded using near-field mobile phones positioned approximately 0.2 meters in front of the speaker. Each sample is stored as a single-channel, 16-bit stream with a sampling rate of 16 kHz. There are four datasets in total: training dataset, practice dataset, feedback dataset, and private dataset. The training dataset, which includes recordings from approximately 100 recorders, is provided for participants to develop their Auto-KWS solutions. The practice dataset contains data from 5 speakers, with each speaker providing 5 registered audio samples and several test audio samples. Alongside a downloadable Docker container, the practice dataset offers examples demonstrating how the platform invokes participants' code. Both the training and practice datasets are available for download to facilitate local debugging. The feedback dataset and private dataset share the same format as the practice dataset, and are used for final evaluation, hence they will be withheld from participants.

搜集汇总

背景与挑战

背景概述

Auto-KWS是一个用于定制关键字发现的数据集，旨在检测口语中的关键词，其数据通过近场手机录制，模拟现实场景。该数据集包含训练、练习、反馈和私有四个子集，用于支持参与者开发解决方案并进行最终评估。