Voice Conversion Challenge 2018

语音转换 (VC) 是一种在保留源语音波形的语言信息的同时将源语音波形中包含的说话者身份转换为不同身份的技术。 2016 年语音转换挑战赛 (VCC) 于 2016 年在 Interspeech 2016 上发起。2016 年挑战赛的目标是更好地理解建立在可免费获得的通用数据集上的不同 VC 技术，以查看一个共同目标，并分享关于未解决的观点当前 VC 技术面临的问题和挑战。 VCC 2016 专注于最基本的 VC 任务，即构建 VC 模型，使用并行清洁训练数据库自动将源说话者的语音身份转换为目标说话者的语音身份，其中源和目标说话者读出相同的内容在专业录音室中的一组话语。 17 个研究小组参加了 2016 年的挑战。该挑战取得了成功，它建立了新的标准评估方法和协议，用于对 VC 系统的性能进行基准测试。 VCC 的第二版于 2018 年推出，即 VCC 2018。在第二版中，对挑战的三个方面进行了修订。首先，用于构建参与者VC系统的语音数据量减少了一半。这是基于上一个挑战参与者的反馈，这对于实际应用也是必不可少的。其次，除了与第一版类似的任务外，还引入了一项更具挑战性的任务，即 Spoke 任务，我们称之为 Hub 任务。在 Spoke 任务中，参与者需要使用非并行数据库构建他们的 VC 系统，在该数据库中源说话人和目标说话人读出不同的话语集。并行和非并行语音转换系统均通过相同的大规模众包听力测试进行评估。第三，还尝试弥合 ASV 和 VC 社区之间的差距。由于为 VCC 2018 开发的新 VC 系统可能是增强 ASVspoof 2015 数据库的有力候选者，因此评估了基于反欺骗分数的 VC 系统的欺骗性能。描述来自：https://datashare.ed.ac.uk/handle/10283/3061

Voice Conversion (VC) is a technique that converts the speaker identity contained in a source speech waveform to a different identity while preserving the linguistic information of the source speech waveform. The 2016 Voice Conversion Challenge (VCC) was launched at Interspeech 2016 in 2016.

The goal of the 2016 challenge was to foster better understanding of diverse VC technologies built on freely accessible general datasets, unify a shared research objective, and disseminate insights into unresolved issues and challenges confronting contemporary VC technologies. VCC 2016 focused on the most fundamental VC task: building a VC model that automatically converts the speech identity of a source speaker to that of a target speaker using a parallel clean training database, where both the source and target speakers utter the same set of utterances in a professional recording studio. Seventeen research teams participated in the 2016 challenge. This challenge was successful as it established new standardized evaluation methods and protocols for benchmarking the performance of VC systems.

The second edition of VCC, namely VCC 2018, was launched in 2018. Three aspects of the challenge were revised in this edition. First, the volume of speech data used to construct participants' VC systems was halved. This adjustment was based on feedback from participants in the previous challenge and is also essential for real-world applications. Second, in addition to tasks similar to those in the first edition, a more challenging task, the Spoke task, was introduced alongside the Hub task. In the Spoke task, participants are required to build their VC systems using non-parallel databases, where the source and target speakers utter different sets of utterances. Both parallel and non-parallel VC systems are evaluated via the same large-scale crowdsourced listening test. Third, efforts were made to bridge the gap between the ASV and VC communities. Since newly developed VC systems for VCC 2018 could serve as strong candidates for augmenting the ASVspoof 2015 database, the spoofing performance of VC systems was evaluated based on anti-spoofing scores.

The description is sourced from: https://datashare.ed.ac.uk/handle/10283/3061

搜集汇总

背景与挑战

背景概述

Voice Conversion Challenge 2018是一个用于评估语音转换技术的数据集，包含并行和非并行语音数据，旨在测试不同语音转换系统的性能，并探索语音转换与反欺骗技术的结合。数据集由爱丁堡大学语音技术研究中心于2018年发布。