CosyVoice 3 Multilingual Dataset

该数据集是一个大规模的多语种音频数据集，来源于互联网上的有声读物、视频和播客。它旨在为语音合成模型提供训练，使其具备零样本多语言处理能力。该数据集涵盖了多样化的语言（包括9种语言和18种中文方言），并跨越了多个领域，如电子商务、导航、金融和教育等。此外，数据集采用了数据处理技术以确保数据的高质量。该数据集的规模达到了一百万小时的训练数据，任务目标是文本到语音（TTS）合成。

This is a large-scale multilingual audio dataset sourced from online audiobooks, videos, and podcasts. It is designed to train speech synthesis models to achieve zero-shot multilingual processing capabilities. The dataset covers a diverse set of languages, including 9 languages and 18 Chinese dialects, and spans multiple domains such as e-commerce, navigation, finance, education, and others. Moreover, data processing technologies are adopted to guarantee the high quality of the dataset. With a total of one million hours of training data, this dataset targets the task of text-to-speech (TTS) synthesis.

搜集汇总

背景与挑战

背景概述

CosyVoice 3是一个大规模多语言语音合成数据集，支持9种语言和18种中国方言，训练数据量达100万小时，模型参数15亿，旨在实现高质量的零样本语音合成。