Vistaar

Vistaar是一个专注于印度语言自动语音识别（ASR）的多样化基准数据集，由AI4Bharat等机构创建。该数据集包含59个基准，覆盖12种印度语言和多种领域/类型，总计超过10,700小时的音频数据。创建过程中，研究人员整合了多个公开可用的训练集，并特别设计了硬基准以增加挑战性。Vistaar的应用领域广泛，旨在通过提供多样化的数据集来评估和改进ASR系统，特别是在印度语言的语音识别技术上。

Vistaar is a diverse benchmark dataset focused on automatic speech recognition (ASR) for Indian languages, developed by institutions including AI4Bharat. It encompasses 59 benchmarks, covering 12 Indian languages and diverse domains/genres, with a total of over 10,700 hours of audio data. During its development, researchers integrated multiple publicly available training datasets and specifically designed hard benchmarks to elevate the task's challenging degree. Boasting wide-ranging application prospects, Vistaar aims to evaluate and improve ASR systems, particularly for Indian language speech recognition technologies, by providing this diverse benchmark dataset.

搜集汇总

背景与挑战

背景概述

Vistaar是一个包含59个基准和训练数据集的集合，涵盖12种印度语言，总时长超过10,700小时，用于自动语音识别（ASR）任务。该数据集支持IndicWhisper模型的训练，在多个基准测试中表现出较低的词错误率，适用于新闻、教育、文学和旅游等多种领域。