SPREDS-D2 (SPeech Recognition Evaluation Data Set - Discourse type 2) ver1.1

What's new

• Released ver 1.1 (Japanese, English, Chinese) on 2025/08/29

Overview

SPREDS-D2 is a dataset of realistic, unscripted dialogue speech provided by NICT under the Creative Commons Attribution 4.0 International License (CC BY 4.0). It comprises Japanese, English, and Chinese, and includes audio recordings of real-time interviews conducted by professional interviewers with two to three experts per language, along with their corresponding transcriptions. The transcriptions include tags specified by NICT. Please refer to the 00README.txt and other documentation located in each language directory for more details.

Download

ver1.1--jpn/eng/zho

Directory structure

The data are compressed in xz format. After extraction, the directory structure is as follows. For details about files under LABEL/ and WAVE/, see 00README.txt.

-------------------------------------------------------------------------------------------
$ver =[version number]
$lang={01_jpn,02_eng,03_zho}

$ver/
  00_doc/ 
  $lang/
    individual/
      unsegmented/
         LABEL/
         WAVE/
      segmented/
         LABEL/
         WAVE/
    mixed/
      LABEL/
      WAVE/
-------------------------------------------------------------------------------------------