SPREDS-U1: SPeech Recognition Evaluation Data Set - Utterance type 1 (ver1.0)

What's new

• 2023/7/4 ver1.0 has been released.

Overview

SPREDS-U1 is a set of evaluation data for multilingual speech recognition released by NICT under the Creative Commons Attribution 4.0 International License (CC BY 4.0). The data set consists of 21 languages from 22 countries/regions including the 12 languages (Japanese, English, Chinese, Korean, etc.) released as "SPREDS2", the 4 languages (Filipino, etc.) released as "SPREDS3", and the newly added languages including the Taiwanese dialect of Chinese, Arabic, German, Italia, Hindi, and Ukrainian. The data set contains audio data recorded under almost the same conditions (domain, number of speakers, recording environment, etc.) and their transcriptions. In addition, the formats of the Speaker IDs (naming conventions, etc.) which differed between SPREDS2 and SPREDS3 have been unified. A high-pass filter has been applied to the audio files to remove low-frequency noises. The transcriptions are raw transcriptions without any tags or labels. For further details, please refer to '00README.txt' in each directory.

Data set of 21 languages from 22 countries and regions

ver1.0--21lang(22regions)

Extracted directory

The files have been compressed in 'xz' format. The extracted directory should look like the following.

-------------------------------------------------------------------------------------------
$ver ={ver1.0}
$lang={01_jpn,02_eng,03_zho,04_kor,05_tha,06_vie,07_ind,08_mya,09_spa,10_fra,11_por_BRA,12_ara,13_rus,14_fil,15_khm,16_nep,17_mon,18_zho_TWN,19_ita,20_deu,21_hin,22_ukr}
$speaker_id=[see 00README.txt]

$ver/
  00_DOC/
    GCP_DialectCode_v1.2.3.xlsx
    README.txt
$lang/
  00README.txt
  LABEL/
    SPREDS-U1.$ver.label
  WAVE/
    $speaker_id/
      *.wav
-------------------------------------------------------------------------------------------