Skip to main navigation Skip to search Skip to main content

AusKidTalk: using strategic data collection and out-of-domain tools to semi-automate novel corpora annotation

Tünde Szalay*, Mostafa Shahin, Tharmakulasingam Sirojan, Zheng Nan, Renata Huang, Kirrie Ballard, Beena Ahmed

*Corresponding author for this work

Research output: Chapter in Book/Report/Conference proceedingConference proceeding contributionpeer-review

Abstract

Annotating speech corpora for novel populations presents a circular problem: eliminating costly manual transcription requires automatic speech recognition (ASR) tools not yet developed; but developing ASR tools requires annotated speech corpora not available. Manual transcription burden was reduced for AusKidTalk, a novel population due to speaker age and accent, by strategic data collection protocol combined with out-of-domain ASR tools for semi-automatic annotation. The data collection protocol inserted tones and timestamps to automatically segment the recordings. Automatic annotation was conducted by out-of-domain tools for diarisation (NeMo) and orthographic transcription (UNSW ASR). Transcription accuracy with 17% word error rate (WER) for single words and 23% WER for continuous speech allowed for hand-correction instead of transcription, reducing annotation burden. The workflow can be adapted for other corpora and updated with new ASR tools as they become available.

Original languageEnglish
Title of host publicationINTERSPEECH 2025
Subtitle of host publicationProceedings of the 26th Annual Conference of the International Speech Communication Association
Place of PublicationBaixas, France
PublisherInternational Speech Communication Association (ISCA)
Pages4268-4272
Number of pages5
DOIs
Publication statusPublished - 2025
EventInterspeech Conference (26th : 2025) - Rotterdam, Netherlands
Duration: 17 Aug 202521 Aug 2025

Publication series

NameProceedings of the Annual Conference of the International Speech Communication Association, INTERSPEECH
PublisherInternational Speech Communication Association
ISSN (Print)2308-457X
ISSN (Electronic)2958-1796

Conference

ConferenceInterspeech Conference (26th : 2025)
Country/TerritoryNetherlands
CityRotterdam
Period17/08/2521/08/25

Keywords

  • Australian English
  • automatic speech recognition
  • child speech
  • orthographic transcription
  • speech corpus

Fingerprint

Dive into the research topics of 'AusKidTalk: using strategic data collection and out-of-domain tools to semi-automate novel corpora annotation'. Together they form a unique fingerprint.

Cite this