A semi-automatic workflow for orthographic transcription of a novel speech corpus: A case study of AusKidTalk

Tünde Szalay, Louise Ratko, Mostafa Shahin, Tharmakulasingam Sirojan, Kirrie Ballard, Felicity Cox, Beena Ahmed

Friday, December 16th, 2022, 10.45am – 11.15am


Developing automatic speech recognition (ASR) tools for AusKidTalk, the novel Australian English (AusE) children’s corpus, presents a circular problem: AusKidTalk is designed to develop adequate ASR for AusE-speaking children; however, orthographic transcription of AusKidTalk requires ASR tools not yet developed. Our semi-automatic workflow augments existing (but inadequate) automatic tools with manual transcription. IBM-Watson diarisation and UNSW ASR orthographic transcription automatically generate Praat textgrids with timealigned orthographic transcriptions. A webtool distributes the textgrids, collects manual corrections, and implements consistency checks. Manual correction is conducted with a custom Praat interface. The output is a searchable, orthographically transcribed, and time-aligned corpus.