AusKidTalk: developing an orthographic annotation workflow for a speech corpus of Australian English-speaking children

Child speech corpora are limited in number and scope, with none available for Australian English (AusE) until now, primarily due to orthographic transcription costs. Therefore, we developed AusKidTalk, the first AusE child speech corpus, a novel population due to speaker age and accent. Annotating AusKidTalk presented a circular problem: eliminating costly manual transcription required automatic speech recognition (ASR) tools not yet developed; but developing ASR tools required annotated speech corpora not available. This paper demonstrates how orthographic transcription burden was reduced in AusKidTalk via strategic data collection combined with out-of-domain ASR tools for automatic annotation augmented by manual correction. 620 children completed a single word production task, orthographic transcriptions were generated for 454 (73%) using the semi-automatic AusKidTalk pipeline and corrected manually for 394 using a custom Praat interface. An additional 58 children were transcribed without ASR assistance for comparison, yielding a total of 461 (74%) transcriptions. Workflow efficiency was evaluated on 380 children, producing a total of 136.6 h of speech. Annotation burden was reduced by (i) automatic orthographic transcription with 20% word error rate, removing the need to transcribe speech from scratch, and by (ii) the Praat interface allowing 136.6 h of speech to be corrected by listening to 18.5 h. Annotator reliability was good, with annotators achieving high similarity on 43/49 ground truth files. Correction time for one child was typically 1.5 h compared to the 4 h for transcription unassisted by ASR.