L’accuratezza della trascrizione ASR sul parlato non-standard. L’italiano nell’OH Portal

This paper aims at evaluating the transcription accuracy of the Google-IT’s ASR service available in the OH Portal. Data on its performance are limited to good quality recordings. Thus, we focus on suboptimal authentic materials, encompassing non-standard conversational speech recorded in noisy environments. We carry out a quanti-qualitative analysis of the linguistics and extra-linguistic parameters affecting accuracy and error distribution. The preliminary results show higher Word Error Rates for non-standard speech and low-quality recordings. Moreover, we seek the error patterns that could ease the transcription correction process for the users.