Abstract

Speech analysis is a crucial tool in discerning the complex cognitive and emotional subtleties of individuals. It holds a significant role in psychiatric research, particularly in the detection and understanding of psychopathological conditions such as psychosis. The process involves computational analysis of speech using natural language processing (NLP) tools, which necessitates a transcription of the speech. However, the manual transcription process is both time-consuming and costly, posing a substantial challenge to large-scale investigations. To address this, we explore the use of “Whisper”, an automated speech recognition (ASR) tool developed by OpenAI©, for transcribing psychiatric interviews in Spanish in heterogeneous environmental conditions. The specific objectives are to compare the transcription accuracy of Whisper with a manual transcription, determine and compare linguistic elements (noun phrases, determiners, and type–token ratio), and examine environmental elements that could alter the quality of the transcription. Sixteen interviews were transcribed using Whisper, and all of them had a manual reference transcription to be compared. A word error ratio (WER, which measures the insertions, deletions, and substitutions that are required to change one word for another) of 7.80% was obtained, with no significant differences by gender. Furthermore, no differences were found in the count and proportionality of nominal phrases, use of determiners, and the type–token ratio (TTR). The findings indicate that Whisper is a precise instrument for transcribing clinical interviews in Spanish. It has a minimal error rate and negligible loss of linguistic data, even in adverse conditions. This could streamline large-scale research endeavors in speech analysis within the clinical domain.