Fine-tuning ASR for corpus linguistics: Singapore English

Abstract Automatic speech recognition (ASR) is increasingly used to build large spoken-language corpora, yet its role is often treated as a purely technical preprocessing step. This paper argues that ASR configuration, especially model fine-tuning, constitutes a form of corpus design that directly shapes what linguistic analyses such resources can support. Focusing on Singapore English, we fine-tune six OpenAI Whisper model sizes on conversational data from Part 3 of Singapore’s National Speech Corpus and evaluate sixteen ASR systems, including baseline Whisper models and two alternative architectures, using word and character error rates on two held-out test sets. The best-performing fine-tuned model is then used to generate YCSEP_v2, an updated version of the YouTube Corpus of Singapore English Podcasts (approximately 620 h of audio). Evaluation goes beyond aggregate accuracy to examine linguistically salient phenomena. We compare corpus-level frequencies of discourse particles and disfluency markers and analyse three morphosyntactic variables (third-person singular -s, past tense zero marking, and copula absence). Fine-tuning yields substantial gains in transcription accuracy while improving the recovery of interactional features and reducing normalization towards mainstream English morphosyntax. The resulting transcripts diverge in systematic ways, demonstrating that ASR choices influence not only transcript quality but also corpus-based generalizations about linguistic feature frequencies and distributions. These findings position ASR adaptation as a methodological decision central to reproducible digital humanities and corpus-based linguistic research.