University admissions increasingly rely on asynchronous video interviews, but manual review and transcription remain time-consuming and sensitive to transcription errors, particularly under accent variability. This paper presents a reproducible, privacy-oriented pipeline for processing authentic admissions interviews, including secure data preparation, automated personally identifiable information (PII) screening with human verification, dialect identification, and dialect-conditioned evaluation of automatic speech recognition (ASR) models. Dialect labels are inferred using a SpeechBrain ECAPA-TDNN accent identification model trained on CommonAccent. Transcription quality is assessed using word error rate (WER) for three transformer-based ASR systems: Whisper, HuBERT, and wav2vec2. Experiments are conducted on real applicant interviews and results are reported for dialect groups with more than 15 recordings. Across all evaluated dialects, Whisper yields the lowest WER (20–27%), while HuBERT and wav2vec2 exhibit higher error rates (31–46% and 44–64%, respectively). Dialect-aware ASR evaluation enhances transcription robustness across dialects, enabling fairer downstream decisions and more reliable early indicators for identifying students who may need support to mitigate attrition risk. The findings demonstrate that ASR accuracy is dialect-dependent in high-stakes educational settings and support dialect-conditioned model selection for more consistent transcription quality.