The Two Gaps That Make Raw Speech-to-Text Hard to Audit

Generic speech-to-text fails in two specific ways, according to AWS: timestamps land at the utterance level, off by several seconds, and there is no reliable answer to “who said what.” The consequences are practical. Transcripts become hard to search, caption, redact or analyze at scale. AWS's write-up describes its preferred fix, WhisperX, as wrapping OpenAI's Whisper with batched inference, adding wav2vec2 forced alignment for per-word timestamps, and adding speaker diarization to label who spoke. The distinction worth keeping is between recognizing words and locating them: word-level timing plus speaker attribution is what makes a transcript usable for review rather than merely readable.