What is transcript-tiler?

This is a tool for cleaning up audio transcripts, still an early work in progress. When speech to text software like Whisper turns audio into text, it gives you a rough guess of when each word starts and stops. Those guesses are often sloppy. This project fixes them, turning them into a precise, gap free labeling where every moment of the audio is marked as either a word, silence, or background noise. Unlike some competing tools that re-align the whole transcript against an acoustic model, this one just corrects the timestamps you already have, using simpler signals: how loud the audio is, a voice activity detector called silero VAD, and a measure of how noise-like versus tone-like a sound is. It does not re-transcribe anything or use a heavy acoustic model. In the author's own testing on one set of podcast recordings, this approach placed word boundaries more accurately than raw speech to text output or other alignment tools, though this has not been checked against a public benchmark. Beyond just marking boundaries, each word gets information about how confidently free its edges are, meaning how close the signal drops to silence right before and after it. That lets other tools decide which words are safe to cut out entirely, for example filler words like um and uh. The project includes a command line tool that can list out those safely cuttable filler words, and a sample script that uses ffmpeg to actually remove them from the audio file, with small fades added to avoid clicks at the cut points. It accepts transcripts from Whisper, WhisperX, or a simple generic word list format, and can export its results in formats compatible with linguistics tools like Praat and ELAN, or as an Audacity label track for manual review. It is built to handle multi hour recordings efficiently and caches results so re-running on the same file is instant. The project grew out of a larger toolchain the author built for editing podcasts.