PublishingNovember 7, 2026
Transcription was only the first step
What the earliest Whisper and FFmpeg experiments taught us about complete creator workspaces.
NGNick GonzalesNovember 7, 2026 · 1 min read
Some of our earliest Bonfÿre experiments were built around audio: Whisper, FFmpeg, segmentation, timing and getting small local tools to work together. Turning a recording into text worked well enough to reveal a more interesting problem.
What does someone actually want to do with the recording?
A podcaster may need clips and show notes. A researcher needs searchable interviews with source timecodes. A resident might want to find a statement in a council meeting. A company may need one summary for staff and another version it can publish.
Keep the recording useful
A transcript is an interpretation of audio. It can miss jargon or speakers. The original recording and its time information are needed for corrections and accurate quotations.
The toolchain expanded from ingest and transcription into cleanup, paragraphing, indexing, narration, editing and publication. Different stages have different accuracy and review needs.
| Stage | Retained material | Human review |
|---|---|---|
| Original recording | Time-coded audio source | Rights and participant context |
| Transcription | Text aligned to source times | Names, jargon and speaker boundaries |
| Editorial arrangement | Paragraphs, excerpts, show notes | Meaning, pacing and factual fidelity |
| Narration and publication | Reviewed script and media sequence | Voice, transitions, loudness and distribution |
Preserve the source
The original audio and time references are required to verify corrections and quotations.
Edit for listening
An article narration needs spoken pacing, section transitions and appropriate production—not a fast raw text-to-speech dump.
Publish as a companion
A waveform, playback position and transcript should remain attached to the same article and editor, when reviewed media actually exists.