Searchable footage: transcribe your rushes and find any sentence in seconds

The maths of unlogged interview footage

Conversational speech runs at 140–160 words per minute. An hour of interview is therefore 8,500–9,500 words — thirty-odd pages of single-spaced text — and a modest corporate or documentary shoot day with four hours of interviews holds well over 35,000 spoken words. Reviewing that material in real time costs four hours before a single logging decision is made; even at 1.5× playback it eats most of a working day per shoot day.

The traditional answers were human ones: a producer logging paper timecode notes during the shoot, an assistant making selects from memory, or a paid transcription service returning documents days later. All three still exist, but the economics have inverted. On-device speech-to-text now transcribes an hour of audio in minutes on an ordinary Mac, at no per-minute cost — which turns the transcript from a luxury deliverable into something every interview-driven job should generate at ingest, automatically, before anyone asks for it.

What the model actually gives you

Modern on-device transcription is built on Whisper-class models — open-source speech recognition that runs locally through implementations such as whisper.cpp. The pipeline is unglamorous: audio is extracted to 16 kHz mono WAV (the model’s expected input), the model chews through it, and out comes a series of timestamped segments — a few seconds of text each, with a start and end time relative to the top of the file. That segment list, not the flowing prose, is the asset: it is what lets a search hit land on a moment rather than on a document.

Accuracy is good enough to change the workflow, and honest expectations help. On a clean lavalier recording of an unaccented single speaker, word error rates in the low single digits are normal — a transcript you can quote from with light checking. Push in background music, heavy accents, crosstalk or a camera-mounted mic three metres from the subject, and the error rate climbs steeply. The practical bar for searchable footage is lower than for publishable quotes: a transcript with a 10% word error rate still finds “the acquisition fell through” nine times out of ten, because search only needs the key words to survive, not the filler between them.

Timestamps are not timecode

The single most important detail in the whole workflow, and the one most casual setups get wrong: the model has never heard of timecode. Whisper timestamps count from 00:00:00 at the top of whatever file it was fed. Your clip does not start at zero — it starts at 10:22:15:00, because that is what the camera or recorder stamped, as covered in the LTC guide. A search hit at segment time 00:14:32 is useless to an editor until something adds the clip’s start timecode and answers 10:36:47:00.

That addition has to be done against the clip’s recorded start — the bext TimeReference in a broadcast WAV, or the camera’s TC track, exactly the identity metadata described in the BWF/iXML guide. Two consequences follow. First, transcribe the original media (or a frame-accurate proxy of it), never a trimmed or re-exported copy, because any head trim silently shifts every timestamp in the file. Second, the transcript must stay attached to the clip’s identity — filename, start TC, reel — so that a text hit resolves to a timeline position in any NLE, not just to an offset in a WAV that may later be renamed.

The workflow: transcribe at ingest, search the day, paper-cut

The version of this that holds up on real jobs is boring and systematic:

1. Transcribe everything at ingest, unattended. Transcription is exactly the kind of work that should happen in the same pass as copying, syncing and proxy generation — machine time, overnight if need be, never editor time. A shoot day that arrives in the edit suite already transcribed cost nobody anything.
2. Feed the model the best channel, not the mix. If the sound recordist delivered ISO tracks, transcribe the lavalier ISO — the cleanest, closest microphone — rather than a camera scratch track or a mixdown where two speakers overlap. Channel choice is the cheapest accuracy upgrade available; it routinely halves the error rate compared to the camera’s onboard mic.
3. Search across the day, not per clip. The payoff arrives when every clip’s segments live in one index: type a phrase, get every take where anyone said it, ranked by clip and timecode. Interview answers repeat across takes and subjects — seeing all five versions of the same story side by side is precisely what a paper log never gave you.
4. Paper-cut before you touch the timeline. For documentary and corporate storytelling, the fastest first assembly is still the radio cut: read the transcripts, highlight the sentences that carry the story, and only then pull those timecode ranges into the NLE. Reading is 3–4× faster than listening, and a producer who cannot drive an NLE can still highlight text.

Where the search lives in each NLE

Premiere Pro has the most complete native story: Speech-to-Text runs on-device, and Text-Based Editing turns the transcript into an editing surface — search the Text panel across a project, select a sentence in the source transcript, and insert exactly that range to the timeline. For interview-heavy work it is the reference implementation of the paper-cut made literal.

DaVinci Resolve Studio transcribes clips from the Media Pool (right-click → Audio Transcription), with text search inside the transcript window and the ability to mark and remove ranges from text. It is clip-oriented rather than project-oriented — strongest when you already know which interview to interrogate.

Avid predates the neural wave with two dedicated paid options: PhraseFind, which indexes clips phonetically so you can search for spoken words without any transcript existing, and ScriptSync, which aligns a prepared script or transcript against takes — the long-standing scripted-television answer.

Final Cut Pro has no built-in transcription; searchable-footage workflows there lean on transcripts generated upstream and attached as markers, captions or keyword ranges. Which is the general point: the NLE does not have to be where transcription happens. A transcript generated once at ingest, carried with the clip’s identity, serves every downstream tool — including the ones with no speech-to-text of their own — and belongs in the editor package alongside the sync report.

Fail modes that break searchable footage

• Transcribing the camera scratch mic. The most common quality killer: the model gets the echoey onboard track while a pristine lav ISO sits in the same folder. If the sync pass already matched sound to picture, transcribe the recorder’s audio, not the camera’s.
• Transcripts detached from identity. A folder of interview_final.txt files with no clip reference, or timestamps relative to a trimmed export, degenerate into documents someone has to read. The transcript is metadata; it must travel with filename, start TC and reel like everything else in the package.
• Auto language detection on a quiet opening. Language is typically detected from the first seconds of audio; room tone, an English “okay, rolling” before a Danish interview, or a bilingual shoot can steer detection wrong and produce fluent nonsense. On single-language jobs, force the language instead of trusting detection.
• Crosstalk and diarisation. Base Whisper-class transcription does not label speakers. On a two-person interview the words are usually all there, but interviewer and subject share segments — fine for search, misleading for quoting. If speaker labels matter editorially, keep the ISO-per-speaker structure so the channel itself identifies who spoke.
• Re-transcribing after a conform. If audio is later speed-conformed — the 0.1% pulldown family — timestamps from the original no longer land frame-accurately on the conformed media. Transcribe the media that will actually be cut, or re-map the timestamps with the conform.
• Sending NDA material to a cloud service. Unannounced products, embargoed interviews, legal and medical subjects: a per-minute cloud API is a data-processing agreement nobody signed. On-device transcription is not just cheaper — for a good share of professional work it is the only version the client can approve.

Where Launchr Post fits

Launchr Post treats transcription as part of ingest, not as a separate job: while it syncs and sorts a shoot day, it also transcribes every spoken word on-device with Whisper, keeps each timestamped segment attached to the clip’s identity, and puts filenames and spoken words in one search box — type a phrase, jump to the timecode where it was said. The transcripts land in the editor package with the sync report, so the search works before the NLE opens and regardless of which NLE it is. Nothing is uploaded anywhere; the audio never leaves your Mac.