two reasons to transcribe a conference
you're covering it. you need quotes from the keynote by the afternoon and accurate attributions from a five-person panel. the transcript is a tool for finding and checking what was said.
you're publishing it. you ran the event and want every session online with captions, a searchable transcript, and maybe a write-up. the transcript is the deliverable.
the workflow is mostly the same. what changes is how much cleanup each file gets.
get the best audio source
the transcript can only be as good as the recording. in order of preference:
- the AV desk feed. the mixing board hears every lavalier and handheld directly. ask the AV team for the board recording or the session video files. at most events they'll hand it over if you ask on day one.
- the livestream or session video. same board feed, compressed. fine for transcription. if all you have is video, extract the audio first so you're not moving gigabytes around.
- a recorder on the press table or a speaker's lectern. usable for keynotes. weak for panels, where speakers sit metres away.
- your phone in the audience. works for one speaker at a podium. falls apart for audience questions. if this is your only option, sit near a PA speaker, not near the stage.
recording on your own laptop or phone? our voice recorder runs in the browser and saves locally, so a long session isn't lost to an upload that fails on conference wifi.
the workflow
- split by session. a full-day board recording is one eight-hour file. cut it at session boundaries before transcribing, using the agenda's start times. one file per session keeps speaker labels and exports sane. trim audio does this in the browser.
- load the agenda as vocabulary. speaker names, company names, product names, and the acronyms the industry uses. these are the words every speech model gets wrong. the program lists them all.
- label speakers from the agenda. a panel has a moderator and three to five panelists, and the program tells you who. listen to the moderator's introductions, then rename each detected speaker once.
- mark audience questions as audience. unless you know who asked, label them "audience member." a wrong name is worse than no name.
- verify every quote you use. click the first word of the quote and listen. panel audio is where models attach a line to the wrong panelist.
- export for where it's going. pulled quotes for a story, captions for the session video, a full transcript for the event site.
what goes wrong in conference audio
- the roaming mic. audience questions are quiet, off-mic, or half-shouted before the mic arrives. the speaker usually repeats the question. quote the repeat, not the original.
- panel crosstalk. two panelists agreeing over each other. most transcripts either drop one or merge them. mark it with a crosstalk tag rather than guessing.
- slides the speaker reads from. "as you can see here" isn't quotable. the transcript is accurate, and it still needs the slide to make sense.
- interpreted sessions. if there's simultaneous interpretation, the board may have both channels. transcribe the original-language channel for quotes and the interpreter channel for your own reference. don't quote the interpreter as the speaker.
publishing sessions after the event
for organizers, a transcript does three jobs: captions make the session video accessible, the text makes the talk findable in search, and a cleaned excerpt makes a better recap post than a summary written from memory.
- SRT and VTT captions for YouTube, Vimeo, or your own player.
- a blog post draft built from the talk, in the speaker's words.
- pull quotes for social posts, each with a timestamp back into the video.
get speakers' sign-off before publishing a full transcript. a talk given to a room isn't always meant for the open web, and some sessions run under the Chatham House rule.
press conferences are a different job
a press conference is one principal, a room of reporters, and a deadline measured in minutes. that workflow is about bulk speaker labels and fast quote checks. it has its own guide.