Transcribing files

Meeting recordings, interviews, voice memos, the audio track of a video call. Drop them in the window and walk away — they queue up, run one at a time, and land in the same list your dictations do.

Last updated 2026-08-31

The Transcribe files card

The right-hand card in the capture band says Transcribe files. Two ways in:

Drop audio or video anywhere in the window. The whole window is a drop target, not just the card — you don't have to aim.
Choose Files… opens a normal file picker if you'd rather browse.

What it accepts

Audiowav · m4a · mp3 · aac · flac · ogg · wma · aiff

Videomp4 · mov · mkv · avi · wmv · webm · m4v

Video files are fine — LymeScribe takes the audio track. If you're recording on a phone, whatever its voice recorder produces will be on this list.

Linux Anything that isn't a WAV needs ffmpeg installed. It's what decodes every other format, and LymeScribe doesn't bundle it. Most desktops already have it; if yours doesn't, the app posts a capability notice rather than failing at the moment you drop a file. Install it from your distribution's packages.

Choosing an output format

There are three, and you pick before the job runs:

Plain textJust the words, no markup. Best when you're pasting into something else, or feeding it to an AI that doesn't need structure.

Speaker labelsEach line prefixed with who said it. The right choice for anything with more than one voice — meetings, interviews, calls. This is the default.

Timed captions (SRT)Subtitle file with timestamps, for captioning video or jumping back to a moment in the recording.

The card itself has an Output selector so you can set it per batch, and a note that it applies to new files — a job already sitting in the queue keeps whatever format it was added with, and you can change that on its own row. The app-wide default lives in Settings ▸ Output.

When to pick speaker labels. Any recording with two or more people. The labels come out as Speaker 1, Speaker 2 and so on — no software can know a stranger's name from their voice — and then you spend thirty seconds turning those into real names. That step is Naming speakers, and it's what makes a transcript usable as a record of who committed to what.

Speaker labels have their own hardware floor, and it's higher. They're a separate pass over the audio, and on a self-hosted machine they need more memory than plain transcription does — on Windows, an NVIDIA GPU specifically. The system requirements on the home page spell out the floors. If you're connecting to somebody else's server, this is their problem rather than yours: the app will simply offer the option if that server supports it.

How the queue behaves

Drop ten files and all ten are accepted immediately, but they are transcribed one at a time. This is deliberate: local speech models want the whole machine while they run, and three jobs sharing a GPU finish later than three jobs taking turns.

The status strip at the bottom of the window carries the queue summary — how many are waiting, what's running now. That strip is the single home for every machine fact in the app, so it's also where you'll find the engine's state, GPU and memory meters when you're hosting, and a link to the Activity log if you want to watch the detail.

The job finishes in place

There is one flat list of transcripts under the capture band, newest first, and an in-flight job is simply the newest row in it. It shows its progress there and then becomes the finished transcript in the same position — nothing moves, nothing jumps to a different screen, and there's no separate "in progress" area to go check.

That list holds roughly today and yesterday. Everything older lives in the History window, which is a second view over the same transcripts rather than a second copy of them — see History & the archive. Every destructive action (delete, clear, export in bulk) lives there and only there, which is why you can't accidentally destroy a transcript from the main window.

Where the work actually happens

Batch jobs go to whichever server you nominated as the batch server under Settings ▸ Servers ▸ Task routing. If you're hosting on your own machine, that's you. If you haven't assigned one, dropping a file tells you so directly and points you at the setting — it fails loudly rather than swallowing your file.

Long recordings take a while. A two-hour meeting on a mid-range machine is a coffee-break job, not an instant one, and the speaker-labels pass adds to that. You can close the window and let it run; on Mac the app lives in the menu bar, so closing the window doesn't quit it.

After it finishes

From a row in the list you can copy the transcript, play the audio if it was kept, reveal the file, or export it. If you chose speaker labels, this is the point at which you name them. If a word comes out wrong in every transcript you produce, the corrections dictionary fixes it going forward — both are covered next.